Structural Health Monitoring
Structural health monitoring instruments a structure so that its condition can be inferred from measurements rather than from a scheduled visual inspection. A bridge, an airframe, a wind turbine blade, or a pipeline carries a permanent network of sensors, and the system reading them attempts to answer one question: has anything changed, and if so, where, how badly, and what follows. The premise is that damage alters how a structure responds to loading, and that a permanent instrument observes that alteration earlier and far more consistently than a person with a flashlight on a five-year cycle.
The premise is sound and the practice is much harder than it sounds. Sensors do not measure damage. They measure strain, acceleration, wavelength, potential, or acoustic energy, and every one of those quantities also responds to temperature, moisture, traffic, and wind. On a real structure the response change caused by a warm afternoon routinely exceeds the change caused by a crack an engineer would want to know about. Separating one from the other is not a detail of the discipline; it is the discipline.
This article sits under energy harvesting because power, more than sensor performance, shapes what a monitoring system can be. Sensors belong where damage initiates, and damage initiates at joints, splices, embedments, and roots — places inconvenient to reach and rarely near a supply. Cabling them costs more than the instruments, and the result must survive a service life measured in decades. Harvested power, buffered storage, duty cycling, and on-node computation are therefore not refinements added at the end of the design. They determine which sensing modalities are usable at all.
What Structural Health Monitoring Sets Out to Do
The field organizes its ambitions into a hierarchy standard since Anders Rytter proposed it in his 1993 doctoral thesis at Aalborg University. Level one asks whether damage is present. Level two asks where it is. Level three asks how severe it is. Level four asks what remains of the structure's useful life. Each level demands strictly more information than the last, and the gaps are wider than the numbering suggests.
Level one needs only a healthy baseline and the detection of a statistically significant departure from it. Level two needs spatial resolution, meaning either many sensors or a feature that carries positional information. Level three ordinarily needs a validated physical model, because converting an observed change into a stiffness loss of stated magnitude is an inverse problem that measurements alone do not close. Level four inherits every uncertainty of level three and adds a damage growth law and a forecast of future loading, and the load forecast is frequently the weakest of the three.
Charles Farrar and Keith Worden reframed the activity as statistical pattern recognition in four stages: operational evaluation, data acquisition and cleansing, feature extraction, and statistical model development for discrimination. Their formulation carries an axiom worth stating plainly. Sensors measure the response of a structure, not its damage; every diagnosis is a feature extracted from that response plus a statistical decision about it. Damage is also defined relative to a reference state, so the baseline is definitional rather than convenient.
Methods divide into global and local families. Global methods interrogate the whole structure through its low-frequency dynamics and are sensitive to anything, anywhere, but only weakly. Local methods use high-frequency ultrasound, electrical impedance, or electrochemistry to find small defects reliably within a few meters. The smallest detectable defect scales with the wavelength used to find it, so covering a whole bridge and resolving millimeter cracks are mutually exclusive goals for any one method.
Strain, Displacement, and Inclination Sensing
Strain is the most direct measure of structural action, because it reports load in the member being measured rather than the behavior of the structure as a whole. The sensor families in routine use differ less in accuracy than in what they are good for over time.
Resistive Foil and Vibrating Wire Gauges
A resistive foil gauge is a serpentine metal pattern bonded to the surface whose resistance changes in proportion to strain. The gauge factor is close to 2 for the constantan and Karma alloys used in practice, so one microstrain changes a 350-ohm gauge by about 0.7 milliohm. Resolving that needs a Wheatstone bridge and an amplifier working at microvolt levels, which makes the measurement sensitive to lead resistance, thermoelectric junctions, and offset drift. Self-temperature-compensated alloys and dummy gauges in adjacent bridge arms are the standard defenses.
A vibrating wire gauge holds a tensioned steel wire between anchors inside a sealed tube. A coil plucks the wire and reads its decay; the resonant frequency follows the tension, and strain follows the square of that frequency, so the output is a frequency rather than a voltage. That property explains the technology's dominance in embedded civil instrumentation, because a frequency survives hundreds of meters of cable, splices, and moderate insulation degradation unchanged. Gauges cast into concrete have produced credible readings decades after placement, and most carry an integral thermistor. The cost is speed: a reading takes about a second, so these instruments measure quasi-static strain, thermal movement, prestress loss, and settlement, and are useless for dynamics.
