Recent Applications of Unsupervised Machine Learning in Structural Health Monitoring

1  Introduction

Structural health monitoring has matured from periodic inspections toward continuous sensing frameworks that record vibration, strain, acoustic and imaging data at high volume, which calls for algorithms that can make sense of changing patterns with minimal manual supervision [1–3]. Many civil and mechanical structures operate under variable temperature, humidity, traffic, and loading, which shapes their measured responses and can mask early signs of condition change if models rely on fixed baselines or labeled events gathered under narrow conditions [4–6]. Data streams from bridges, dams, buildings and pipelines are often incomplete or noisy, which makes label creation expensive and time consuming, while the event of interest may be rare or evolving in form [7–9].

Supervised learning typically assumes that representative labels exist for the conditions of interest and that class definitions remain stable over time. In structural health monitoring, these assumptions are difficult to satisfy in practice because true damage events are rare, labels are expensive to obtain, and imbalance between healthy and damaged states can dominate training and evaluation. Monitoring campaigns may also experience evolving damage manifestations, changing operational regimes, and sensor or network modifications that alter the data distribution over long horizons. These constraints motivate interest in unsupervised learning for novelty detection, clustering, and representation learning, where models can be trained primarily on abundant ambient measurements without relying on curated damage labels [10–12].

Unsupervised approaches can search for departures from healthy behavior, group similar states, and build compressed features that are less sensitive to variability, which suits output-only conditions common in civil infrastructure [13–15]. Recent studies emphasize that operational viability depends on robustness to environmental and operational variability, resilience to missing data, and stability as sensors age or monitoring networks evolve [16–18]. Work across the literature also shows that performance can degrade when feature distributions drift with seasons or operational patterns, so many studies focus on stabilizing decision rules under such drift [19–21].

Representation learning has emerged as a central mechanism for forming compact indicators from high-dimensional response histories, while anomaly scoring and thresholding remain core elements for operational decision making under unlabeled conditions [22,23]. In addition, long-term deployments often face missing segments and nonstationary disturbances, which has encouraged the use of generative and reconstruction-oriented ideas to preserve continuity and improve reliability in downstream assessment [7,24,25]. A parallel direction considers monitoring scenarios where permanent instrumentation is limited, which includes indirect and drive-by monitoring as well as image-based inference of structural response quantities [26–28].

Applications span long-span bridges, arch dams, rail systems, and offshore assets, where temperature variability and operational shifts present recurring hurdles for stable alarms across seasons and sites [29–31]. Population-based monitoring and domain adaptation aim to transfer knowledge across nominally similar structures, which is essential for regional programs that seek consistent decision rules across a fleet of bridges or wind assets [32–34]. Unsupervised learning is also increasingly adopted in guided waves and ultrasonic testing, where thermal cycles and access constraints make extensive calibration difficult [35–37]. Streaming settings further highlight the need for methods that adjust to concept drift while maintaining sensitivity to rare events, which links algorithm design to online updating and computational efficiency [13–15].

Interpretability and performance metrics remain central for practical adoption because operators must distinguish spurious alarms from meaningful condition changes under changing environments [4,38,39]. Advances in indirect monitoring, including drive-by methods and computer vision with learned features, also show that unsupervised learning can function with heterogeneous data sources and sparse labels, which broadens the reach of monitoring to resource-limited operators [40–42].

Despite this momentum, the literature remains dispersed across sensing modalities, structural types, and algorithm families, which makes it difficult to gauge prevailing methods, adoption patterns, and common limitations at a glance [43–45]. Prior reviews have tended to focus on specific sensors or single algorithmic themes, which leaves open questions regarding cross-cutting progress, reporting practices, and evidence strength across field and laboratory studies [12,46,47]. Bibliometric mapping has rarely been combined with a structured technical synthesis for unsupervised learning in structural health monitoring, which limits visibility into prolific venues, influential studies, and collaboration networks needed for coherent advancement. The literature lacks a consolidated account that draws together sensing domains, algorithm classes, evaluation practices, and adoption trends for unsupervised learning in structural health monitoring. The aim of this study is to conduct a systematic and bibliometric review that characterizes recent applications of unsupervised learning across structural systems, methods, and evaluation choices, while mapping trends and identifying recurring limitations.

2  Bibliometric Analysis

This study assembled a Scopus dataset to map research that links unsupervised machine learning with structural health monitoring (SHM). The search required the terms to appear in the title, abstract, or keywords and returned 221 records for 2000–2026. VOSviewer was then used to process the set and extract publication patterns and keyword relations.

Document types show a strong tilt toward journal outlets. Fig. 1 reports 158 articles, 58 conference papers, and 5 book chapters, which together match the 221-record total. The dominance of articles points to a field that favors archival reporting and method validation, while conferences remain a secondary route for early-stage results.

Figure 1: Types of indexed documents in Scopus database.

Publishing activity gathered pace only recently. Early years were quiet, with one paper in 2010 and one in 2011, followed by small steps between 2013 and 2016. A steady climb appears from 2018 onward, reaching 21 items in 2021, then 19 in 2022, and a clear surge in 2023 and 2024 with 41 and 39 items. The peak of the period is 2025 with 46 records. A small bar for 2026, three records, likely reflects partial indexing at the time of retrieval. These counts trace a rapid expansion of interest in unsupervised methods within SHM, as seen in Fig. 2.

Figure 2: Number of yearly publications in Scopus database.

Keyword co-occurrence brings the focus of the field into view. “Structural health monitoring” appears 193 times and “damage detection” 123 times. Method-oriented terms remain frequent, with “unsupervised learning” 67, “deep learning” 42, “auto encoders” 40, and “anomaly detection” 41. General tags include “machine learning” 34. Classic tools are present through “principal component analysis” 19 and “clustering algorithms” 20. Application terms include “bridges” 19, “environmental conditions” 16, and “guided electromagnetic wave propagation” 15. The phrase “unsupervised machine learning” itself occurs 17 times. These counts place SHM and damage detection at the center, with method clusters around autoencoders, principal components analysis (PCA), clustering, and convolutional approaches, and with application cues related to bridges, guided waves, cables, vibration, and crack detection. The network view in Fig. 3 shows these hubs and their connections.

Figure 3: Co-occurrence of keywords in previous studies in Scopus database.

From an engineering perspective, the rise of terms such as “anomaly detection,” “auto encoders,” and “deep learning” alongside persistent references to environmental conditions suggests that the field is converging on practical workflows that can sustain reliable alarms under real operational variability rather than relying on short, controlled datasets. The strong presence of unsupervised and deep feature-learning terms is consistent with the increased availability of dense sensing and long-duration monitoring archives, where manual labeling remains costly and often incomplete. In parallel, the appearance of application cues such as bridges and guided waves reflects the operational demand for methods that can function under output-only measurements and temperature-driven drift, which are common deployment constraints. These bibliometric patterns therefore indicate an applied research trajectory that prioritizes field robustness, environmental compensation, and computationally feasible representation learning for continuous monitoring.

3  Fundamentals of Unsupervised Machine Learning for Structural Health Monitoring

Unsupervised learning in structural health monitoring focuses on extracting condition information from unlabeled measurements gathered during routine operation. Civil infrastructure often runs under changing temperature, load and boundary conditions, so methods that organize responses, reduce dimensionality and indicate departures from normal behavior suit field monitoring where labeled events are scarce [1,13,43,48,49]. Reported studies show three model families at the center of current use. Clustering assigns unlabeled samples to consistent groups while isolating points that differ from the bulk of historical data, and this idea appears across vibration, acoustic emission and imaging datasets for a range of structures [50–53]. Novelty detection learns a boundary that characterizes healthy response, then scores future data against that boundary, which fits cases where only undamaged measurements exist [1,39,54,55]. Representation learning compresses high-dimensional measurements and tracks reconstruction error or latent changes as indicators of condition shift, with applications to bridges and laboratory structures using autoencoders and variational autoencoders [22,23,56].

Feature extraction and dimensionality reduction sit at the core of these systems. Classical approaches such as principal component analysis and frequency-response features remain common, and recent work adopts autoencoders and variational autoencoders to form low-dimensional descriptors that remain stable under seasonal variation [22,57,58]. Hybrid encoders with temporal models have been proposed for state evolution, where latent trends and reconstruction statistics serve as compact indicators across long horizons. Studies on benchmark bridges report that such encoders can compress many channels into a reliable latent space for continuous assessment [23,59].

Clustering remains a central device in practice because it organizes operating states while revealing outliers. Density-based and consensus strategies have grouped vibration and acoustic responses, often after a preliminary reduction step, and have been applied to masonry, bridges and composite components [3,52,60]. Healthy subspace and hypersphere formulations define compact regions of normal response and detect departures in a way that is resilient to moderate environmental drift, an idea extended to populations of similar structures [20,32,61]. Studies have also combined modal analysis with density-based spatial clustering of applications with noise (DBSCAN) for bridge scour assessment and tested immune-inspired recognition for composite and steel members, which shows the breadth of unsupervised grouping tools under field variability [62–64].

Generative modeling and domain adaptation address gaps that arise in long campaigns. Long-term datasets often contain outages, so imputation with generative adversarial networks has been adopted before detection and localization, and recurrent-generative hybrids have been used for damage localization under dynamic excitation [7–9,65]. Domain adaptation transfers features across structures and conditions, which is central in drive-by settings where vehicle characteristics differ between sites, and in high-arch dams with strong state changes after earthquakes [18,33,34,66]. Ultrasonic and guided-wave studies report that attention mechanisms and temperature-aware encoders help stabilize decisions across thermal cycles and operational shifts, and related work treats corrosion in underwater pipes and response shifts in arch dams within the same unsupervised transfer view [29,36,37,65,67].

