Data Processing and Feature Extraction
Data processing in structural health monitoring (SHM) begins with the acquisition of raw sensor data, which often includes vibration, strain, or acoustic signals contaminated by noise, environmental variations, and operational influences. Preprocessing is essential to enhance signal quality and prepare data for analysis, involving steps such as filtering to remove outliers, normalization to standardize scales, and denoising using techniques like wavelet thresholding or Kalman filtering. These steps mitigate artifacts and ensure reliable input for subsequent stages, as unprocessed data can lead to false positives in damage detection. For instance, in bridge monitoring, Gaussian noise reduction via empirical mode decomposition (EMD) has been shown to improve signal-to-noise ratios significantly in experimental setups.
Feature extraction transforms raw or preprocessed time-series data into a reduced set of damage-sensitive descriptors, enabling efficient pattern recognition while preserving critical information about structural integrity. This process addresses the high dimensionality of sensor networks, where thousands of data points per second are common, by focusing on attributes that change predictably under damage scenarios such as cracks or fatigue. Widely adopted methods draw from signal processing and machine learning, prioritizing features robust to environmental and operational variability (EOV). Seminal work emphasizes that effective features should exhibit low sensitivity to benign changes while amplifying damage indicators, as validated in benchmarks using real-world datasets like the Z24 Bridge, where feature sets reduced data volume substantially without significant loss in detection accuracy.[19]
Time-domain features, extracted directly from signal waveforms, are computationally simple and provide intuitive measures of amplitude and variability. Common examples include root mean square (RMS) value, which quantifies overall energy levels and detects stiffness reductions; kurtosis, sensitive to impulsive damage events like impacts; and crest factor, the ratio of peak to RMS amplitude, useful for identifying nonlinearities in structures. These features perform well in baseline comparisons, achieving high classification accuracies for progressive damage in laboratory beams when combined with statistical thresholds. However, they may overlook frequency-specific changes in non-stationary vibrations.
Frequency-domain features leverage transforms like the fast Fourier transform (FFT) to reveal shifts in spectral content, such as reductions in natural frequencies indicative of mass or stiffness losses. Power spectral density (PSD) estimates energy distribution across frequencies, while spectral centroid tracks the "center of mass" of the spectrum, proving effective for global damage localization in civil structures. In the Z24 Bridge dataset, frequency-based features like dominant peaks distinguished 16 damage states with F1 scores exceeding 85% under forced excitations, outperforming time-domain alone due to their robustness to amplitude variations. Limitations include assumptions of stationarity, which fail in time-varying loads.[19]
Time-frequency domain methods address non-stationary signals by jointly analyzing temporal and spectral evolution, essential for capturing transient damage effects in dynamic environments. The short-time Fourier transform (STFT) divides signals into overlapping windows for localized spectra, though it suffers from fixed resolution trade-offs; improvements via adaptive windowing have enhanced crack detection in beams. Wavelet transforms (WT), particularly continuous (CWT) and discrete (DWT), decompose signals into multi-scale components, extracting features like wavelet energy or coefficients that highlight localized anomalies, such as delaminations in composites, with good noise robustness. Empirical mode decomposition (EMD), introduced by Huang et al., adaptively sifts signals into intrinsic mode functions (IMFs), enabling Hilbert-Huang spectra for instantaneous frequency analysis; applications in bridge SHM have identified modal shifts with high accuracy in noisy conditions. Variants like ensemble EMD (EEMD) and variational mode decomposition (VMD) further mitigate mode mixing, achieving superior performance in bearing fault analogs for structural joints.[20][21]
Advanced feature extraction often integrates modal parameters, derived from output-only identification techniques like stochastic subspace methods, including natural frequencies, damping ratios, and mode shapes. These physics-based features link directly to structural dynamics, with changes in the first few modes signaling 5-10% stiffness losses in real bridges. Statistical features, such as principal component analysis (PCA) projections, reduce multicollinearity in multi-sensor data, retaining substantial variance with fewer components in benchmarks. For selection, wrapper methods like recursive feature elimination (RFE) with random forests have shown optimal subsets yielding high F1 scores on the S101 Bridge dataset, emphasizing spectral time-frequency hybrids over pure domains. Machine learning enhancements, including autoencoders for unsupervised extraction, are increasingly adopted to handle big data from IoT sensors.[19]
This table illustrates representative techniques, prioritizing those with high citation impact in SHM literature. Overall, hybrid approaches combining domains via fusion yield the most robust systems, as demonstrated in population-based SHM frameworks.