Displacement and Inclination
Displacement instruments answer questions strain cannot. Crackmeters and joint meters, typically linear variable differential transformers or draw-wire sensors bridging a gap, report directly on expansion joint travel, bearing seizure, and the opening of a known crack under load. These are among the most diagnostically valuable channels on a bridge precisely because they are unambiguous: a joint that has stopped moving with temperature has seized, and no statistical model is needed to say so.
Absolute displacement of a span is harder, because nothing stationary is nearby to measure against. Satellite receivers in real-time kinematic mode give centimeter-level absolute position, adequate for a long suspension span but not for a stiff girder. Tiltmeters offer the other route: inclinometers resolving fractions of a milliradian, chained along a span and integrated spatially into a deflected shape with no external reference. Their limitation is that thermal offset drift inside the sensing element is indistinguishable from a genuine rotation.
Fiber Optic Sensing
Optical fiber occupies a distinct position because it removes electricity from the sensing element entirely. A fiber carries no current, creates no ground loop, is immune to electromagnetic interference and lightning-induced transients, tolerates high temperature, and is safe in explosive atmospheres. On a tall structure, a substation, a fuel tank, or a pipeline right of way those properties are decisive rather than incidental.
Fiber Bragg Gratings
A fiber Bragg grating is a periodic modulation of refractive index written into the fiber core, a few millimeters long, which reflects a narrow band centered on a wavelength set by twice the effective index times the grating period. Strain stretches the period and temperature changes both period and index, so the reflected wavelength shifts with each. Near the 1550-nanometer band the strain response is on the order of 1.2 picometers per microstrain and the thermal response on the order of 10 picometers per degree Celsius, so installations pair every sensing grating with a companion held mechanically free nearby.
Multiplexing is the decisive advantage. Gratings written at different nominal wavelengths along one fiber occupy different parts of the spectrum, so a single interrogator and a single fiber read many points in series. The count is the interrogator's usable spectral range divided by the window each grating needs, and arrays of a few tens of gratings on one fiber are routine. On a large structure one armored fiber replaces dozens of shielded pairs.
Three limitations deserve equal weight. The interrogator is a precision optical instrument consuming watts, exactly the wrong shape for a harvested node, which confines Bragg systems to installations with mains power or a substantial solar array. Fiber is mechanically vulnerable at terminations and connectors, and field splicing needs equipment and skill that are not universally available. A break anywhere silences every grating beyond it unless the array can be interrogated from both ends.
Distributed Fiber Sensing
Distributed sensing dispenses with discrete gratings and treats the fiber itself as a continuous sensor by analyzing backscattered light. Raman scattering yields temperature; Brillouin scattering yields strain and temperature together; phase-sensitive Rayleigh backscatter, marketed as distributed acoustic sensing, yields dynamic strain and therefore the acoustic field along the fiber. Brillouin optical time-domain analysis, using a counter-propagating pump and probe, reaches tens of kilometers of range with spatial resolution of a meter or so in commercial instruments and down to centimeters in laboratory work that trades range for resolution.
For a linear asset this rewrites the economics. A pipeline, tunnel lining, levee, or rail alignment is instrumented continuously by laying one cable, with a single interrogator serving the whole length, so the cost per monitored meter falls toward the cost of cable and installation. The trade is that the interrogator is a large, powered, environmentally protected instrument in a cabinet: distributed sensing suits assets with a serviceable end point and suits isolated harvested nodes not at all.
Accelerometers and the Low-Frequency Noise Floor
Dynamic monitoring lives or dies on the accelerometer, and the criterion is narrower than most datasheets suggest: what matters is the noise floor within the band containing the modes of interest. A short footbridge may have a first vertical mode at several hertz, but a long-span suspension bridge has fundamental modes below one hertz, and the lowest modes of the largest spans approach a tenth of a hertz, where ambient excitation produces accelerations of micro-g to milli-g. An instrument that resolves a milli-g comfortably records nothing but its own noise there.
Piezoelectric Instruments
Seismic piezoelectric accelerometers, usually built to the integrated-electronics piezoelectric convention with sensitivities around 10 volts per gravity, achieve broadband resolution near one micro-g and serve as the reference instruments of the field. Their limitation is intrinsic: the element produces charge in response to strain, and that charge leaks through the amplifier input impedance, so the response rolls off at low frequency and there is none at all to a constant acceleration. Power is the other cost. A constant-current supply of 2 to 20 milliamperes at 18 to 30 volts, drawn whenever the sensor is on, is tens of milliwatts per channel, two to four orders of magnitude above what a harvester sustains.
