Source-linked AI summary
A Review of Vibration-Based Damage Detection in Civil Structures: From Traditional Methods to Machine Learning and Deep Learning Applications
Onur Avci, Osama Abdeljaber, Serkan Kiranyaz, Mohammed Hussein, Moncef Gabbouj, Daniel J. Inman
TL;DR
Civil structures need monitoring because damage from environmental and human-induced factors can shorten their service life, while prior reviews did not describe the transition from traditional vibration-based methods to ML and DL. This paper reviews traditional, parametric and nonparametric ML, and DL approaches for vibration-based structural damage detection, emphasizing features and classifiers. The review finds that DL methods, particularly CNNs, can detect and locate damage directly from raw acceleration time histories without manual feature extraction, while 1D CNNs offer lower complexity than 2D CNNs in limited-signal applications.
Problem
Civil structures require ongoing damage monitoring, but prior reviews did not comprehensively describe the transition from traditional vibration-based methods to ML and DL approaches.
Method
The paper comprehensively reviews traditional, parametric and nonparametric ML, and DL vibration-based structural damage detection studies, comparing features, classifiers, structures, excitations, and damage types.
Results
Recent 2D and 1D CNN studies detect and locate damage directly from raw acceleration time histories without data preprocessing or hand-crafted feature extraction.
Takeaways & Limitations
Compact 1D CNNs are preferable for dedicated applications with limited 1D vibration signals because they are easier to train and have lower computational complexity than 2D CNNs.
Takeaways & Limitations
ML-based detection performance depends substantially on selecting features and classifiers, and no fixed combination is guaranteed to suit every structure or damage type.
Abstract
from arXiv · showhide
Monitoring structural damage is extremely important for sustaining and preserving the service life of civil structures. While successful monitoring provides resolute and staunch information on the health, serviceability, integrity and safety of structures; maintaining continuous performance of a structure depends highly on monitoring the occurrence, formation and propagation of damage. Damage may accumulate on structures due to different environmental and human-induced factors. Numerous monitoring and detection approaches have been developed to provide practical means for early warning against structural damage or any type of anomaly. Considerable effort has been put into vibration-based methods, which utilize the vibration response of the monitored structure to assess its condition and identify structural damage. Meanwhile, with emerging computing power and sensing technology in the last decade, Machine Learning (ML) and especially Deep Learning (DL) algorithms have become more feasible and extensively used in vibration-based structural damage detection with elegant performance and often with rigorous accuracy. While there have been multiple review studies published on vibration-based structural damage detection, there has not been a study where the transition from traditional methods to ML and DL methods are described and discussed. This paper aims to fulfill this gap by presenting the highlights of the traditional methods and provide a comprehensive review of the most recent applications of ML and DL algorithms utilized for vibration-based structural damage detection in civil structures.
1. Introduction
Structural damage can accumulate from environmental and human-induced factors, shortening service life and motivating continuous monitoring. This review traces vibration-based damage detection from traditional approaches to ML and DL methods for civil structures.
- Damage and structural monitoring: Damage can accumulate gradually or suddenly from environmental and mechanical factors, reducing structural performance, safety, reliability, and operational life.Corrosion and fatigue typically develop over long periods, whereas earthquakes and fire can produce rapid failure.
- Damage and structural monitoring: Structural Health Monitoring integrates experimental testing, system identification, data acquisition, data management, and long-term environmental and operational measurements.It is applied across mechanical, aerospace, and civil engineering fields.
- Damage detection systems: A typical damage detection system combines sensing and data acquisition hardware with software that processes signals and recognizes patterns reflecting structural condition.Hardware may include accelerometers, velocimeters, strain gauges, load cells, fiber-optic sensors, and acquisition modules.
- Vibration-based damage detection: Civil infrastructure requires continuous monitoring because creep, corrosion, shrinkage, fatigue, and scour shorten structural life, while visual inspection is difficult for large or covered members.Vibration-based systems seek to identify damage presence, location, and severity systematically from structural vibration responses.