Environmental and operational variability remains a major concern for vibration-based monitoring. Reported solutions include kernel null-space mappings, local metric learning and feature weighting that separate environmental influence from structural change [19,21,68]. Healthy boundaries that evolve with measurement periods maintain performance over slow shifts in temperature and loading, while hybrid designs link physical response expectations with learned features to stabilize decisions [17,20,32,69]. Bridge and dam case studies document the routine use of such normalization in long-term settings [30,70].

Unsupervised methods span several sensing modalities. Vibration and strain dominate, although guided waves, acoustic emission and imaging are increasingly present. Self-organizing maps and independent component analysis appear in vibration-based work, while acoustic emission studies apply shape-based clustering to sort fatigue activity in composites. Infrared and camera-based observations analyzed with deep clustering or unsupervised segmentation support assessment of reinforced concrete and cable systems [27,41,71–74]. Indirect and drive-by monitoring combine these ideas with vehicle responses, forming bridge indicators where permanent instrumentation is limited [18,34,40].

Streaming operation places attention on online learning and computational efficiency. Real-time novelty indices, moving windows and incremental encoders have been tested on railway bridges and laboratory structures, while fast algorithms address dense networks and long records [5,13–15,47,75,76]. Evaluation practice reflects these constraints. Receiver operating characteristic curves, precision–recall analysis and reconstruction statistics appear frequently, with benchmark datasets such as Z24 supporting comparative tests of autoencoders, one-class models and clustering [4,23,43,59]. Studies highlight that simple distance-based rules coupled with careful normalization often match more complex models in stability and transparency, which continues to drive adoption in the field [10,12,19,30,44,51].

Evaluation practice commonly relies on receiver operating characteristic and precision–recall curves for threshold setting in unlabeled contexts, and several comparisons find that simple algorithms paired with careful normalization achieve stable results [4,10,43,46]. Historical work on unsupervised statistical diagnostics provided early templates for these approaches and remains relevant for modern clustering designs [77–81].

4  Clustering-Based Approaches for Damage Detection and Pattern Recognition

Clustering organizes unlabeled data into groups that mirror consistent structural states, then isolates observations that depart from those groups. Early work introduced this idea for vibration data and event identification, laying the groundwork for later systems that couple clustering with feature learning and environmental normalization [1,3,44,82,83]. Density-based and consensus schemes are common because they do not require a preset number of clusters and can absorb irregular sampling, which suits vibration and acoustic emission contexts where operating patterns change across seasons or usage [52,62,73]. Case studies on bridges and masonry structures report condition changes that remain visible after grouping, even when measurements vary with temperature and traffic [60,84].

Prototype-based clustering, including self-organizing maps and k-means variants, continues to appear in vibration work owing to its computational economy and ease of visualization. Applications include unsupervised vibration-based detection with self-organizing maps and grouping of frequency response and strain features with k-means or genetic search [41,57,85]. Hierarchical clustering and immune-inspired recognition have been tested for composite and steel members, where a layered view of patterns helps relate grouped features to likely degradation paths [50,63,64].

Environmental and operational variability shapes cluster quality, so many contributions place normalization before or within the clustering stage. Healthy subspace and hypersphere models help separate normal and abnormal behavior across temperature cycles, and related designs apply kernel mappings and local metric adjustment to refine distances under drift [5,17,19–21,32,68,99]. Latent-space clustering has gained traction, where deep encoders first construct compact embeddings from vibration or deflection records and grouping occurs in that reduced space, often with improved separation of states [22,59]. Hybrid pipelines pair encoders with one-class classifiers or deep clustering networks when sensitivity to unseen conditions is required [39,54,100].

Table 1 highlights that clustering-based pipelines often achieve stable separation of operating states when the feature space is explicitly normalized for environmental and operational variability, whereas the same clustering rules can fragment the healthy state when seasonal drift enlarges dispersion or when sensor faults introduce non-structural outliers. Density-based clustering is generally tolerant of noise and irregular cluster shapes, yet it can misinterpret slow environmental drift as a new “cluster” and raise false alarms if thresholds are not drift-aware. Prototype-based methods can remain computationally economical, although their decision boundaries may become brittle when the response distribution is multi-modal or when the number of operating regimes changes over time. Consensus and ensemble designs tend to improve robustness to initialization and noisy channels, but the added computation and tuning overhead can reduce their attractiveness for resource-constrained deployments. These patterns motivate a recurring trade-off in field SHM, where modest increases in model complexity are valuable only when they yield demonstrably lower false-alarm rates under realistic environmental shifts and sensor degradation.

Guided-wave applications make extensive use of clustering with learned features. Temperature-compensated encoders followed by grouping have been reported for defect localization, and attention mechanisms or wavelet features extend this approach to environments with strong thermal variation [36,37,65,101,102]. Acoustic emission and infrared imaging provide further examples, where consensus clustering separates emission signatures over long fatigue tests and unsupervised segmentation partitions thermal images of reinforced concrete [72–74].

Indirect and drive-by monitoring adapt clustering to vehicle-induced responses. Feature spaces obtained from adversarial autoencoders or domain-adapted encoders are grouped to reveal bridge state variation that is less sensitive to vehicle type and speed, and this pattern appears across multiple sites and platforms [18,26,34,40,103]. Feature selection remains important for these pipelines. Principal component analysis, manifold learning and feature weighting remove redundant variability before grouping, while time-series modeling helps craft features that remain responsive under output-only conditions [11,13,57,58,104–106].

Probabilistic perspectives support decision making when data quality varies. Bayesian formulations and reliability-oriented clustering express uncertainty in group assignments, and recent work integrates probabilistic features from neural encoders to strengthen anomaly indication [91,95,107,108]. Full-scale experiments have validated many pipelines that combine encoders and clustering, including studies on operational bridges and streaming algorithms that update clusters in real time [15,38,47,87,109,110]. Work on imputation before clustering addresses missing data and maintains continuity, while recurrent-generative schemes supply embeddings suited for localization [7–9,65]. Data fusion with strain, acceleration and imaging features has also been tested, as have fleet-level applications in offshore wind where shared cluster definitions assist scheduling across assets [31,61,69,111,112].

Evaluation practice commonly relies on receiver operating characteristic and precision–recall curves for threshold setting in unlabeled contexts, and several comparisons find that simple algorithms paired with careful normalization achieve stable results [4,10,43,46]. Historical work on unsupervised statistical diagnostics provided early templates for these approaches and remains relevant for modern clustering designs [77–81].

5  Dimensionality Reduction and Feature Extraction Techniques in SHM

Feature extraction and dimensionality reduction convert large sensor datasets into compact indicators suitable for detection and tracking under unlabeled and variable conditions. Matrix factorization methods were among the first tools in this setting. Singular value decomposition has been used for virtual strain sensing and anomaly indication in operational bridges, and frequency-domain decomposition coupled with clustering has separated modal shifts from environmental effects in bridge monitoring [62,113]. Principal component analysis remains common for low-dimensional vibration features, including population-based projections that support cross-comparison within groups of similar structures [57,61,114]. Probabilistic decompositions such as Bayesian factor analysis have supported one-class assessments for cables, and factor models with genetic clustering have sorted operating conditions from strain measurements [55,115].

Nonlinear reduction has expanded as datasets increased in size and complexity. Kernel null-space learning and manifold-based selection aim to separate normal and abnormal behavior when temperature and boundary conditions vary, and local metric learning with feature weighting helps maintain stable relations across measurement periods [5,16,19,104]. Autoencoders play a central role in current practice. Deep and variational encoders trained on vibration data produce latent spaces and reconstruction errors that track condition change in bridges and laboratory structures, with results reported on the Z24 benchmark and other datasets [22,23,46,59,116]. Hybrid encoders combined with one-class support vector machines or probabilistic layers refine extracted features for unsupervised decisions [39,95].

Time-frequency and wavelet features remain important for guided waves and broadband vibration. Deep convolutional autoencoders have been trained on wavelet transmissibility patterns for bridge assessment, and temperature-compensated guided-wave encoders have run on edge devices for local inspection [36,37,97]. Multi-head attention encoders provide temporal focus on informative segments of vibration signals during variable conditions [117]. Sequential and generative models add further options. Bidirectional long short term memory (LSTM) encoders with distribution modeling support quantitative estimates of change, and generative adversarial network (GAN)-based designs supply reconstruction and localization support for dynamic tests [65,118]. Domain adaptation extends these representations to dams, pipelines and composite structures where operating conditions differ across sites [29,65,119].

Dimensionality reduction also appears as a preparatory step for cleaning and imputation. Generative approaches restore missing segments in bridge datasets, while convolutional cleaning with clustering removes noisy samples before feature extraction, which stabilizes downstream detection [7–9,90,120]. Population-based and transfer designs rely on shared latent spaces to compare structures and reuse information when labeled damage data are unavailable, as shown in bridges and pedestrian crossings under domain shifts [33,99,121]. Edge and embedded computing motivate compact encoders that meet power and bandwidth limits, with field programmable gate array (FPGA) and TinyML studies reporting real-time feasibility [87,122]. Data fusion through unsupervised encoders integrates strain, acceleration and imaging features, and related indirect methods merge visual and vibration cues for bridges where direct instrumentation is limited [26,28,112]. Imaging tasks have used encoders for deflection and tension estimation, and unsupervised transfer learning supports cross-dataset crack segmentation in composite members, while infrared thermography has been grouped through clustering features [27,74,119].