Micro-Electromechanical Instruments
Capacitive micro-electromechanical accelerometers respond down to direct current, occupy a few square millimeters, and draw microamperes to a few milliamperes. They also measure static inclination, so one device serves both the dynamic and the tilt channels. Historically their noise density disqualified them from civil work, since general-purpose consumer parts sit in the hundreds of micro-g per root hertz or worse and bury the ambient response of a large structure. Devices designed for precision inertial and monitoring use have narrowed the gap, with the better commercial parts now specified in the region of a few tens of micro-g per root hertz.
Whether that suffices is settled by an arithmetic every designer should perform explicitly. Noise density integrates over the analysis bandwidth, so the noise contribution is the density multiplied by the square root of the bandwidth actually used. A device at 25 micro-g per root hertz, analyzed in a band one-tenth of a hertz wide around a mode, contributes roughly 8 micro-g of noise to that estimate. If wind and traffic excite the mode to tens of micro-g the measurement is usable; if they excite it to a few, it is not.
The lever is bandwidth, and bandwidth is bought with record length. Longer records narrow the effective analysis band and pull the noise down, provided the structure and its environment stay stationary throughout. That is unusually convenient for harvested systems, because a long record at a modest sample rate costs far less energy than a short record at a high one.
Force-balance servo accelerometers fill the remaining niche, holding the proof mass stationary in a closed loop and reporting the restoring current, which gives direct-current response with sub-micro-g resolution at the price of size, cost, and steady power. Whatever the technology, offset drift with temperature deserves as much attention as noise, because a slow offset change is indistinguishable from a slow tilt and will be reported as one.
Acoustic Emission
Acoustic emission monitoring listens rather than interrogates. When a crack advances, a fiber breaks, a bond delaminates, or a prestressing wire fails, the sudden release of stored elastic energy launches a transient stress wave. Piezoelectric sensors coupled to the surface — commonly resonant types centered near 150 kilohertz for metals, and lower frequencies for concrete, where attenuation is far more severe — detect it as a short burst above the ambient noise floor.
What distinguishes acoustic emission from every other method here is that it observes the event of damage growth rather than the accumulated state of damage. That is an advantage and a limitation in equal measure. It catches active deterioration at the moment it occurs, which no periodic inspection can do, and it is entirely blind to a defect that is not currently growing. A large dormant crack emits nothing.
Source location comes from arrival-time differences across an array, which requires propagation velocity and geometry to be known, and attenuation limits sensor spacing to a few meters in concrete and tens of meters in favorable steel plate. The dominant practical difficulty, though, is discrimination. A bridge in service produces emission-like signals from traffic, rain, thermal ticking, bearing fretting, and loose components rubbing. Separating genuine crack emission from that background is approached by classification on hit parameters such as amplitude, rise time, duration, and energy, or by full waveform analysis, often with guard sensors placed to veto known noise sources.
Acoustic emission is the most demanding modality on the power budget, because meaningful monitoring means listening continuously at hundreds of kilohertz and no harvested node can digitize continuously at that rate. The standard architecture is an analog threshold detector, a low-power comparator watching the conditioned sensor output, that consumes microwatts, holds the digitizer and processor asleep indefinitely, and wakes them within microseconds when a hit exceeds the threshold. Setting that threshold is the whole design: too low and the node spends its energy budget on rain, too high and it sleeps through the crack.
Guided Waves and Electromechanical Impedance
Where acoustic emission listens, guided wave methods transmit. Both use the same piezoelectric materials and both target the local end of the coverage spectrum, but they answer different questions.
Guided Ultrasonic Waves
In a plate, shell, or pipe wall the boundaries guide ultrasound into modes, Lamb modes in a plate and longitudinal, torsional, and flexural modes in a pipe, that travel long distances with modest attenuation. That is the appeal: a sparse array interrogates a large area rather than the small footprint a conventional probe covers. Bonded piezoelectric wafer transducers, thin lead zirconate titanate discs several millimeters across, both transmit and receive at the tens to hundreds of kilohertz where the useful modes live. Pitch-catch compares a signal against a stored baseline, pulse-echo listens for reflections, and sparse-array imaging combines many paths into a map of scattering strength.
