Telling Glue Drift from Real Damage in Fiber-Optic Strain Sensors

Fiber Bragg gratings (FBGs) are short stretches of optical fiber that reflect one narrow color of light. The reflected color shifts when the fiber is stretched, so they serve as hair-thin strain gauges that can be glued onto or buried inside aircraft parts, composite panels and buildings. The reported results look excellent. Landing-gear loads are reconstructed within 4%, and internal damage jumps are caught before composite laminates fail. Almost all of that evidence, however, comes from one specimen tested once, over minutes or a few months. Over a long service life, the glue or coating that carries strain from the structure into the fiber slowly creeps, relaxes and debonds. A single sensor on a single specimen then cannot tell whether a slowly changing reading means the structure is weakening or the bond is. The clinical literature has a name for this trap: a short follow-up on one case can show that something is possible but cannot show a pattern.

We propose to fix this with sensor design rather than more data alone. At each measurement point, place two or three FBGs that reach the structure through deliberately different bonding paths, for example embedded, bonded with a soft adhesive, and bonded with a stiff adhesive. Then test many nominally identical specimens for long enough that the bonds actually age. Real substrate strain reaches all co-located sensors in fixed proportions. Bond aging changes those proportions. Watching the ratios therefore separates "the glue is going" from "the part is going," and pooling specimens in a hierarchical model turns a collection of single-case results into an estimate with a spread.

Core Claim

Model the reading of sensor $i$ at one location as $\varepsilon_i(t) = \alpha_i(t)\,\varepsilon_s(t) + \beta_i\,\Delta T(t) + n_i(t)$. Here $\varepsilon_s$ is the true substrate strain, $\alpha_i$ is that sensor's strain transfer ratio (the fraction of substrate strain that actually reaches the fiber core), $\beta_i$ is its temperature sensitivity, $\Delta T$ is the temperature change measured by a strain-free reference grating, and $n_i$ is noise. After temperature compensation, the ratio between two co-located sensors is

$$r_{ij}(t) = \frac{\alpha_i(t)}{\alpha_j(t)},$$

and this ratio does not depend on the unknown in-service load $\varepsilon_s$.

The claim has two parts. First, a drift in $r_{ij}$ is a load-free signature of interface degradation, because the two bonding paths age at different rates. Changes common to every sensor at a location, judged against remote sensors and occasional reference loads, are attributable to the substrate. Second, a model with layup-level effects and specimen-level random effects, fitted across many specimens, predicts the long-horizon drift of an unseen specimen better than any single-specimen model. It also tells us how short a follow-up can be before the prediction breaks.

Neither part exists in the file's literature. The analytical strain-transfer model of Strain Transfer Characteristic of a Fiber Bragg Grating Sensor Bonded to the Surface of CFRP Laminates derives $\alpha$ from shear-lag theory, the theory of how shear stress carries load across a thin bond layer. It shows that $\alpha$ depends on stacking sequence, falling from 85.1% for $[0]$ layups to 73.7% for $[90]$. It assumes the adhesive is linear-elastic and the interface never slips. We use it as the time-zero prior on $\alpha_i$ and relax exactly the assumption it makes. The embedded-FBG damage studies Damage and Failure Monitoring of Aerospace Insulation Layers Based on Embedded FBG Sensors and Structural Health Monitoring of Defective CFRP Composites Based on Multi-Sensor Technology supply the motivation. Each rests on one or two specimens per condition in a single monotonic pull. The first explicitly lists sensor creep and long-term stability as unverified. The second reads damage from a strain-load slope change, from about 0.51 intact to 0.71 defective. That is precisely the kind of slow slope change bond aging could imitate.

Research Program

The program has three parts, all built on one physical testbed.

The first part is the instrumented specimens. We build carbon-fiber laminate coupons in the three layups of the strain-transfer study, $[0]_{4s}$, $[\pm45]_{4s}$ and $[90]_{4s}$, at least six per layup across two manufacturing batches. Six per layup is a minimum for variance estimation, not a power calculation. At two gauge locations, each coupon carries a co-located triplet: an embedded FBG, a surface FBG under a compliant adhesive, and a surface FBG under a stiff adhesive. Each coupon also carries a strain-free reference FBG for temperature and an electrical strain gauge for comparison.

Specimens are assigned to a two-by-two design that plants known causes. One factor is a substrate defect, the 50 × 5 mm material-loss notch of the multi-sensor CFRP study, or none. The other is a deliberately weakened bond, made with surface contamination or under-cured adhesive, or a standard bond. Because the cause of drift is planted rather than guessed, the decomposition can be scored against truth. That is the evaluation the single-specimen studies could not perform.

The second part is a long follow-up with periodic ground truth. Coupons go through cyclic fatigue loading at room and elevated temperature, interleaved with long static holds, for at least $10^6$ cycles and several months of wall-clock time. At scheduled intervals we run a known-load proof test and full-field optical strain measurement (digital image correlation) to obtain $\varepsilon_s$ independently. Non-destructive ultrasonic scans run at the same intervals. Selected specimens are sectioned at end of life to confirm whether debonding or substrate cracking occurred. The released dataset holds about 36 specimen histories with dense FBG streams, sparse ground truth and causal labels. To our knowledge from the file, it would be the first long-duration, multi-specimen FBG aging dataset.