Table 2 shows that imaging and localization tasks often rely on representation learning to convert high-dimensional wavefields or images into compact latent descriptors that can drive probabilistic mapping or pixel-level segmentation without dense manual labels. A recurring contrast is that localization accuracy is strongly tied to whether the representation is stable under temperature and operational drift, which is why several studies embed explicit temperature compensation or baseline reconstruction steps before localization and segmentation decisions. Another visible trend is the movement from handcrafted tomography features toward end-to-end learned reconstructions that reduce artifacts and improve repeatability, although these gains are typically conditional on careful design of the reconstruction target so that damage-sensitive differences remain visible in the residual or latent space. These patterns clarify that the central practical challenge is not only extracting compact features, but also ensuring that the learned compression does not remove the subtle, damage-related variations that localization depends on under field variability.

Several studies discuss the trade-off between compression and information retention. Results indicate that moderate reduction can preserve sensitivity while improving generalization and interpretability for long-term monitoring, a point that aligns with the steady preference for clear features and simple thresholds in field deployment [58,106,131]. Early statistical diagnostics on response surfaces and delamination detection established this logic and continue to inform current designs that combine deep encoders with careful normalization [78,132].

6  Emerging Unsupervised and Hybrid Frameworks for Structural Condition Assessment

Recent work shows a shift toward unsupervised and hybrid frameworks that process large unlabeled datasets while accommodating variable environments. These emerging frameworks commonly treat environmental normalization, missing-data imputation, and domain adaptation as enabling layers that stabilize clustering, novelty detection, and representation learning when models face drift, outages, and cross-site variability in real deployments. Besides, these frameworks couple reconstruction-based learning with one-class decisions, transfer across sites through domain adaptation and incorporate probabilistic layers for uncertainty handling. Bridges, dams, composite structures and pipelines appear frequently in this stream of studies [45]. Autoencoders with one-class modules provide compact features and alarms that remain stable across environmental shifts, while recurrent-generative designs aim to support localization as well as detection [25,39,65,89].

Population-based monitoring extends these ideas from single structures to fleets. Shared latent spaces and meta-learning enable cross-comparison and adaptation under missing data or irregular sampling, which is relevant for regional programs and long corridors [61,70,99,133,134]. Hybrid designs couple statistical normalization and physical knowledge with learned features. Locally unsupervised adjustments remove environmental effects across measurement periods, and feature weighting or metric tuning treats mitigation as a learning task in its own right [16,17,21,135]. Physics-informed hybrids combine finite element expectations with unsupervised grouping or one-class detection for field bridges, seeking reliable decisions when conditions change widely [136,137].

Domain adaptation and transfer learning have become a regular part of these frameworks. Reported cases include high-arch dams after earthquakes, corrosion in underwater steel pipes and composite crack detection, as well as transfer between pedestrian and highway bridges [33,65,66,119,121]. Generative and sequential architectures supply the internal engines for these transfers. Variational and adversarial encoders learn compact dynamics, and bidirectional LSTM units describe temporal relations that support condition estimation and response forecasting with attention to normal variability and potential damage [25,65,118,123].

Hardware-aware learning brings these models closer to continuous deployment. FPGA studies show real-time autoencoder inference on operational bridges, while TinyML guided-wave implementations handle temperature compensation on low-power devices [37,87,122]. Probabilistic reasoning appears in several designs. Bayesian factor analysis has been used for cables under ambient vibration, and unsupervised deep models that quantify uncertainty in latent features supply more informative alarms for anomaly detection [55,95]. Hypersphere and subspace learning provide clear decision boundaries for robust operation under changing conditions [20,53].

Data reconstruction has moved from a convenience step to a standard layer. Generative imputation networks fill gaps before detection and localization, and convolutional cleaning with clustering eliminates noise and outliers that would otherwise bias encoders [7–9,90,120]. Guided-wave and vibration studies continue to report benefits from hybrids that combine learned features with physics-inspired representations, including multi-head attention LSTM encoders, wavelet transmissibility features and probabilistic imaging for localization under changing temperatures [97,101,117].

Multimodal fusion appears often in this literature. Encoders that merge strain, acceleration and imaging features supply richer condition descriptions for bridges, and computer vision systems based on adversarial encoders estimate quasi-static displacements and image-based indicators in indirect settings [26,28,112]. Infrared and camera data have been segmented in an unsupervised way for reinforced concrete and composite components, which extends SHM beyond purely vibration-based sensing [74,102]. Drive-by monitoring continues to benefit from unsupervised and hybrid adaptation. Domain-adapted encoders trained on vehicle responses form invariant features across vehicles and spans, adversarial and hierarchical learning support condition inference without direct instrumentation, and regression with dynamic time warping has been used on vehicle-induced signals for unsupervised damage assessment [18,34,40,88].

Research threads that test new computation forms and system views also appear. Unsupervised quantum machine learning has been proposed for vibration signals, and digital-twin models combine feature fusion with unsupervised learning to maintain virtual representations of bridges in service [138]. Outlier ensembles and deep clustering variants continue to serve as robust anomaly detectors under noise, while novelty indices and symbolic representations remain useful for online early-damage indication [3,14,38,139]. Recent contributions on context-free forecasting and spatio-temporal sparse integration aim to manage evolving data streams without dense labeling [25,86]. Hybrid detection informed by mechanical models creates links between numerical predictions and learned features, as seen in computer aided engineering (CAE)-aided encoders and deep one-class detection for in-situ bridges [136,140].

Score-guided regularization and covariance-based metrics have improved transparency in decision rules, and correlation-based assessment supplies indicators such as stiffness trends from periodic monitoring data [30,68,118,141]. Long-standing statistical work on unsupervised diagnosis for composite beams and civil structures established the idea of tracking response-surface change without labels, an idea that remains visible in modern latent-feature designs [78,80]. The overall direction points toward adaptable, integrated and interpretable systems that operate under minimal supervision while coping with missing data, environmental variation and multiple sensing modes. Continued refinement of normalization, generative modeling and edge deployment indicates sustained movement toward field-ready unsupervised monitoring.

Table 3 indicates that deployment-oriented progress is increasingly driven by methods that explicitly address missing data, cross-domain shifts, uncertainty reporting, and computational constraints rather than focusing only on isolated detection accuracy. A consistent contrast across these studies is that models with strong predictive or reconstruction capacity still require careful calibration of thresholds and uncertainty to avoid overconfident alarms under drift, while lightweight edge implementations must trade representation richness for power, memory, and latency limits. The table also highlights that transfer learning and domain adaptation are becoming central for network-wide monitoring, where the cost of per-site tuning can dominate operational budgets. These trends reinforce that future advances are likely to be judged by reliability under field variability, resilience to imperfect data streams, and feasibility of continuous operation rather than by peak accuracy on a single controlled dataset.

7  Discussion and Future Implications

This review indicates a steady shift from isolated algorithm trials toward integrated monitoring pipelines that can cope with variable environments, gaps in data, and limited labels. Three families shape current use, namely clustering, novelty detection, and representation learning. Clustering remains common because it groups operating states while exposing departures from routine behavior. Novelty detection suits long records that mainly capture undamaged conditions. Representation learning with autoencoders and related encoders supplies compact features and reconstruction measures that are less sensitive to seasonal drift. A clear theme across studies is the need to manage environmental and operational variability.

Kernel mappings, metric learning, and healthy subspace or hypersphere models appear frequently, often in combination with simple thresholding. Results from bridges and dams show that careful normalization before decision making can stabilize alarms across seasons, and several papers report that distance measures tuned to local conditions match or exceed deeper models in stability. Another consistent pattern is the use of transfer across sites. Domain adaptation is now routine in drive-by monitoring and in large civil assets, such as high arch dams, where conditions change after major events. These methods cut down on site specific tuning and make population based assessments possible. Related work uses shared latent spaces to compare nominally similar structures, which suits regional programs and fleets. Generative modeling and data cleaning have moved closer to the core of SHM pipelines. Imputation networks fill gaps that arise in long campaigns, while convolutional cleaning and outlier removal improve the reliability of later stages. Sequential encoders add support for damage localization under dynamic excitation. These additions matter in records with outages and noise, which are common in field settings.

Streaming operation is another focus. Online novelty indices, moving windows, and incremental encoders have been validated on rail bridges and laboratory structures. Edge and embedded implementations suggest that real time processing is feasible within power and bandwidth limits, which supports continuous monitoring without large data transfers. This trend links algorithm design with deployment constraints. Evaluation practice is improving. The reviewed literature still spans highly heterogeneous datasets, sensing layouts, damage definitions, and reporting conventions, which makes direct comparison of numerical performance across papers inherently uncertain even when the same model family is used. Differences in sampling rates, monitoring duration, environmental range, and the availability of true “ground-truth” damage states can shift reported scores substantially, while metrics such as accuracy, F1, ROC, and precision–recall may not be directly comparable when anomaly prevalence and thresholding rules differ. This heterogeneity reinforces the value of shared benchmarks and standardized evaluation protocols, including multi-season tests, open splits, and consistent reporting of false-alarm behavior under environmental drift, so that future studies can provide clearer evidence of generalization and operational reliability.

Receiver operating characteristic curves and precision–recall analysis appear often, which helps tune thresholds under class imbalance and changing environments. Studies that use public benchmarks, such as the Z24 bridge, allow direct comparisons between one class models, clustering, and autoencoder based methods. Reports continue to advise caution, since metrics can overstate performance if normalization and drift are not tested across seasons. Interpretability and decision clarity remain priorities for field use. Many authors favor compact features with simple rules, provided that preprocessing removes the main drivers of variability. Where deeper models are adopted, probabilistic layers and uncertainty estimates help operators judge alarm strength. The overall picture points to SHM systems that combine clear features, site adaptation, and streaming readiness, so that alarms remain reliable when sensors age, networks change, and operating patterns shift.