The complications explain why deployment has lagged the coverage advantage. Multiple dispersive modes propagate at any frequency, so a transmitted pulse arrives spread out and interleaved. Temperature changes wave speed enough to break baseline subtraction outright, making compensation by baseline signal stretch, or by selecting the closest match from a library of baselines, mandatory. The adhesive bond degrades, and a debonding transducer produces a change indistinguishable from damage. Real structures are also full of fasteners, stiffeners, and thickness changes that scatter as strongly as any crack, which is why the method works best on well-characterized hot spots.
Electromechanical Impedance
A piezoelectric wafer bonded to a structure has an electrical impedance that depends on the mechanical impedance beneath it, because the wafer's motion is constrained by what it is attached to. Sweeping the drive frequency from tens to a few hundred kilohertz produces a spectrum of sharp features corresponding to local resonances, and nearby damage such as a loosening bolt or a debonding repair patch shifts and reshapes them. Comparison against a baseline reduces to a scalar such as the root-mean-square deviation of the spectrum. The high frequency buys sensitivity to very small damage and costs coverage, confining the sensing region to a radius of centimeters.
Electromechanical impedance is unusually friendly to constrained power budgets. The measurement is a low-voltage swept impedance measurement over a narrow band, performed in milliseconds, and integrated impedance converter devices paired with a small microcontroller carry it out without a laboratory analyzer. A short sweep every hour or every day, yielding tens of bytes, fits comfortably inside a harvested budget, which is why the method appears so often in battery-free bolt-loosening and patch-monitoring demonstrations.
Corrosion and Environmental Sensors
For reinforced and prestressed concrete the dominant deterioration mechanism is not mechanical overload but corrosion of embedded steel, driven by chloride ingress from de-icing salt or seawater and by carbonation of the cover concrete. Corrosion monitoring addresses the cause rather than waiting for the mechanical consequence, and it does so with electrochemistry rather than dynamics.
Half-cell potential measurement compares the potential of the reinforcement against an embedded reference electrode, commonly silver-silver chloride or manganese dioxide; ASTM C876 provides the standard framework, with more negative potentials indicating a higher probability of active corrosion. Linear polarization resistance estimates an instantaneous corrosion rate from a small potential perturbation. Anode-ladder probes place steel elements at graded depths through the cover, so successive elements depassivate as the chloride front advances. In pipelines and tanks, electrical resistance probes and mass-loss coupons serve the same purpose internally.
Environmental instrumentation is a prerequisite for interpreting the structural channels rather than decoration around them. A useful bridge system records air temperature, structure temperature at several depths through a deck or girder section, relative humidity, wind speed and direction, and at least a traffic count. The temperature gradient through a section matters more than the mean, because a gradient produces curvature and therefore strain and deflection that mimic load. On many well-designed installations the environmental channels outnumber the mechanical ones, and that ratio is a sign of designers who understood the problem rather than of misplaced priorities.
Operational Modal Analysis
Classical experimental modal analysis measures both excitation and response, computes frequency response functions, and extracts natural frequencies, damping ratios, and mass-normalized mode shapes. It is the right method in a laboratory and the wrong method in the field. A portable shaker cannot meaningfully excite a long-span bridge, a drop-weight test on a highway structure requires closing it, and closing the structure defeats the purpose of continuous monitoring. Controlled excitation is equally impractical on an operating wind turbine or an aircraft in service.
Operational modal analysis solves the problem by discarding the input. It assumes ambient excitation — wind, traffic, wave action, micro-tremor, machinery, rotor loads — is broadband and approximately stationary across the band of interest, so its spectrum is roughly flat there and does not color the identified parameters. Under that assumption, natural frequencies, damping ratios, and mode shapes follow from the response alone. Mode shapes emerge unscaled, because without knowing the input magnitude there is no way to normalize by mass, which limits some downstream uses unless a mass-perturbation test or a model supplies the scaling.
Several identification methods are in general use. Peak picking on averaged output spectra is crude and assumes well-separated lightly damped modes, and remains popular because it is robust. Frequency domain decomposition takes a singular value decomposition of the output spectral density matrix at each frequency line, separating closely spaced modes that peak picking merges. Stochastic subspace identification fits a discrete-time stochastic state-space model directly to the measured series and is the workhorse of automated permanent systems. The natural excitation technique with the eigensystem realization algorithm converts cross-correlations into functions proportional to free decays.