- Transition to ML and DL: Machine learning has enabled parametric and nonparametric vibration-based damage detection using tools such as ANNs, SVMs, SOMs, and CNNs.The review notes that improved computational power and chip and sensor technology enabled broader ML use in engineering applications.
- Review scope: The paper reviews traditional, ML, and DL vibration-based damage detection methods, covering feature extraction, classification, and recent applications in civil structures.Its reviewed literature includes 112 journal and conference articles: 58 on traditional methods and 54 on ML and DL methods.
3. Vibration-based structural damage detection methods that are not based on Machine-Learning
Traditional vibration-based structural damage detection methods use measured structural responses to infer condition, and are commonly divided into parametric and nonparametric approaches. Parametric methods estimate dynamic properties, whereas nonparametric methods analyze signal characteristics directly, with applications spanning laboratory and full-scale structures.
- Classification and advantages: Vibration-based SDD methods use structural response to assess overall condition and can operate with limited sensors without knowing damage location beforehand.These methods are categorized as global methods because they use the monitored structure’s vibration response.
- Classification and advantages: Vibration-based techniques are categorized as parametric or nonparametric according to whether they use extracted structural parameters or signal characteristics.The review treats these categories in separate subsections.
- Parametric methods: Parametric methods compare damaged and undamaged dynamic properties, including modal frequencies, damping, stiffness, modal mass, and mode shapes, to indicate damage.Input-output and output-only modal identification methods are used to estimate these parameters.
- Parametric methods: Output-only methods estimate structural behavior from ambient response; examples include CMIF with Random Decrement for flexibility and Random Decrement signatures for damping estimation.These approaches address civil-structure settings where artificial excitation is impractical.
- Applications and interpretation: Variations in modal damping ratio correlated with seismic damage severity and stiffness reduction, while weather and operational effects must be separated from principal-component indicators for accurate SDD.The damping result came from experiments on a large-scale concrete structure; the environmental-effects finding concerned principal components.
- Nonparametric methods: A genetic-algorithm clustering method, GADBA, achieved better classification performance than GMM and MSD on the Z-24 and Tamar Bridge datasets.The review attributes this advantage to genetically guided characteristic selection increasing the chance of approaching a global optimum.
4. Vibration-based structural damage detection methods based on Machine-Learning
Machine-learning vibration-based SDD methods use measured or model-derived modal features with supervised classifiers to detect and localize damage. The reviewed studies show promising results, while performance depends on feature/classifier choices, structure, damage type, and validation scope.
- ML-based vibration SDD methods are supervised approaches that classify extracted structural features for damage detection and localization.The review distinguishes these methods from earlier automatic and unsupervised vibration-based approaches.
- ANN-based parametric SDD: ANN studies use modal characteristics, acceleration-derived features, mode-shape changes, or stiffness-related parameters to identify damage in bridges and building models.Applications include truss bridges, five-story structures, benchmark models, and a steel moment-frame building.
- ANN-based parametric SDD: Bayesian ANN design selects network complexity systematically, including an optimum architecture with 6 hidden neurons for a five-story structure.The study used 32 input–target pairs and examined damage scenarios involving 20% to 80% inter-story stiffness reductions.
- ANN-based parametric SDD: ANN-based methods reported satisfactory or successful detection across benchmark cases, including 10.96% predicted versus 11.31% actual first-story horizontal stiffness reduction.The latter method used modal frequencies and mass-normalized mode-shapes to identify stiffness reductions in two orthogonal directions.
- Alternative ML architectures and classifiers: Other reviewed systems combine ANN with genetic algorithms, Bayesian optimization, GRNN and fuzzy ART, OSELM, or SVMs for feature classification and model updating.Reported findings include successful ANN/GA matching, noise-tolerant GRNNFA operation, and 100% accuracy for RE-based SVM in all examined cases.
- Drawbacks and limitations: ML-based SDD commonly requires hand-crafted features and a suitable classifier, but no feature/classifier combination is guaranteed to work best across structures or damage types.The review notes that researchers often search for effective combinations through trial and error, and broader generalization requires further verification.