8  Conclusion

This study aims to assemble a systematic and bibliometric review of recent unsupervised learning methods in structural health monitoring, addressing a gap left by prior reviews that treated single sensors or narrow algorithm classes. It maps how clustering, novelty detection, and representation learning have been applied across structures and sensing modes, and how researchers handle environmental drift, missing data, and site transfer. The review also compiles evaluation choices and adoption patterns across field and laboratory studies.

Unsupervised learning in SHM clusters into three main uses, namely grouping of operating states, boundary-based alarms for healthy behavior, and compact representations through autoencoders and related encoders. Environmental and operational variation is managed most effectively when normalization and distance learning are placed before decisions, with healthy subspace and hypersphere models providing clear boundaries. Transfer across sites has matured, with domain adaptation supporting drive-by settings and large civil assets, which reduces tuning costs and enables fleet level assessment. Generative imputation and data cleaning are now core steps in long term records, since outages and noise would otherwise bias features and thresholds. Streaming readiness is improving through online indices, incremental encoders, and edge implementations, while evaluation practice increasingly relies on receiver operating characteristic and precision-recall analysis.

Practical deployment remains tightly linked to sensor reliability, data continuity, and operational scalability. Long-term monitoring campaigns often face sensor aging, intermittent dropouts, calibration drift, and changing network configurations, and these issues can mimic damage signatures if preprocessing and thresholds are not robust to such non-structural disturbances. Dense networks also raise computational and bandwidth constraints, which makes the reliability-complexity trade-off decisive when selecting models for field use. Edge and embedded implementations, together with compact feature representations, are therefore important not only for latency reduction but also for reducing dependence on continuous data transfer and centralized storage. These considerations reinforce that reliable SHM decisions depend on end-to-end feasibility, including the stability of sensing hardware, the resilience of the pipeline to missing data and environmental drift, and the interpretability of alarm logic for operators.

Finally, future work should prioritize multi season field trials with open data and code, shared protocols for normalization and domain shift tests, and side by side comparisons that include simple baselines. Further study on uncertainty reporting and alarm calibration will support decision making, and hardware aware designs will help connect research outcomes with continuous operation in bridges, dams, offshore assets, and similar systems.

Acknowledgement: Not applicable.

Funding Statement: The authors received no specific funding.

Author Contributions: The authors confirm their contribution to the paper as follows: study conception and design: Abdullah Alariyan, Abdulhadi Alzabout, Mohammed Alariyan, Anas Alaryan, Mahmoud Alhashash, Abdulrahman Ahmed, Mohammed Abdulaal, Ahed Habib; analysis and interpretation of results: Abdullah Alariyan, Abdulhadi Alzabout, Mohammed Alariyan, Anas Alaryan, Mahmoud Alhashash, Abdulrahman Ahmed, Mohammed Abdulaal, Ahed Habib; draft manuscript preparation: Abdullah Alariyan, Abdulhadi Alzabout, Mohammed Alariyan, Anas Alaryan, Mahmoud Alhashash; manuscript review & editing: Abdulrahman Ahmed, Mohammed Abdulaal, Ahed Habib. All authors reviewed and approved the final version of the manuscript.

Availability of Data and Materials: The data supporting this study are available from the corresponding author on reasonable request.

Ethics Approval: Not applicable.

Conflicts of Interest: The authors declare no conflicts of interest.

References

1. Oh CK, Sohn H. Damage diagnosis under environmental and operational variations using unsupervised support vector machine. J Sound Vib. 2009;325(1–2):224–39. doi:10.1016/j.jsv.2009.03.014. [Google Scholar] [CrossRef]

2. Entezami A, Shariatmadar H. An unsupervised learning approach by novel damage indices in structural health monitoring for damage localization and quantification. Struct Health Monit. 2018;17(2):325–45. doi:10.1177/1475921717693572. [Google Scholar] [CrossRef]

3. Bull LA, Worden K, Fuentes R, Manson G, Cross EJ, Dervilis N. Outlier ensembles: a robust method for damage detection and unsupervised feature extraction from high-dimensional data. J Sound Vib. 2019;453(3):126–50. doi:10.1016/j.jsv.2019.03.025. [Google Scholar] [CrossRef]

4. Giglioni V, García-Macías E, Venanzi I, Ierimonti L, Ubertini F. The use of receiver operating characteristic curves and precision-versus-recall curves as performance metrics in unsupervised structural damage classification under changing environment. Eng Struct. 2021;246(4):113029. doi:10.1016/j.engstruct.2021.113029. [Google Scholar] [CrossRef]

5. Sarmadi H, Entezami A, Behkamal B, De Michele C. Partially online damage detection using long-term modal data under severe environmental effects by unsupervised feature selection and local metric learning. J Civ Struct Health Monit. 2022;12(5):1043–66. doi:10.1007/s13349-022-00596-y. [Google Scholar] [CrossRef]

6. Yon KC, Bakhary N, Padil KH, Shamsudin MF. Unsupervised environmental operating condition compensation strategies in a guided ultrasonic wave monitoring system: evaluation and comparison. J Civ Struct Health Monit. 2024;14(4):861–84. doi:10.1007/s13349-024-00761-5. [Google Scholar] [CrossRef]

7. Jiang H, Wan C, Yang K, Ding Y, Xue S. Continuous missing data imputation with incomplete dataset by generative adversarial networks-based unsupervised learning for long-term bridge health monitoring. Struct Health Monit. 2022;21(3):1093–109. doi:10.1177/14759217211021942. [Google Scholar] [CrossRef]

8. Xie X, Lei Y, Xiang C, Li Y, Liu L. An enhanced generative adversarial imputation network with unsupervised learning for random missing data imputation of all sensors. Struct Control Health Monit. 2025;2025(1):8419570. doi:10.1155/stc/8419570. [Google Scholar] [CrossRef]

9. Zheng W, Li J, Hao H, Fan G. Missing data imputation for structural health monitoring using unsupervised domain adaptation and pretraining techniques. Eng Struct. 2025;328(2):119694. doi:10.1016/j.engstruct.2025.119694. [Google Scholar] [CrossRef]

10. Wang Z, Cha YJ. Unsupervised novelty detection techniques for structural damage localization: a comparative study. In: Model validation and uncertainty quantification. Vol. 3. New York, NY, USA: River Publishers; 2025. p. 125–32. doi:10.1007/978-3-319-54858-6_13. [Google Scholar] [CrossRef]

11. Soleimani-Babakamali MH, Sepasdar R, Nasrollahzadeh K, Lourentzou I, Sarlo R. Toward a general unsupervised novelty detection framework in structural health monitoring. Comput Aided Civ Infrastruct Eng. 2022;37(9):1128–45. doi:10.1111/mice.12812. [Google Scholar] [CrossRef]

12. Eltouny KA, Liang X. Bayesian-optimized unsupervised learning approach for structural damage detection. Comput Aided Civ Infrastruct Eng. 2021;36(10):1249–69. doi:10.1111/mice.12680. [Google Scholar] [CrossRef]

13. Entezami A, Shariatmadar H, Mariani S. Fast unsupervised learning methods for structural health monitoring with large vibration data from dense sensor networks. Struct Health Monit. 2020;19(6):1685–710. doi:10.1177/1475921719894186. [Google Scholar] [CrossRef]

14. de Almeida Cardoso R, Cury A, Barbosa F, Gentile C. Unsupervised real-time SHM technique based on novelty indexes. Struct Control Health Monit. 2019;26(7):e2364. doi:10.1002/stc.2364. [Google Scholar] [CrossRef]

15. Hoeltgebaum H, Adams N, Lau FD. Unsupervised streaming anomaly detection for instrumented infrastructure. Ann Appl Stat. 2021;15(3):1101–25. doi:10.1214/20-aoas1424. [Google Scholar] [CrossRef]

16. Sarmadi H, Entezami A, Magalhães F. Unsupervised data normalization for continuous dynamic monitoring by an innovative hybrid feature weighting-selection algorithm and natural nearest neighbor searching. Struct Health Monit. 2023;22(6):4005–26. doi:10.1177/14759217231166116. [Google Scholar] [CrossRef]

17. Entezami A, Mariani S. Mitigation of unmeasured environmental and operational variability in long-term bridge health monitoring by unsupervised statistical learning. In: Proceedings of the 11th European Workshop on Structural Health Monitoring (EWSHM 2024); 2024 Jun 10–13; Potsdam, Germany. doi:10.58286/29847. [Google Scholar] [CrossRef]

18. Liu J, Xu S, Bergés M, Noh HY. HierMUD: hierarchical multi-task unsupervised domain adaptation between bridges for drive-by damage diagnosis. Struct Health Monit. 2023;22(3):1941–68. doi:10.1177/14759217221081159. [Google Scholar] [CrossRef]

19. Sarmadi H, Yuen KV. Early damage detection by an innovative unsupervised learning method based on kernel null space and peak-over-threshold. Comput Aided Civ Infrastruct Eng. 2021;36(9):1150–67. doi:10.1111/mice.12635. [Google Scholar] [CrossRef]

20. Kyriakos VS, Spilios F, John S. An automated hypersphere-based healthy subspace method for robust and unsupervised damage detection via random vibration response signals. Struct Health Monit. 2022;21(2):465–84. doi:10.1177/14759217211004429. [Google Scholar] [CrossRef]

21. Daneshvar MH, Sarmadi H, Yuen KV. A locally unsupervised hybrid learning method for removing environmental effects under different measurement periods. Measurement. 2023;208:112465. doi:10.1016/j.measurement.2023.112465. [Google Scholar] [CrossRef]