Automation is where field systems succeed or fail. A permanent installation identifies modes thousands of times a year with nobody watching, so the model order selection an engineer would do by eye must be replaced by clustering and consistency rules that separate physical modes from spurious mathematical ones. Harmonic excitation from rotating machinery appears as a very sharp line with implausibly small apparent damping, and on a wind turbine, where rotor harmonics sweep across the structural modes, rejecting them is a first-order design problem.
One measured quantity deserves a warning. Damping is far less repeatable than frequency, commonly scattering by tens of percent between consecutive records on the same structure because of amplitude dependence, method, and record length. The intuition that cracking dissipates energy and should raise damping is physically reasonable and practically unusable on its own, because the natural scatter swamps the effect.
Damage-Sensitive Features and Statistical Classification
A feature is any quantity computed from measurements that changes when damage appears and stays reasonably constant otherwise. The choice determines everything downstream and is where the field's ingenuity has concentrated.
Natural frequencies are the most used and the most disappointing. They are global, easy to identify accurately, and need few sensors. They are also insensitive, because a frequency responds to a local stiffness loss in proportion to the strain energy that particular mode stores at that particular location, so a crack at a modal node is invisible to that mode entirely. Shifts well under one percent are routine for damage an engineer would call serious, which is exactly why the environmental confounding described below is fatal rather than merely inconvenient.
Mode shapes carry spatial information and support localization. The modal assurance criterion compares whole shapes between records, and its coordinate-wise variant flags the locations contributing most to a mismatch. Mode shape curvature localizes better still, since bending curvature relates directly to local flexural stiffness, but obtaining it means differentiating a spatially sampled shape twice, which amplifies noise dramatically and demands dense instrumentation. Flexibility matrices assembled from identified modes converge faster with the number of measured modes than the modal parameters themselves, so changes in assembled flexibility localize damage with fewer modes than curvature methods need.
Features requiring no modal identification at all suit autonomous nodes. Fitting an autoregressive model to the acceleration series at each sensor and monitoring its coefficients, or the residual when a model trained on healthy data meets new data, condenses a long record into a handful of numbers within a few kilobytes of memory.
Model-Based and Data-Driven Approaches
The model-based route builds a finite element representation, updates its parameters until predicted modal properties match those measured on the healthy structure, and explains later measurements by the pattern of stiffness reduction that best accounts for them. The answer is physical and feeds directly into structural assessment. The difficulties are those of any inverse problem: candidate damage parameters outnumber independent measurements, many patterns explain the same data equally well, and model errors such as boundary condition stiffness are exactly the parameters updating adjusts, so they are readily absorbed as fictitious damage.
The data-driven route learns healthy behavior statistically and flags departures. Because damaged data essentially never exist for a real structure, the honest formulation is novelty detection rather than classification. Mahalanobis distance against the healthy feature distribution, one-class support vector machines, autoencoders judged by reconstruction error, and Gaussian process models with predictive variance all serve, with the threshold set from the healthy distribution to achieve a stated false-alarm rate.
The rare cases where damaged data do exist are correspondingly influential. Structures already condemned have been instrumented and then progressively damaged before demolition, notably the I-40 bridge over the Rio Grande in Albuquerque and the Z24 bridge in Switzerland, and those data sets have been reanalyzed for decades because so little else exists. Population-based structural health monitoring instead transfers knowledge across a fleet of nominally similar structures.
Environmental and Operational Variability
This is the central difficulty of the discipline and deserves stating without hedging. On a real structure in a real climate, temperature routinely moves natural frequencies more than structurally significant damage does. Any scheme comparing a frequency today against a frequency last year, without accounting for the conditions under which each was measured, will produce false alarms in winter and miss damage in summer.
The Z24 bridge remains the canonical evidence. It was a concrete highway bridge in Switzerland, monitored continuously for a year and then subjected to a sequence of deliberate damage scenarios before demolition in 1998. Published analyses of that data set report the first vertical bending frequency rising by roughly fifteen percent during cold periods, with a torsional mode rising considerably more, and a distinctly bilinear relation between frequency and temperature whose kink sits at the freezing point. The mechanism is the asphalt deck surfacing, which contributes almost nothing to stiffness when warm and a great deal when frozen. Several of the deliberate damage scenarios introduced later produced smaller frequency changes than the winter had produced by itself.