Damage Detection and Assessment
Damage detection and assessment in structural health monitoring (SHM) involves systematically identifying the presence, location, type, and severity of structural damage using sensor data and analytical techniques. This process follows a hierarchical framework proposed by Rytter, which includes four levels: Level 1 detects the existence of damage; Level 2 localizes it; Level 3 characterizes the type and extent; and Level 4 predicts remaining service life.[22] Achieving higher levels requires robust data processing to distinguish damage-induced changes from environmental or operational variabilities, such as temperature fluctuations or loading effects.[23]
Vibration-based methods dominate damage detection due to their non-invasive nature and ability to assess global structural integrity. These techniques analyze changes in dynamic properties, like natural frequencies, mode shapes, and damping ratios, which decrease in stiffness when damage occurs. For instance, a reduction in resonant frequency can indicate damage presence, while curvature in mode shapes helps localize it, with severity estimated from the percentage change in frequency (e.g., %Δf = (f_undamaged - f_damaged)/f_undamaged × 100). Traditional parametric approaches, such as modal analysis using output-only methods like Stochastic Subspace Identification, compare baseline models to current responses but are sensitive to noise and environmental factors, limiting accuracy for small damages (e.g., less than 5% stiffness loss).[23][24] Non-parametric methods, including time series modeling with autoregressive moving average (ARMA) models, detect anomalies via statistical residuals without needing a finite element model, offering advantages in real-world applications like bridge monitoring. However, they often struggle with localization beyond presence detection.[24]
Advancements in machine learning (ML) and deep learning (DL) have enhanced vibration-based assessment by automating feature extraction and handling complex data patterns. Supervised ML techniques, such as artificial neural networks (ANNs), classify damage using features like modal parameters or acceleration variances; for example, a multi-layer perceptron (MLP) ANN achieved near-100% accuracy in numerical simulations of truss bridges with simulated stiffness reductions. Support vector machines (SVMs) combined with AR modeling reported errors of 2.6-3.4% in damage localization on laboratory beams. DL methods, particularly convolutional neural networks (CNNs), process raw time-series data directly, eliminating manual feature selection; a 1D-CNN detected damage in 31 scenarios with 100% precision and processed signals 5000 times faster than real-time, demonstrating scalability for online monitoring. These approaches excel in noisy environments but require large labeled datasets for training, posing challenges for rare damage events. Seminal works, including those by Farrar and Worden, emphasize statistical pattern recognition paradigms to validate ML outputs against baselines.[24]
Advanced Techniques
Advanced techniques in structural health monitoring (SHM) leverage artificial intelligence (AI), machine learning (ML), and probabilistic frameworks to enhance damage detection, localization, and prognosis beyond traditional signal processing methods. These approaches address challenges such as environmental variability, data scarcity, and computational complexity by integrating domain knowledge with data-driven models, enabling real-time analysis and predictive maintenance for critical infrastructure like bridges and buildings.[25] Seminal work has established ML paradigms that treat SHM as a multi-level inference process, from data cleaning to reliability assessment, emphasizing the use of big data from sensors to uncover structural patterns.[25]
Machine learning techniques, particularly supervised and unsupervised methods, form the backbone of advanced damage identification. Supervised learning, such as support vector machines and convolutional neural networks (CNNs), excels in classifying damage types like cracks or corrosion from labeled vibration or image data, achieving high accuracy in applications to bridges and aircraft by automating feature extraction and reducing manual inspections. Unsupervised methods, including autoencoders, detect anomalies in unlabeled datasets from buildings or pipelines, offering efficiency for real-time monitoring without prior damage knowledge, though they may produce false alarms in variable conditions. These ML approaches outperform classical methods in handling high-dimensional sensor data, with hybrid models combining supervised and unsupervised elements for robust localization and severity assessment.[26]
Deep learning extends these capabilities through architectures like recurrent neural networks (RNNs) and CNN variants, which process sequential vibration signals or acoustic emissions for precise damage prognosis. In bridge monitoring, deep learning integrated with wireless sensor networks enables automated anomaly detection and predictive modeling, improving structural integrity evaluation by capturing nonlinear patterns that traditional methods miss. For instance, U-Net-based models with self-attention mechanisms segment cracks in images with limited training data via meta-learning, transferring knowledge across damage classes to enhance adaptability.[27] Advantages include scalability to complex structures and reduced computational overhead compared to finite element simulations, though challenges like data quality and model interpretability persist.
Physics-informed machine learning (PIML) represents a high-impact advancement by embedding physical laws, such as partial differential equations (PDEs), into neural networks to overcome data limitations in SHM. Physics-informed neural networks (PINNs), pioneered for solving forward and inverse problems, integrate governing equations directly into loss functions, allowing accurate damage quantification from sparse, noisy sensor data with less training required than pure data-driven models. Applications include crack detection via wave propagation modeling and digital twin creation for real-time seismic response prediction, where PIML achieves improved generalization errors compared to traditional ML.[28] This fusion of physics and data enhances reliability under operational variability, making it widely adopted for civil infrastructure prognosis.[28] Recent developments as of 2025 include integration with edge computing for on-device processing and large language models for interpretive diagnostics in SHM systems.