22. Ma X, Lin Y, Nie Z, Ma H. Structural damage identification based on unsupervised feature-extraction via variational auto-encoder. Measurement. 2020;160:107811. doi:10.1016/j.measurement.2020.107811. [Google Scholar] [CrossRef]

23. Giglioni V, Venanzi I, Baia AE, Poggioni V, Milani A, Ubertini F. Deep autoencoders for unsupervised damage detection with application to the Z24 benchmark bridge. In: European Workshop on Structural Health Monitoring. Berlin/Heidelberg, Germany: Springer; 2022. p. 1048–57. doi:10.1007/978-3-031-07258-1_105. [Google Scholar] [CrossRef]

24. Luo Y, Guo X, Wang LK, Zheng JL, Liu JL, Liao FY. Unsupervised structural damage detection based on an improved generative adversarial network and cloud model. J Low Freq Noise Vib Act Control. 2023;42(3):1501–18. doi:10.1177/14613484221150804. [Google Scholar] [CrossRef]

25. Gao S, Zou Z, Zhou Z, Wan C, Xie L, Xue S. An unsupervised context-free forecasting method for structural health monitoring by generative adversarial networks with progressive growing and self-attention. Struct Health Monit. 2025;24(5):3233–54. doi:10.1177/14759217241269702. [Google Scholar] [CrossRef]

26. Calderon Hurtado A, Kaur K, Makki Alamdari M, Atroshchenko E, Chang KC, Kim CW. Unsupervised learning-based framework for indirect structural health monitoring using adversarial autoencoder. J Sound Vib. 2023;550(4):117598. doi:10.1016/j.jsv.2023.117598. [Google Scholar] [CrossRef]

27. Xu Y, Tian Y, Zhang Y, Li H. Deep-learning-based bridge condition assessment by probability density distribution reconstruction of girder vertical deflection and cable tension using unsupervised image transformation model. In: European Workshop on Structural Health Monitoring. Berlin/Heidelberg, Germany: Springer; 2021. p. 35–45. doi:10.1007/978-3-030-64908-1_4. [Google Scholar] [CrossRef]

28. Hurtado AC, Alamdari MM, Atroshchenko E, Chang KC, Kim CW. A data-driven methodology for bridge indirect health monitoring using unsupervised computer vision. Mech Syst Signal Process. 2024;210:111109. doi:10.1016/j.ymssp.2024.111109. [Google Scholar] [CrossRef]

29. Cao X, Chen L, Chen J, Li J, Lu W, Liu H, et al. Seismic damage identification of high arch dams based on an unsupervised deep learning approach. Soil Dyn Earthq Eng. 2023;168:107834. doi:10.1016/j.soildyn.2023.107834. [Google Scholar] [CrossRef]

30. Wei YT, Yi TH, Yang DH, Liu H, Deng Y. Unsupervised stiffness evaluation of high-speed railway bridges using periodic monitoring data. J Bridge Eng. 2024;29(3):04024002. doi:10.1061/jbenf2.beeng-6561. [Google Scholar] [CrossRef]

31. Yeter B, Garbatov Y, Guedes Soares C. Life-extension classification of offshore wind assets using unsupervised machine learning. Reliab Eng Syst Saf. 2022;219:108229. doi:10.1016/j.ress.2021.108229. [Google Scholar] [CrossRef]

32. Vamvoudakis-Stefanou KJ, Sakellariou JS, Fassois SD. Vibration-based damage detection for a population of nominally identical structures: unsupervised Multiple Model (MM) statistical time series type methods. Mech Syst Signal Process. 2018;111(3):149–71. doi:10.1016/j.ymssp.2018.03.054. [Google Scholar] [CrossRef]

33. Souza L, Yano MO, da Silva S, Figueiredo E. A comprehensive study on unsupervised transfer learning for structural health monitoring of bridges using joint distribution adaptation. Infrastructures. 2024;9(8):131. doi:10.3390/infrastructures9080131. [Google Scholar] [CrossRef]

34. Aggarwal J, Liu J, Noh HY. Vehicle-invariant drive-by monitoring across multiple bridges through bootstrappingenhanced unsupervised domain adaptation. In: Proceedings of the 14th International Workshop on Structural Health Monitoring; 2023 Sep 12–14; Stanford, CA, USA. doi:10.12783/shm2023/36909. [Google Scholar] [CrossRef]

35. Yang K, Kim S, Harley JB. Guidelines for effective unsupervised guided wave compression and denoising in long-term guided wave structural health monitoring. Struct Health Monit. 2023;22(4):2516–30. doi:10.1177/14759217221124689. [Google Scholar] [CrossRef]

36. Sawant S, Sethi A, Banerjee S, Tallur S. Unsupervised learning framework for temperature compensated damage identification and localization in ultrasonic guided wave SHM with transfer learning. Ultrasonics. 2023;130(3):106931. doi:10.1016/j.ultras.2023.106931. [Google Scholar] [PubMed] [CrossRef]

37. Kashyap P, Shivgan K, Patil S, Raja BR, Mahajan S, Banerjee S, et al. Unsupervised deep learning framework for temperature-compensated damage assessment using ultrasonic guided waves on edge device. Sci Rep. 2024;14(1):3751. doi:10.1038/s41598-024-54418-w. [Google Scholar] [PubMed] [CrossRef]

38. Santos JP, Crémona C, Calado L, Silveira P, Orcesi AD. On-line unsupervised detection of early damage. Struct Control Health Monit. 2016;23(7):1047–69. doi:10.1002/stc.1825. [Google Scholar] [CrossRef]

39. Abu Zouriq MF, Linzell DG, Eftekhar Azam S. Bridge damage identification using unsupervised deep learning through variational autoencoders. Struct Health Monit. 2026. doi:10.1177/14759217261417539. [Google Scholar] [CrossRef]

40. Calderon Hurtado A, Makki Alamdari M, Atroshchenko E, Chang KC, Kim CW. An unsupervised learning method for indirect bridge structural health monitoring. In: Experimental vibration analysis for civil engineering structures. Berlin/Heidelberg, Germany: Springer; 2023. p. 121–31. doi:10.1007/978-3-031-39117-0_13. [Google Scholar] [CrossRef]

41. Avci O, Abdeljaber O. Self-organizing maps for structural damage detection: a novel unsupervised vibration-based algorithm. J Perform Constr Facil. 2016;30(3):04015043. doi:10.1061/(asce)cf.1943-5509.0000801. [Google Scholar] [CrossRef]

42. Xu Y, Tian Y, Li H. Unsupervised deep learning method for bridge condition assessment based on intra-and inter-class probabilistic correlations of quasi-static responses. Struct Health Monit. 2023;22(1):600–20. doi:10.1177/14759217221103016. [Google Scholar] [CrossRef]

43. Eltouny K, Gomaa M, Liang X. Unsupervised learning methods for data-driven vibration-based structural health monitoring: a review. Sensors. 2023;23(6):3290. doi:10.3390/s23063290. [Google Scholar] [PubMed] [CrossRef]

44. Colace F, Gupta BB, Lorusso A, Troiano A, Santaniello D, Valentino C. Unsupervised learning techniques for vibration-based structural health monitoring systems driven by data: a general overview. In: Handbook of research on AI and ML for intelligent machines and systems. Hershey, PA, USA: IGI Global; 2023. p. 305–47. doi:10.4018/978-1-6684-9999-3.ch013. [Google Scholar] [CrossRef]

45. Boccagna R, Bottini M, Petracca M, Amelio A, Camata G. Unsupervised deep learning for structural health monitoring. Big Data Cogn Comput. 2023;7(2):99. doi:10.3390/bdcc7020099. [Google Scholar] [CrossRef]

46. Chalapathy R, Khoa NLD. Comparison of unsupervised shallow and deep models for structural health monitoring. Proc Inst Civ Eng Bridge Eng. 2024;177(2):107–17. doi:10.1680/jbren.21.00016. [Google Scholar] [CrossRef]

47. Meixedo A, Santos J, Ribeiro D, Calçada R, Todd MD. Online unsupervised detection of structural changes using train-induced dynamic responses. Mech Syst Signal Process. 2022;165(3):108268. doi:10.1016/j.ymssp.2021.108268. [Google Scholar] [CrossRef]

48. Worden K, Manson G. The application of machine learning to structural health monitoring. Philos Trans Math Phys Eng Sci. 2007;365(1851):515–37. doi:10.1098/rsta.2006.1938. [Google Scholar] [PubMed] [CrossRef]

49. Al Houri A, Habib A, Al-Sadoon ZA. Artificial intelligence-based design and analysis of passive control structures: an overview. J Soft Comput Civ Eng. 2025;9(3):137–68. [Google Scholar]

50. Chen B, Zang C. Unsupervised structure damage classification based on the data clustering and artificial immune pattern recognition. In: Artificial immune systems. Berlin/Heidelberg, Germany: Springer; 2009. p. 206–19. doi:10.1007/978-3-642-03246-2_21. [Google Scholar] [CrossRef]

51. Tibaduiza DA, Torres-Arredondo MA, Mujica LE, Rodellar J, Fritzen CP. A study of two unsupervised data driven statistical methodologies for detecting and classifying damages in structural health monitoring. Mech Syst Signal Process. 2013;41(1–2):467–84. doi:10.1016/j.ymssp.2013.05.020. [Google Scholar] [CrossRef]

52. Ramasso E, Placet V, Boubakar ML. Unsupervised consensus clustering of acoustic emission time-series for robust damage sequence estimation in composites. IEEE Trans Instrum Meas. 2015;64(12):3297–307. doi:10.1109/TIM.2015.2450354. [Google Scholar] [CrossRef]