The third part is the decomposition model. It is a hierarchical state-space model with three pieces: a slowly varying transfer ratio $\alpha_i(t)$ for each sensor, a substrate strain $\varepsilon_s(t)$ shared within a location, and specimen random effects on initial $\alpha$ and drift rate. Initial $\alpha$ priors come from the shear-lag model. The drift dynamics can be parametric, such as a creep-relaxation curve. They can also be learned with a small neural ordinary differential equation (a network that defines the rate of change of the state), in the spirit of A Data-Driven Method for Modeling Creep-Fatigue Stress-Strain Behavior Using Neural ODEs. That work showed a black-box neural ODE captures creep-fatigue features a classical constitutive model misses, but only in one dimension and for one metal alloy. We borrow the technique for the interface rather than the bulk material.

Baselines

The naive reference point is static calibration: fit each sensor's gain once at time zero, then read strain from it forever. This is what the landing-gear system in Fiber Bragg Grating Based Load Monitoring for Carrier-Based Aircraft Main Landing Gear effectively does, and on one ground test frame it reports average relative load errors of 1.14–2.90% and a maximum of 4.00% under bidirectional loading. We apply it per sensor on our coupons. Its error against the proof-load and image-correlation truth over time shows how much aging costs.

The strongest physics baseline is shear-lag compensation from the strain-transfer study. That paper reports under 0.5% maximum model-versus-coupon error and an 11.6–19.8% reduction in measurement error on a wing skin. We apply it with layup-specific constants at time zero, using the same sensors and the same ground truth. It handles layup but not aging. Beating it means lower strain error at late time points, with the gap widening over the follow-up.

The strongest data-driven baseline is a single-specimen drift model. It takes the temperature-compensated reading of one sensor plus linear or neural ODE detrending, fitted per specimen with the same tuning budget and the same access to periodic proof loads as our method. No reported result on FBG drift exists on Lacuna for this. The neural ODE paper's MSE of 0.000516 on Alloy 617 is a different quantity and is quoted only as evidence the model class fits creep curves.

Two ablations isolate the key ingredient. In the first, the co-located triplet uses the same adhesive three times. The redundancy stays but the differential bonding path disappears. If this matches the full design, the differential path is not what matters. In the second, the hierarchical model is replaced by independent per-specimen fits, which isolates the pooling across cases. Our method counts as beating the field if two things hold. On held-out specimens it classifies the planted cause, interface versus substrate, clearly above both ablations. It also cuts late-horizon strain error relative to shear-lag compensation and to single-specimen detrending.

Why not just…?

Why not just recalibrate periodically against a known load? This is the real threat, and it works wherever a known load can be applied. We therefore include it as an oracle upper bound, not as a straw man. It fails where FBGs are most attractive: in-service structures such as flight hardware or buildings, which cannot be proof-loaded on a schedule. The ratio $r_{ij}$ requires no known load. The decisive comparison is how close the load-free method gets to the oracle as the recalibration interval is lengthened. If it gets nowhere near, the proposal's practical value collapses to "recalibrate more often," and we will say so.

Why not just add a temperature reference grating and detrend? Reference gratings remove thermal cross-sensitivity, as in A Fiber Bragg Grating Sensing Structure for the Design, Simulation and Stress Strain Monitoring of Human Puncture Surgery. Detrending a single sensor, however, removes slow substrate damage along with slow bond creep, since both look like trends. The notched-plus-good-bond arm tests this directly. A detrender that "fixes" those specimens has erased real damage.

Why not just test more specimens with existing methods? Replication fixes the single-case problem but not the confound. Each replicate still has one sensor path per point, so substrate and interface drift remain inseparable within every specimen; more specimens only give a better average of a mixed signal. The same-adhesive ablation tests whether redundancy plus replication is enough without differential paths.

Why not just compare against the electrical strain gauge? Gauges drift too, and they add local stiffness. The ship-model work Designing a Hydro-Structural Ship Model to Experimentally Measure its Vertical Bending and Torsional Vibrations cites this added stiffness as the reason to prefer FBGs. We record gauges anyway, as one more aging path.

Decisive Experiments

Four checks decide the proposal.

The first is causal recovery. On held-out specimens, under leave-one-specimen-out and leave-one-batch-out splits, can the ratio signal and the common-mode signal recover the planted two-by-two cause, scored by balanced accuracy against sectioning? The prediction is that $r_{ij}$ drifts in weak-bond arms and stays flat in notched-only arms. One failure mode would falsify the idea: the notch's strain gradient might change the ratios even with good bonds, because co-located sensors sit millimetres apart. If notched good-bond specimens show ratio drift comparable to weak-bond ones, the ratio is not a clean interface signature.

The second is a truncation study, the direct test of the limited-follow-up concern. We fit every method on the first 5%, 10%, 25% and 50% of each history and predict the rest. We report late-horizon strain error as a function of observed fraction. The point at which the hierarchical model stops improving gives an empirical minimum follow-up, and we expect the single-specimen models to keep linearizing curves that later bend.

The third is cross-unit calibration transfer, addressing the concern the landing-gear paper raises about moving calibration between units. A calibration learned on one batch is applied to the other with and without the specimen random effects, and we measure strain error at time zero and at end of life.

The fourth is the oracle gap: error versus proof-load recalibration interval, for every method.