From Fourier to Deep Learning: How Bridge Health Monitoring Signals Are Evolving

Bridges live hard lives. They flex under rush-hour traffic, expand and contract with daily temperature swings, sway in the wind, and occasionally endure earthquakes — all of which gradually produce cracking, corrosion, and fatigue. A new comprehensive review published in Advances in Bridge Engineering examines how engineers read these hidden signals, surveying four decades of signal-processing methods used in structural health monitoring (SHM) of bridges, from classical Fourier analysis to modern machine-learning pipelines.

The review, authored by P. Khosravi Hajivand, S. M. M. Emami, A. Safaeimehr, H. Hosseini, and S. Mahboubi and published on 26 June 2026, is not a single experiment but a systematic synthesis. Using a four-stage bibliometric process, the authors searched Scopus and cross-checked Web of Science, initially retrieving roughly 386 publications. After excluding non-journal documents, non-English papers, and studies not directly tied to damage detection or performance monitoring, and after assessing full-text availability and citation quality, the final dataset comprised 201 studies published between 1985 and 2025. Each was categorized in a taxonomy linking sensing methods, signal-processing workflows, and diagnostic models.

The historical arc the authors trace begins well before computers dominated the field. Early SHM grew out of non-destructive evaluation techniques such as ultrasonic testing and impact-echo analysis, which could reveal localized defects but did not scale to long-span structures. A pivotal moment came in 1978, when Adams and colleagues demonstrated that shifts in natural frequencies could indicate stiffness loss, formalizing vibration-based assessment as a central SHM concept. From the 1990s onward, global dynamic parameters — natural frequencies, mode shapes, and damping ratios — became standard indicators of structural condition, aided by advances in wireless sensor networks and MEMS devices that made distributed sensing on large bridges economically feasible.

At the heart of the classical toolkit sits the Fourier Transform, which converts time-domain vibration measurements into frequency-domain spectra from which modal parameters can be extracted. The Fast Fourier Transform, introduced by Cooley and Tukey in 1965, slashed computation time and made frequency-domain analysis practical for real-time monitoring. Yet Fourier methods assume stationarity, an assumption bridges routinely violate as temperatures fluctuate and traffic comes and goes. The Short-Time Fourier Transform patched this partially by windowing the analysis, but the fundamental trade-off between time and frequency resolution remains.

Wavelet-based methods address that weakness with localized basis functions that adapt over both time and frequency, making them well suited to capturing transients — impact responses, crack formation, sudden changes driven by vehicular or seismic excitation. The review distinguishes the Continuous Wavelet Transform, which offers fine time-frequency localization at high computational cost, from the Discrete and Wavelet Packet Transforms, which decompose signals hierarchically for denoising and heightened sensitivity to subtle, distributed damage such as bearing deterioration. The Hilbert–Huang Transform takes a different route: Empirical Mode Decomposition breaks a signal into intrinsic mode functions, and Hilbert Spectral Analysis then yields a time-frequency-energy picture effective for detecting sudden stiffness loss and nonlinear behavior, though the method is noise-sensitive and computationally demanding.

Alongside these transforms, the review catalogs autoregressive (AR) and related parametric models, which express a signal as a linear combination of its past values, with ARX and ARMAX variants adding input-output relationships. Changes in model coefficients or prediction-error patterns can flag early stiffness degradation at low computational cost, but performance suffers under strong nonlinearity and environmental influence. System identification techniques — Frequency Response Functions, Frequency Domain Decomposition, and Stochastic Subspace Identification — extract modal parameters from output-only measurements, a significant advantage when operational traffic, not controlled shakers, provides the excitation. Statistical decomposition methods such as Principal Component Analysis, Independent Component Analysis, and Singular Value Decomposition condense high-dimensional sensor data and help separate damage-related features from traffic, temperature, and wind effects.