53. Vamvoudakis-Stefanou K, Fassois S, Sakellariou J. An automated non-parametric healthy subspace method for unsupervised robust vibration-based damage detection under uncertainty. In: Proceedings of the Twelfth International Workshop on Structural Health Monitoring; 2019 Sep 10–12; Stanford, CA, USA. doi:10.12783/shm2019/32494. [Google Scholar] [CrossRef]

54. Wang Z, Cha YJ. An unsupervised deep auto-encoder with one-class support vector machine for damage detection. In: Data science in engineering. Vol. 9. Abingdon, UK: Talylor Francis Group; 2025. p. 99–104. doi:10.1007/978-3-030-76004-5_12. [Google Scholar] [CrossRef]

55. Wang X, Li L, Tian W, Du Y, Hou R, Xia Y. Unsupervised one-class classification for condition assessment of bridge cables using Bayesian factor analysis. Smart Struct Syst. 2022;29(1):41–51. doi:10.12989/sss.2022.29.1.041. [Google Scholar] [CrossRef]

56. Finotti RP, da Silva CF, Oliveira PHE, de Souza Barbosa F, Cury AA, Silva RC. Novelty detection on a laboratory benchmark slender structure using an unsupervised deep learning algorithm. Lat Am J Solids Struct. 2023;20(9):e512. doi:10.1590/1679-78257591. [Google Scholar] [CrossRef]

57. Siow PY, Ong ZC, Khoo SY, Lim KS. Damage sensitive PCA-FRF feature in unsupervised machine learning for damage detection of plate-like structures. Int J Str Stab Dyn. 2021;21(2):2150028. doi:10.1142/s0219455421500280. [Google Scholar] [CrossRef]

58. Soleimani-Babakamali MH, Soleimani-Babakamali R, Sarlo R, Farghally MF, Lourentzou I. On the effectiveness of dimensionality reduction for unsupervised structural health monitoring anomaly detection. Mech Syst Signal Process. 2023;187:109910. doi:10.1016/j.ymssp.2022.109910. [Google Scholar] [CrossRef]

59. Giglioni V, Venanzi I, Poggioni V, Baia A, Milani A, Ubertini F. Deep neural networks for unsupervised damage detection on the Z24 bridge. In: Proceedings of the 8th European Congress on Computational Methods in Applied Sciences and Engineering; 2022 Jun 5–9; Oslo, Norway. doi:10.23967/eccomas.2022.080. [Google Scholar] [CrossRef]

60. Barontini A, Masciotta MG, Mendes PA, Ramos LF. Performance assessment of a bio-inspired anomaly detection algorithm for unsupervised SHM: application to a Manueline masonry church. Int J Mason Res Innov. 2020;5(4):468. doi:10.1504/ijmri.2020.111798. [Google Scholar] [CrossRef]

61. Bel-Hadj Y, Weijtjens W. Population-based shm under environmental variability using a classifier for unsupervised damage detection. In: Proceedings of the 14th International Workshop on Structural Health Monitoring; 2023 Sep 12–14; Stanford, CA,USA. doi:10.12783/shm2023/36895. [Google Scholar] [CrossRef]

62. Belmokhtar M, Schmidt F, Chevalier C, Ture Savadkoohi A. Monitoring of a bridge experiencing scour using frequency domain decomposition mixed with DBSCAN algorithm: unsupervised modal analysis of output-only system. Arch Civ Mech Eng. 2025;25(4):193. doi:10.1007/s43452-025-01235-1. [Google Scholar] [CrossRef]

63. Chen B, Zang C. A hybrid immune model for unsupervised structural damage pattern recognition. Expert Syst Appl. 2011;38(3):1650–8. doi:10.1016/j.eswa.2010.07.087. [Google Scholar] [CrossRef]

64. Hu TY, Tian XB, Chen J, Yang X, Fu WW. Damage mode identification for bonding interface of steel-UHPC composite structures based on acoustic emission and fuzzy C-means unsupervised clustering. Structures. 2026;84(9):111060. doi:10.1016/j.istruc.2026.111060. [Google Scholar] [CrossRef]

65. Lu H, Chinchilla SC, Rotsaert B, Croxford A, Gryllias K, Chronopoulos D. Damage identification using ultrasonic Lamb waves with multi-scale adaptive attention Transformer-based unsupervised domain adaptation. Mech Syst Signal Process. 2025;234:112772. doi:10.1016/j.ymssp.2025.112772. [Google Scholar] [CrossRef]

66. Cao X, Chen L, Chen J, Li J, Lu W, Liu H, et al. Unsupervised domain adaptation damage identification approach of high arch dams after earthquakes. Struct Control Health Monit. 2023;2023:6349167. doi:10.1155/2023/6349167. [Google Scholar] [CrossRef]

67. Shao F, Li J, Ouyang L, Ren Y, Qiu L, Yuan S. Unsupervised adaptive feature enhancement network for reliable damage alarm using guided waves under time-varying conditions. Int J Smart Nano Mater. 2025;16(3):612–30. doi:10.1080/19475411.2025.2545443. [Google Scholar] [CrossRef]

68. Nikdel A, Shariatmadar H. Unsupervised damage assessment under varying ambient temperature based on an adjusted artificial neural network and new multivariate covariance-based distances. Struct Health Monit. 2024;23(6):3814–31. doi:10.1177/14759217231225402. [Google Scholar] [CrossRef]

69. Salar M, Entezami A, Sarmadi H, Behkamal B, De Michele C, Martinelli L. Vibration-based structural health monitoring of bridges based on a new unsupervised machine learning technique under varying environmental conditions. In: Current perspectives and new directions in mechanics, modelling and design of structural systems. Boca Raton, FL, USA: CRC Press; 2022. p. 1748–53. doi:10.1201/9781003348443-286. [Google Scholar] [CrossRef]

70. Mahmoudkelayeh M, Adhami B, Saeedi Razavi B. Continuous health assessment of bridges under sudden environmental variability by local unsupervised learning. J Perform Constr Facil. 2024;38(5):04024034. doi:10.1061/jpcfev.cfeng-4323. [Google Scholar] [CrossRef]

71. Farzampour A, Kamali-Asl A, Hu JW. Unsupervised identification of arbitrarily-damped structures using time-scale independent component analysis: part II. J Mech Sci Technol. 2018;32(9):4413–22. doi:10.1007/s12206-018-0839-8. [Google Scholar] [CrossRef]

72. Doan DD, Ramasso E, Placet V, Zhang S, Boubakar L, Zerhouni N. An unsupervised pattern recognition approach for AE data originating from fatigue tests on polymer-composite materials. Mech Syst Signal Process. 2015;64:465–78. doi:10.1016/j.ymssp.2015.04.011. [Google Scholar] [CrossRef]

73. Ramasso E, Butaud P, Placet V, Jeannin T, Sarasini F. Clustering acoustic emission time-series using unsupervised-shapelets. In: Proceedings of the Twelfth International Workshop on Structural Health Monitoring; 2019 Sep 10–12; Stanford, CA, USA. doi:10.12783/shm2019/32263. [Google Scholar] [CrossRef]

74. Pedram M, Taylor S, Hamill G, Robinson D. Quantification of subsurface defects in reinforced concrete of bridges by unsupervised segmentation of IR images. In: Proceedings of the IABSE Symposium Istanbul 2023: Long Span Bridges; 2023 Apr 26–28; Istanbul, Turkey. doi:10.2749/istanbul.2023.0835. [Google Scholar] [CrossRef]

75. Gopalakrishnan K, Mathews VJ. A fast unsupervised online learning algorithm to detect structural damage in time-varying environments. In: Proceedings of the 2021 48th Annual Review of Progress in Quantitative Nondestructive Evaluation; 2021 Jul 28–30. doi:10.1115/qnde2021-75247. [Google Scholar] [CrossRef]

76. Santos JP, Calado LM, Orcesi AD, Crémona CF. Adaptive detection of structural changes based on unsupervised learning and moving time-windows. In: Proceedings of the 7th European Workshop on Structural Health Monitoring; 2014 Jul 8–11; Nantes, France. [Google Scholar]

77. Iwasaki A, Todoroki A, Shimamura Y, Kobayashi H. Unsupervised structural damage diagnostic method using judgement of change of response surface by statistical tool. application for damage detection of composite structure. Trans Jpn Soc Mech Eng Ser A. 2002;68(673):1292–7. doi:10.1299/kikaia.68.1292. [Google Scholar] [CrossRef]

78. Iwasaki A, Todoroki A, Shimamura Y, Kobayashi H. An unsupervised statistical damage detection method for structural health monitoring (applied to detection of delamination of a composite beam). Smart Mater Struct. 2004;13(5):N80–5. doi:10.1088/0964-1726/13/5/n02. [Google Scholar] [CrossRef]

79. Abraham JN, Tran MQ, Jayaraj JS, Matos JC, Valluzzi MR, Dang SN. Unsupervised learning-based anomaly detection for bridge structural health monitoring: identifying deviations from normal structural behaviour. Sensors. 2026;26(2):561. doi:10.3390/s26020561. [Google Scholar] [PubMed] [CrossRef]

80. Iwasaki A, Todoroki A, Sugiya T, Izumi S, Sakai S. Unsupervised statistical damage diagnosis for structural health monitoring of existing civil structures. Smart Mater Struct. 2005;14(3):S154–61. doi:10.1088/0964-1726/14/3/018. [Google Scholar] [CrossRef]

81. Mohanty S, Chattopadhyay A, Wei J, Peralta PD. Unsupervised time-series fatigue damage state estimation of complex structure using ultrasound based narrowband and broadband active sensing. Struct Durab Health Monit. 2009;5(3):227–50. doi:10.1115/smasis2009-1432. [Google Scholar] [CrossRef]