The confounding mechanisms are numerous and structure-specific. The modulus of asphalt, and to a lesser degree of concrete, depends on temperature. Thermal expansion closes gaps and locks bearings that were free at another temperature, and freezing soil around a foundation raises every frequency at once. Traffic mass lowers the frequencies of a light structure measurably. On an aircraft, fuel state, payload, altitude, and airspeed shift modal properties; on a wind turbine, rotor speed, pitch, yaw, and ice accretion do the same.
Compensation Strategies
The simplest and most interpretable approach measures the confounder and regresses the feature against it. A model relating identified frequency to measured temperature, usually with lag terms because a concrete deck's thermal state trails air temperature by hours, is fitted on baseline data, and the residual after subtracting the prediction becomes the monitored quantity. This requires the confounder to be measured, the relationship to be identified across the full range of conditions, and therefore at least one complete annual cycle of baseline before alarms carry meaning. That last requirement is unwelcome to project schedules and is nonetheless real.
Latent-variable methods dispense with environmental sensors. Principal component analysis on a vector of features across many records finds the directions of greatest variance, which in healthy data are dominated by environmental variation because that is what drives the healthy structure. Projecting out the leading components leaves a residual subspace in which environmental effects are largely suppressed. The attraction is that no confounder need be measured or even identified; the risk is symmetrical, since any damage effect aligned with the discarded components is discarded with the environment.
Cointegration, borrowed from econometrics, is a more principled version of the same idea: feature series that are individually nonstationary because they follow a common environmental trend may have a linear combination that is stationary, cancelling the shared trend by construction. Operational variability is often better handled by conditioning than compensation, restricting comparisons to records collected within a narrow band of speed, pitch, and yaw, which discards most of the data and is honest about what it can compare.
Sensor Placement, Networks, and Long Life
Sensor Placement Optimization
Where to put a limited number of sensors is a formal optimization problem with established answers. The effective independence method, introduced by Daniel Kammer, iteratively discards candidate locations contributing least to the linear independence of a set of target mode shapes. Modal kinetic energy weighting biases placement toward locations where those modes carry most energy, and information-theoretic criteria maximize expected information gain. All require a model and a nominated set of target modes, so placement embeds an assumption about what damage will look like before any has occurred, which is why the analytical result is combined with judgment about known hot spots.
Wired Against Wireless, and Synchronization
A wired system with a central acquisition chassis gives synchronization for free, imposes no energy limit, and supports high channel counts and sample rates. Its drawback is installation: conduit, cable trays, traffic management, and labor commonly exceed the instrumentation cost several times over. Wireless removes the cable and pays in four currencies: synchronization, bandwidth, reliability, and energy. For quasi-static monitoring at intervals of minutes it is straightforwardly better; for synchronized dynamic monitoring at high channel counts, wired systems still dominate major installations.
Mode shape estimation depends entirely on relative phase between channels, so sampling must be aligned to a small fraction of the period of the highest mode of interest, which for civil structures means an error budget on the order of a tenth of a millisecond. Free-running crystal oscillators drift by tens of parts per million, accumulating milliseconds within a minute, so nodes must synchronize before or during each burst. The remedies are beacon-based protocols in the tradition of the flooding time synchronization protocol, post-facto correction fitting a linear clock model across a timestamped record, satellite-disciplined timing, and the IEEE 1588 precision time protocol on wired segments.
Reduction at the Node Against Streaming
A single triaxial accelerometer sampled at two hundred samples per second per axis with 16-bit resolution generates roughly seventy thousand bytes per minute, so a twenty-node network streaming continuously produces on the order of two gigabytes per day. Computing on the node changes the arithmetic entirely: a windowed spectral estimate, an autoregressive fit, or a local modal identification reduces a multi-minute burst to tens or hundreds of bytes. The cost is loss of the raw record, which cannot be reanalyzed when a better method appears or an anomaly needs forensic examination.
Calibration Drift Over Decades
A system commissioned with a bridge is expected to outlive several generations of the electronics inside it. Amplifier offsets and sensor sensitivities drift, adhesives creep and absorb moisture, cable insulation embrittles, connectors corrode, reference electrodes are consumed by the measurements they enable, and radio standards and software platforms become unsupportable long before the structure does. A slow drift is more dangerous than an outright failure, because a dead channel is obvious while a drifting one quietly poisons a trend.