5. Vibration-based structural damage detection by Deep-Learning
Deep Learning (DL) methods extend vibration-based structural damage detection by learning representations from raw signals and supporting complex pattern recognition. The reviewed applications include autoencoders and CNNs, with compact 1D CNNs offering accurate damage localization and low computational demands.
- Deep-Learning architectures: Deep Learning networks use multiple hidden processing layers to learn increasingly abstract data representations and handle complex patterns.Unlike shallow ANNs, DL can increase learning ability through greater depth and can learn representations directly from data.
- Deep-Learning architectures: DL architectures reviewed for vibration-based SDD include autoencoders, CNNs, recurrent neural networks, and recursive neural networks.Autoencoders, DBNs, and GANs are grouped under unsupervised pretrained networks; autoencoders are the only UPN architecture identified as used in vibration-based SDD applications.
- Autoencoders: Autoencoder-based SDD learns compact representations and relationships from modal and structural stiffness parameters, while addressing uncertainty from temperature and noise in related work.The reviewed studies include dimensionality reduction, relationship learning, ensemble classification, and experimental validation on an aluminum beam.
- Convolutional Neural Networks: CNNs learn features and classification jointly from raw vibration signals, motivating their use to address hand-crafted feature and classifier limitations in conventional ML-based SDD.CNN applications reviewed in the paper target vibration-based damage detection in civil structures and can process large-scale raw-signal datasets.
- Applications of CNNs: 1D CNNs require lower computational complexity than 2D CNNs, can operate on single CPUs, and are well-suited to real-time applications with scarce training data.Their forward and back-propagation operations use simple array operations, while compact architectures can reduce overfitting risk on limited datasets.
- Applications of CNNs: 5000x faster than the “real-time” requirement was the reported overall SDD speed, with 60 msec processing for 12 acceleration signals in one CNN study.The study reported 5 msec computation for a 300 sec signal and satisfactory prediction across all damage scenarios despite training on samples from only two damage cases.
- Applications of CNNs: 100% of damage scenarios were successfully detected and localized using compact 1D CNNs tested on a laboratory frame with 31 damage cases.An individual CNN processed local data from each accelerometer location, and testing included numerous single- and double-damage cases.
6. Conclusions and recommendations for future work
The review traces vibration-based structural damage detection from traditional and conventional ML methods to deep learning, emphasizing feature extraction, classification, and CNN-based approaches. It identifies practical limitations in labeled data and hand-crafted features while recommending simulated-data and unsupervised-learning strategies.
- Review scope: The review covers non-ML and ML-based parametric and nonparametric vibration-based methods, conventional ML algorithms, and recent 1D and 2D CNN applications.The reviewed studies are compared by extracted features, classification method, test structure, excitation, and damage type.
- Conventional ML: Most ML-based methods perform feature extraction and feature classification, making them more generic than non-ML methods for vibration-based structural damage detection.
- Feature limitations: Modal characteristics are widely used as damage-sensitive features, but temperature, moisture, and low sensitivity to some damage types limit their suitability.The review recommends avoiding modal properties as damage-sensitive features in ML-based detection.
- Feature limitations: Hand-crafted feature and classifier combinations remain structure- and scenario-dependent, while techniques such as model identification, PCA, and AR modeling require substantial computational effort.
- Deep learning: Both 2D and 1D CNNs can detect and locate damage directly from raw acceleration time-histories without preprocessing or hand-crafted feature extraction.The review describes these deep learning approaches as promising for vibration-based structural damage detection.
- Deep learning: 1D CNNs are easier to train and have lower computational complexity than 2D CNNs, making compact 1D models preferable for limited 1D vibration signals in dedicated applications.The review also notes that supervised ML methods require labeled undamaged and damaged data, which are rarely available for civil structures.
- Future work: The review recommends combining real undamaged-structure data with simulated damage data and expanding semi-supervised or unsupervised ML and DL methods for minimally labeled or unlabeled vibration data.Simulated damage data may come from accurate finite-element models or downscaled laboratory models.