82. Card L, McNeill DK. Novel event identification for SHM systems using unsupervised neural computation. In: Proceedings of the NDE for Health Monitoring and Diagnostics; 2004 Mar 14–18; San Diego, CA, USA. doi:10.1117/12.548114. [Google Scholar] [CrossRef]

83. de Lautour OR, Omenzetter P. Damage classification in structural systems using time series analysis and supervised and unsupervised clustering methods. In: Proceedings of the 20th Australasian Conference on the Mechanics of Structures and Materials; 2008 Dec 2–5; Toowoomba, QLD, Australia. doi:10.13140/2.1.3574.4329. [Google Scholar] [CrossRef]

84. Gagliardi V, Tosti F, Bianchini Ciampoli L, D’Amico F, Alani AM, Battagliere ML, et al. Monitoring of bridges by MT-InSAR and unsupervised machine learning clustering techniques. In: Proceedings of the Earth Resources and Environmental Remote Sensing/GIS Applications XII; 2021 Sep 13–18; Online. doi:10.1117/12.2597509. [Google Scholar] [CrossRef]

85. Silva M, Santos A, Figueiredo E, Santos R, Sales C, Costa JCWA. A novel unsupervised approach based on a genetic algorithm for structural damage detection in bridges. Eng Appl Artif Intell. 2016;52:168–80. doi:10.1016/j.engappai.2016.03.002. [Google Scholar] [CrossRef]

86. Wan K, Zhang W, Wang Y, Wang J, Ren J. Unsupervised data anomaly detection based on spatio-temporal sparse integration framework. J Comput Civ Eng. 2026;40(1):04025126. doi:10.1061/jccee5.cpeng-7002. [Google Scholar] [CrossRef]

87. Hoda MA, Mazzanti A, Shojaii J, Azam YE, Linzell DG. Unsupervised machine learning for bridge SHM using FPGA: proof of concept via full-scale experiments. Measurement. 2025;256:118717. doi:10.1016/j.measurement.2025.118717. [Google Scholar] [CrossRef]

88. Talukder R, Siddique F, Rana S. Regression (REG-1DCNN) based unsupervised bridge damage detection using vehicle induced acceleration response and dynamic time warping algorithm. Measurement. 2025;256(6):118380. doi:10.1016/j.measurement.2025.118380. [Google Scholar] [CrossRef]

89. Gunes B, Gunes O. An unsupervised hybrid approach for detection of damage with autoencoder and one-class support vector machine. Appl Sci. 2025;15(8):4098. doi:10.3390/app15084098. [Google Scholar] [CrossRef]

90. Deng Y, Zhao Y, He Y, Ju H, Li A. Unsupervised learning-based data cleaning framework for structural health monitoring using GoogLeNet and clustering models. Struct Health Monit. 2025;34(12):4491. doi:10.1177/14759217251348394. [Google Scholar] [CrossRef]

91. Liu J, Li Q, Li L, An S. Unsupervised structural damage detection and severity assessment via U-GraphFormer. Struct Health Monit. 2025. doi:10.1177/14759217251348421. [Google Scholar] [CrossRef]

92. Xue S, Shi Q, Xie L, Zhou S, Lu W, Zhang M. Outlier detection of monitored data and unsupervised recognition of construction activities during seismic performance enhancement of historic stone monuments. Eng Res Express. 2025;7(1):015116. doi:10.1088/2631-8695/adad33. [Google Scholar] [CrossRef]

93. Boratto T, Bernardino HS, Vieira AB, Gontijo TS, Bodini M, Martyushev DA, et al. An agglomerative clustering combined with an unsupervised feature selection approach for structural health monitoring. Infrastructures. 2025;10(2):32. doi:10.3390/infrastructures10020032. [Google Scholar] [CrossRef]

94. Wei C, Chu X, Huang M, Liu D. An unsupervised graph learning network for early damage intelligent detection of CFRP. IEEE Sens J. 2025;25(13):24320–34. doi:10.1109/JSEN.2025.3571600. [Google Scholar] [CrossRef]

95. Wan HP, Zhu YK, Luo Y, Todd MD. Unsupervised deep learning approach for structural anomaly detection using probabilistic features. Struct Health Monit. 2025;24(1):3–33. doi:10.1177/14759217241226804. [Google Scholar] [CrossRef]

96. Fang L, Zhu D, Zhang J, Dong Y. An unsupervised learning framework for health diagnosis by incorporating multiscale data. Structures. 2024;69(2):107396. doi:10.1016/j.istruc.2024.107396. [Google Scholar] [CrossRef]

97. Li S, Cao Y, Gdoutos EE, Tao M, Faisal Alkayem N, Avci O, et al. Intelligent framework for unsupervised damage detection in bridges using deep convolutional autoencoder with wavelet transmissibility pattern spectra. Mech Syst Signal Process. 2024;220:111653. doi:10.1016/j.ymssp.2024.111653. [Google Scholar] [CrossRef]

98. Shi S, Du D, Mercan O, Kalkan E, Wang S. A novel unsupervised real-time damage detection method for structural health monitoring using machine learning. Struct Control Health Monit. 2022;29(10):e3042. doi:10.1002/stc.3042. [Google Scholar] [CrossRef]

99. Entezami A, Sarmadi H, Behkamal B. Long-term health monitoring of concrete and steel bridges under large and missing data by unsupervised meta learning. Eng Struct. 2023;279:115616. doi:10.1016/j.engstruct.2023.115616. [Google Scholar] [CrossRef]

100. Liu J, Li Q, Li L, An S. Structural damage detection and localization via an unsupervised anomaly detection method. Reliab Eng Syst Saf. 2024;252(6):110465. doi:10.1016/j.ress.2024.110465. [Google Scholar] [CrossRef]

101. Song R, Sun L, Gao Y, Wei J, Peng C, Fan L, et al. Unsupervised temperature-compensated damage localization method based on damage to baseline autoencoder and delay-based probabilistic imaging. Mech Syst Signal Process. 2025;230(1):112649. doi:10.1016/j.ymssp.2025.112649. [Google Scholar] [CrossRef]

102. Liao Y, Wang Y, Fang C, Yang X, Zeng X, Chronopoulos D, et al. Baseline-free damage detection and localization on composite structures with unsupervised Kolmogorov-Arnold autoencoder and guided waves. Mech Syst Signal Process. 2025;239(5):113356. doi:10.1016/j.ymssp.2025.113356. [Google Scholar] [CrossRef]

103. Jiang GF, Wang SM, Ni YQ, Liu WQ. Unsupervised discrepancy-based domain adaptation network to detect rail joint condition. IEEE Trans Instrum Meas. 2023;72:3532319. doi:10.1109/TIM.2023.3316221. [Google Scholar] [CrossRef]

104. Wang T. Manifold learning-based unsupervised feature selection for structural health monitoring. In: Proceedings of the 10th Asia-Pacific Workshop on Structural Health Monitoring (10APWSHM); 2024 Dec 8–10; Sendai, Japan. doi:10.21741/9781644903513-9. [Google Scholar] [CrossRef]

105. Entezami AR, Shariatmadar H, Karamodin A. Improving feature extraction via time series modeling for structural health monitoring based on unsupervised learning methods. Sci Iran. 2020;27(3A):1001–18. doi:10.24200/SCI.2018.20641. [Google Scholar] [CrossRef]

106. Soleimani-Babakamali MH, Lourentzou I, Sarlo R. Does the curse of dimensionality apply to unsupervised SHM? Investigating the trade-off between loss of information and generalizability to unseen structural conditions. In: Proceedings of the 13th International Workshop on Structural Health Monitoring; 2022 Mar 15–17; Stanford, CA, USA. doi:10.12783/shm2021/36311. [Google Scholar] [CrossRef]

107. Mohammadi-Ghazi R, Büyüköztürk O. Bayesian inference for damage detection in unsupervised structural health monitoring. In: Model validation and uncertainty quantification. Vol. 3. Abingdon, Oxfordshire, UK: Taylor & Francis; 2015. doi:10.1007/978-3-319-15224-0_30. [Google Scholar] [CrossRef]

108. Soleimani-Babakamali MH, Sepasdar R, Nasrollahzadeh K, Sarlo R. A system reliability approach to real-time unsupervised structural health monitoring without prior information. Mech Syst Signal Process. 2022;171(9):108913. doi:10.1016/j.ymssp.2022.108913. [Google Scholar] [CrossRef]

109. Akintunde E, Eftekhar Azam S, Rageh A, Linzell DG. Unsupervised machine learning for robust bridge damage detection: full-scale experimental validation. Eng Struct. 2021;249(2):113250. doi:10.1016/j.engstruct.2021.113250. [Google Scholar] [CrossRef]

110. Ardani S, Akintunde E, Linzell D, Eftekhar Azam S, Alomari Q. Evaluating pod-based unsupervised damage identification using controlled damage propagation of out-of-service bridges. Eng Struct. 2023;286(2):116096. doi:10.1016/j.engstruct.2023.116096. [Google Scholar] [CrossRef]

111. Xiang W, Li X, Zhang FL. Bridge safety state classification based on unsupervised machine learning. In: Proceedings of the 3rd International Civil Engineering and Architecture Conference; 2024 Mar 17–20; Kyoto, Japan. doi:10.1007/978-981-99-6368-3_81. [Google Scholar] [CrossRef]

112. García-Macías E, Ubertini F. Integrated SHM systems: damage detection through unsupervised learning and data fusion. In: Structural health monitoring based on data science techniques. Berlin/Heidelberg, Germany: Springer; 2021. p. 247–68. doi:10.1007/978-3-030-81716-9_12. [Google Scholar] [CrossRef]