The countermeasures that work are unglamorous. Prefer sensors whose output is intrinsically stable, such as a vibrating wire frequency or a Bragg wavelength, over those depending on long-term amplifier stability. Provide on-board reference channels and a periodic self-test with a known stimulus. Install redundant sensors so a drifting channel can be identified against its neighbors. And design on the assumption that the electronics will be replaced at least once, which argues for accessible enclosures and documented interfaces.
Powering a Monitoring Network
Sensors belong where damage initiates, and damage initiates where power does not reach. That single observation explains why energy harvesting and structural health monitoring developed together, and why the power architecture governs the sensing choices rather than following from them.
Available sources depend on the asset. Solar is the highest-yield option by a wide margin on any outdoor structure with a mounting surface, and the design question is survival through the darkest week rather than the annual average. Vibration harvesting suits structures that move, but requires tuning a resonant transducer to a modal frequency that must be measured rather than assumed and that drifts as the structure ages. Thermoelectric generators exploit the gradient between a sun-warmed and a shaded surface and reliably deliver less than the ambient record suggests, because mounting and heat sink consume most of the difference.
The honest yield across these sources is microwatts to a few milliwatts on average, punctuated by droughts, while the instantaneous demand of a node that is measuring and transmitting is orders of magnitude higher. Every design is therefore an exercise in duty cycling around a buffer sized for the longest credible interruption. The typical architecture sleeps at sub-microampere levels, wakes on a schedule for a synchronized burst, and wakes asynchronously when an analog threshold detector reports an event. Those detectors deserve emphasis, because they make event-driven modalities feasible at all: a comparator watching a piezoelectric sensor consumes microwatts, holds a digitizer off for months, and responds within microseconds.
The consequence that shapes everything else is this. Harvested power almost always forces the processing onto the node, because transmitting a byte over a low-power radio costs far more energy than a microcontroller spends on many arithmetic operations. That is not a stylistic preference but an inequality between two energy costs, and it explains why fiber Bragg systems, excellent sensors by every other measure, sit awkwardly in harvested deployments: the interrogator is a watts-class instrument that cannot be duty-cycled down to a microwatt average.
Two escapes are worth naming. Passive interrogation builds the sensor as a wirelessly powered device, embedded in concrete or bonded under a doubler and read by an interrogator brought to the asset, which removes the storage problem at the cost of requiring a visit. The second is to admit that harvesting need not be the sole supply: many deployed nodes run on primary lithium cells sized for several years, with a harvester and a supercapacitor extending life rather than replacing the cell.
Applications
Civil Infrastructure
Long-span bridges are the flagship application, and the largest carry instrumentation installed during construction and operated continuously since opening. Hong Kong's Tsing Ma Bridge, in service since 1997, has been monitored from the outset by a wind and structural health monitoring system spanning hundreds of channels of accelerometers, strain gauges, anemometers, temperature sensors, displacement transducers, and satellite positioning receivers. The replacement I-35W St. Anthony Falls Bridge in Minneapolis, opened in 2008, was instrumented during construction with several hundred embedded sensors including vibrating wire strain gauges, thermistors, and accelerometers.
Beyond the flagships, the applications with the clearest value are narrower. Weigh-in-motion instrumentation turns a bridge into an instrument that measures the traffic loading it, the single largest uncertainty in a fatigue assessment. Post-earthquake instrumentation of buildings answers within hours whether a structure may be reoccupied. Scour monitoring at piers addresses a failure mode that inspection of the superstructure cannot see at all, since the erosion happens underwater.
Aerospace Airframes
Aircraft structures face fatigue cracking and corrosion in metallic assemblies and impact damage, disbonding, and delamination in composites. The physics of detecting these is comparatively well understood; the barrier is certification. Any sensor whose output substitutes for a scheduled inspection becomes part of the approved maintenance program and must be qualified accordingly, including a demonstrated probability of detection and an understanding of what happens when the sensor itself fails.
Progress has therefore concentrated on narrowly defined hot spots. Comparative vacuum monitoring, which detects a crack by loss of vacuum in a fine gallery pattern bonded across a critical detail, went through a long qualification campaign involving Sandia National Laboratories fatigue testing and airline trials, and reporting in the aerospace maintenance literature indicates that the Federal Aviation Administration issued a supplemental type certificate in 2022 covering its use on a Boeing 737-800 structural inspection, described as the first approval of its kind for a crack-detecting monitoring sensor. Load monitoring is more widely deployed because it supplements rather than replaces inspection: strain gauge systems such as the life-time monitoring system fitted to the Airbus A400M sharpen individual aircraft fatigue tracking without changing any inspection interval.