The most striking shift in the literature, the authors find, is toward machine learning and artificial intelligence. Support Vector Machines, K-Nearest Neighbors, and Random Forests classify structural states from hand-crafted features, performing reasonably on small and medium datasets. Deep learning goes further: convolutional neural networks excel at image-based inspection of deck cracks and corrosion from UAV imagery, while LSTM recurrent networks model long-term temporal trends in acceleration and temperature data from suspension and cable-stayed bridges. Graph neural networks treat sensors and structural elements as interconnected nodes, capturing load-transfer behavior and spatial damage propagation more naturally than grid-based models. Computer vision paired with these classifiers enables automated, scalable inspections that reduce manual, hazardous field surveys.

Real-world deployments illustrate the promise. Bao and colleagues applied a computer vision and deep learning anomaly-detection framework to acceleration data from a densely instrumented long-span bridge, converting time-series into image-like representations to detect sensor malfunctions and data loss. Tran-Ngoc and colleagues optimized an artificial neural network with the Cuckoo Search algorithm and validated it on the Nam O Railway Bridge in Vietnam, a full-scale steel truss bridge, improving convergence speed, noise robustness, and damage quantification. Barros and colleagues demonstrated an interpretable decision-tree algorithm on a steel truss bridge in Spain, combining vibration measurements with finite-element model updating to distinguish healthy and damaged states with low false-alarm rates.

The bibliometric analysis quantifies the field’s momentum and its blind spots. Roughly 26.5 percent of the reviewed studies appeared between 1990 and 2009, while publications from 2023 to 2025 alone constitute more than 20 percent of the corpus — a surge the authors attribute to cheaper sensing hardware, better computing, and open-source machine-learning libraries. Bridges, buildings, and laboratory specimens dominate the research landscape, accounting for 16.1, 9.7, and 11.6 percent of studies respectively, while dams, offshore platforms, and wind turbines remain under-represented. Steel and concrete structures comprise roughly 58 and 25.7 percent of reviewed work, with composites just over 12 percent, leaving emerging materials poorly covered. In comparative evaluations, hybrid and deep-learning architectures — FFT-DCNN, WT-LSTM, and WT+CNN combinations — achieved over 93 percent accuracy on building tasks and over 98 percent on bridge vibration-pattern detection, exceeding classical methods such as ARMAX models, which reached about 85 percent on building structures. Classical techniques, however, retain advantages in interpretability, computational cost, and performance when data are scarce.

The review closes with a candid assessment of what still blocks field-scale deployment: limited large-scale validation, poor interpretability of machine-learning models, scarce labeled damage data, and the absence of standardized benchmark datasets. The authors propose remedies — a benchmarking protocol with time-based data splitting, environmental stratification by temperature and traffic, cross-site validation on different bridges, and recommended metrics including damage-detection accuracy, false-alarm rate, localization error, and a cross-structure transferability score. They also call for open code, documented preprocessing pipelines, and a unified public repository of long-duration, multi-sensor, damage-labeled bridge data. Keyword co-occurrence analysis points to where the field is heading: fiber-optic sensing, vision-based inspection, and hybrid data-fusion schemes are emerging themes, signaling a gradual move from isolated experiments toward integrated, networked monitoring — and, ultimately, from diagnostic tools toward proactive decision-support systems for the world’s aging bridge stock.

Subject of Research: Signal processing methods for structural health monitoring of bridges

Article Title: Signal processing methods for structural health monitoring of bridges: a comprehensive review of classical, data-driven, and hybrid approaches

Article References: Hajivand, P. K., Emami, S. M. M., Safaeimehr, A., Hosseini, H., & Mahboubi, S. (2026). Signal processing methods for structural health monitoring of bridges: a comprehensive review of classical, data-driven, and hybrid approaches. Advances in Bridge Engineering, 7(1), Article 20. https://doi.org/10.1186/s43251-026-00214-7

Image Credits: AI Generated

DOI: 10.1186/s43251-026-00214-7

Keywords: structural health monitoring, bridges, signal processing, Fourier transform, wavelet analysis, machine learning, deep learning, damage detection, system identification, computer vision, sensor networks, bibliometric review