113. Akintunde E, Azam SE, Linzell DG. Singular value decomposition and unsupervised machine learning for virtual strain sensing: application to an operational railway bridge. Structures. 2023;58(4):105417. doi:10.1016/j.istruc.2023.105417. [Google Scholar] [CrossRef]

114. Vamvoudakis-Stefanou KJ, Sakellariou JS, Fassois SD. On the use of unsupervised response-only vibration-based damage detection methods for a population of composite structures. In: Proceedings of the 8th European Workshop on Structural Health Monitoring (EWSHM); 2016 Jul 5–8; Bilbao, Spain. [Google Scholar]

115. Perafán-López JC, Sierra-Pérez J. An unsupervised pattern recognition methodology based on factor analysis and a genetic-DBSCAN algorithm to infer operational conditions from strain measurements in structural applications. Chin J Aeronaut. 2021;34(2):165–81. doi:10.1016/j.cja.2020.09.035. [Google Scholar] [CrossRef]

116. Domaneschi M, Cucuzza R, Martinelli L, Noori M, Marano GC. A probabilistic framework for the resilience assessment of transport infrastructure systems via structural health monitoring and control based on a cost function approach. Struct Infrastruct Eng. 2026;22(1):107–19. doi:10.1080/15732479.2024.2318231. [Google Scholar] [CrossRef]

117. Ghazimoghadam S, Hosseinzadeh SAA. A novel unsupervised deep learning approach for vibration-based damage diagnosis using a multi-head self-attention LSTM autoencoder. Measurement. 2024;229:114410. doi:10.1016/j.measurement.2024.114410. [Google Scholar] [CrossRef]

118. Lu Y, Tang L, Chen C, Zhou L, Liu Z, Liu Y, et al. Unsupervised structural damage assessment method based on response correlations. Eng Struct. 2024;302:117413. doi:10.1016/j.engstruct.2023.117413. [Google Scholar] [CrossRef]

119. Zhao P, Xu W, Qi D, Yuan B. Cross-dataset semantic segmentation for composite crack detection using unsupervised transfer learning. Compos Struct. 2025;362:119071. doi:10.1016/j.compstruct.2025.119071. [Google Scholar] [CrossRef]

120. Ye J, Zhang Z, Cheng K, Tan X, Du B, Chen W. Investigation on identification of structural anomalies from polluted data sets using an unsupervised learning method. Front Struct Civ Eng. 2024;18(10):1479–91. doi:10.1007/s11709-024-1065-3. [Google Scholar] [CrossRef]

121. Marasco G, Moldovan I, Figueiredo E, Chiaia B. Unsupervised transfer learning for structural health monitoring of urban pedestrian bridges. J Civ Struct Health Monit. 2024;14(6):1487–503. doi:10.1007/s13349-024-00786-w. [Google Scholar] [CrossRef]

122. Kashyap P, Shivgan K, Patil S, Mahajan S, Banerjee S, Tallur S. TinyML-enabled unsupervised ultrasonic guided wave SHM under varying thermal conditions. In: Proceedings of the 2023 Joint Conference of the European Frequency and Time Forum and IEEE International Frequency Control Symposium (EFTF/IFCS); 2023 May 15–19; Toyama, Japan. doi:10.1109/eftf/ifcs57587.2023.10272087. [Google Scholar] [CrossRef]

123. Junges R, Rastin Z, Lomazzi L, Giglio M, Cadini F. Convolutional autoencoders and CGANs for unsupervised structural damage localization. Mech Syst Signal Process. 2024;220(1):111645. doi:10.1016/j.ymssp.2024.111645. [Google Scholar] [CrossRef]

124. Lomazzi L, Junges R, Giglio M, Cadini F. Unsupervised data-driven method for damage localization using guided waves. Mech Syst Signal Process. 2024;208:111038. doi:10.1016/j.ymssp.2023.111038. [Google Scholar] [CrossRef]

125. Han C, Dong Y, Gao MY, Dong L, Yang Y. Unsupervised anomaly segmentation model for rail damage based on image-inpainting and cold diffusion. Autom Constr. 2025;177(2):106342. doi:10.1016/j.autcon.2025.106342. [Google Scholar] [CrossRef]

126. Tanveer M, Cho S. Unsupervised deep learning method for concrete and asphalt crack segmentation using vision transformer and probability thresholding. Smart Struct Syst. 2025;35(5):267–84. doi:10.12989/sss.2025.35.5.267. [Google Scholar] [CrossRef]

127. Chun PJ, Kikuta T. Self-training with bayesian neural networks and spatial priors for unsupervised domain adaptation in crack segmentation. Comput Aided Civ Infrastruct Eng. 2024;39(17):2642–61. doi:10.1111/mice.13315. [Google Scholar] [CrossRef]

128. Meng S, Gu L, Zhou Y, Jafari A. A label-free high precision automated crack detection method based on unsupervised generative attentional networks and swin-crackformer. Smart Struct Syst. 2024;33(6):449–63. doi:10.12989/sss.2024.33.6.449. [Google Scholar] [CrossRef]

129. Huang B, Kang F, Li X, Zhu S. Underwater dam crack image generation based on unsupervised image-to-image translation. Autom Constr. 2024;163(1):105430. doi:10.1016/j.autcon.2024.105430. [Google Scholar] [CrossRef]

130. Zhang X, Li L, Ren X. Unsupervised local damage diagnosis leveraging substructure isolation technology and time series models. Structures. 2025;77:109122. doi:10.1016/j.istruc.2025.109122. [Google Scholar] [CrossRef]

131. Ayankoso S, Dutta A, He Y, Gu F, Ball A, Pal SK. Performance of vibration and current signals in the fault diagnosis of induction motors using deep learning and machine learning techniques. Struct Health Monit. 2026;25(1):196–212. doi:10.1177/14759217241289874. [Google Scholar] [CrossRef]

132. Omenzetter P, de Lautour OR. Classification of damage in structural systems using time series analysis and supervised and unsupervised pattern recognition techniques. In: Proceedings of the Sensors and Smart Structures Technologies for Civil, Mechanical, and Aerospace Systems 2010; 2010 Mar 8–11; San Diego, CA, USA. doi:10.1117/12.852573. [Google Scholar] [CrossRef]

133. Morais J, Meixedo A, Fortunato E, Ribeiro D, Mendes J. Unsupervised data-driven tool for track support conditions assessment–first results. In: Proceedings of the 9th European Congress on Computational Methods in Applied Sciences and Engineering; 2024 Jun 3–7; Lisbon, Portugal. doi:10.23967/eccomas.2024.163. [Google Scholar] [CrossRef]

134. Entezami A, Sarmadi H, Behkamal B, De Michele C. On continuous health monitoring of bridges under serious environmental variability by an innovative multi-task unsupervised learning method. Struct Infrastruct Eng. 2024;20(12):1975–93. doi:10.1080/15732479.2023.2166538. [Google Scholar] [CrossRef]

135. Sarmadi H, Entezami A. Short-term health monitoring and damage assessment of a concrete cable-stayed bridge by an integrated unsupervised learning approach. In: Damage detection and structural health monitoring of concrete and masonry structures. Berlin/Heidelberg, Germany: Springer; 2025. p. 177–98. doi:10.1007/978-981-97-8975-7_6. [Google Scholar] [CrossRef]

136. Park S, Kim S. Hybrid CAE-DSVDD for unsupervised vibration-based damage detection in in situ steel truss bridge. Struct Health Monit. 2025. doi:10.1177/14759217251369727. [Google Scholar] [CrossRef]

137. Zhang X, Wang Z, Zhao D, Wang J, Li Z. Unsupervised structural damage identification based on covariance matrix and deep clustering. Struct Des Tall Spec Build. 2024;33(11):e2115. doi:10.1002/tal.2115. [Google Scholar] [CrossRef]

138. Mousavi V, Rashidi M, Ghazimoghadam S, Samali B. A data-driven digital twin model for bridge health monitoring using feature fusion and unsupervised deep learning. Inf Fusion. 2026;126:103534. doi:10.1016/j.inffus.2025.103534. [Google Scholar] [CrossRef]

139. Chen C, Tang L, Xiao Q, Zhou L, Wang H, Liu Z, et al. Unsupervised anomaly detection for long-span bridges combining response forecasting by deep learning with Td-MPCA. Structures. 2023;54:1815–30. doi:10.1016/j.istruc.2023.06.033. [Google Scholar] [CrossRef]

140. Zhang M, Guo T, Zhu R, Zong Y, Pan Z. Vibration-based structural health monitoring using CAE-aided unsupervised deep learning. Smart Struct Syst. 2022;30(6):557–69. doi:10.12989/sss.2022.30.6.557. [Google Scholar] [CrossRef]

141. Que Y, Zhong S, Chen K. A score-guided regularization strategy-based unsupervised structural damage detection method. Appl Sci. 2022;12(10):4887. doi:10.3390/app12104887. [Google Scholar] [CrossRef]

142. Yang K, Zhang C, Yang H, Wang L, Kim NH, Harley JB. Improving unsupervised long-term damage detection in an uncontrolled environment through noise-augmentation strategy. Mech Syst Signal Process. 2025;224:112076. doi:10.1016/j.ymssp.2024.112076. [Google Scholar] [CrossRef]

143. Shi S, Du D, Mercan O, Kalkan E, Parol J. Contrastive and self-supervised learning for open-set damage classification in structural health monitoring with incomplete and imbalanced vibration data. Expert Syst Appl. 2025;293:128731. doi:10.1016/j.eswa.2025.128731. [Google Scholar] [CrossRef]