Wind Turbine Blades and Pipelines
A wind turbine blade is a large composite structure loaded tens of millions of times, inspected by rope access or unmanned aircraft, and ruinously expensive to repair offshore. Fiber Bragg arrays embedded during layup measure root and spanwise strain. Accelerometers track blade natural frequencies, which shift with degradation but also with rotor speed through centrifugal stiffening, with pitch angle, and with ice accretion, making the blade a textbook case of operational variability dominating the signal. Acoustic emission detects bond-line disbonds and is standard in full-scale blade testing under IEC 61400-23.
Pipelines are the clearest case for distributed fiber, because the asset is linear, the fiber can be laid in the same trench, and one interrogator covers tens of kilometers. Brillouin scattering measures strain along the pipe, detecting the ground movement and subsidence that place a line at risk long before it leaks. Raman scattering measures temperature, which reveals a leak by the thermal signature of escaping product. Phase-sensitive Rayleigh backscatter detects both the acoustic signature of a leak and the more common threat of an excavator working near the right of way.
What the Field Has and Has Not Delivered
Load and usage monitoring works, and it is the least glamorous and most valuable result the field has produced. A structure recording its own load history removes the single largest uncertainty from a remaining-life calculation, because the assumed load spectrum in a design fatigue analysis is invariably conservative and invariably wrong in detail. Aircraft, ships, offshore platforms, and a growing number of bridges do this in service today, and the benefit is realized through a revised inspection interval or an extended certified life rather than through any damage diagnosis at all.
Post-event assessment works. After an earthquake, vessel impact, or overload, an instrumented structure answers whether anything changed far better and faster than an inspector can. Local non-destructive evaluation made permanent works too, in narrowly defined hot spots where the damage location is known in advance, which is exactly where aviation certification has landed: when the location is known, a probability of detection can be established and the statistical problem becomes tractable.
What has not been delivered is the original ambition: unattended, model-free, whole-structure damage quantification and remaining-life prognosis on ordinary civil structures. The obstacles are the ones named throughout this article. Confounding environmental and operational variability produces changes of the same magnitude as the damage being sought. Damaged training data do not exist for the specific structure of interest. Global dynamic features are intrinsically insensitive to local damage, and the local methods that are sensitive cover only small areas. None of these is a temporary limitation awaiting better electronics; they are properties of the physics and the statistics.
The economics disappoint at least as often as the physics. Many systems have been installed, produced data for a few years, and then been abandoned, because nobody held a budget for maintaining the instrumentation or an obligation to act on its output. A monitoring system that does not feed a decision is an expense with a data archive attached. The practical recommendation is to define the decision first, then the feature that supports it, then the sensor that yields the feature.
Conclusion
Structural health monitoring turns a structure into an instrument that reports on itself. The available modalities are diverse and complementary: foil and vibrating wire gauges for strain over short and long horizons respectively, fiber Bragg gratings for multiplexed strain with no electrical power at the sensor, distributed fiber for continuous coverage of linear assets, accelerometers whose usefulness is decided by their low-frequency noise floor rather than their nominal range, acoustic emission for damage that is actively growing, guided waves and electromechanical impedance for local interrogation of known hot spots, and electrochemical sensors for the corrosion driving most concrete deterioration.
The analysis that converts those measurements into a diagnosis rests on operational modal analysis, which extracts modal properties from ambient excitation alone and is therefore the only practical field method, and on a hierarchy of ambition running from detection through localization and quantification to prognosis, each step markedly harder than the last. Above all it rests on confronting environmental and operational variability honestly, because temperature routinely shifts a natural frequency more than real damage does, and no amount of sensing resolution substitutes for a baseline spanning a full annual cycle and a compensation scheme suited to the structure.
Power ties the discipline to energy harvesting. Sensors go where cable cannot, must last decades, and must live on microwatts to milliwatts averaged across droughts. That constraint pushes processing onto the node, favors features computable in kilobytes, rules out continuously powered interrogators, and rewards event-driven architectures built around microwatt threshold detectors. The field has delivered load monitoring, post-event assessment, permanent hot-spot inspection, and construction verification, and it has not delivered autonomous whole-structure prognosis. Recognizing which of those a given project is actually attempting is the most useful piece of engineering judgment available in the field.