Spending the Sampling Budget Where the Material Damps

Battery-powered vibration sensors on bridges and towers last years longer if they send only a fraction of their samples. This works through compressive sensing: a sensor records far fewer samples than classical sampling theory requires, and a server rebuilds the vibration spectrum from them. Field studies show that about 11% of the data is enough to recover a structure's resonant frequencies. Current practice gives every sensor the same compression rate, and the recovery step assumes each spectrum consists of a few sharp peaks. That assumption depends on the material. Lightly damped steel rings with narrow peaks. Viscoelastic polymers, rubberized concrete, and fluid-loaded parts smear each resonance across many frequency bins, so their signals need more samples to rebuild.

This proposal claims that on multi-material structures, the right number of samples depends on location. Each sensor's sampling rate should be chosen together with where the sensors go, using a damping-aware prediction of how much compression each location can tolerate. At a fixed energy budget, this should recover the heavily damped modes much better than uniform compression, and it should cost little accuracy on the stiff, lightly damped parts. The search runs over sensor positions and per-node rates jointly, under a clustered-network energy model. We solve it with multi-objective particle swarm optimization and test that solver head-to-head against a greedy method. The central contribution is the damping-aware allocation, not the swarm.

Core Claim and Why It Is Not Already Done

Two lines of work come closest. Gkoktsi and colleagues evaluated two sub-Nyquist schemes for output-only modal identification on a highway overpass and a wind turbine, using fixed sensor grids and one compression level for all nodes. The first is random-sampling compressive sensing (Output-Only Vibration-Based Monitoring of Civil Infrastructure via Sub-Nyquist/Compressive Measurements Supporting Reduced Wireless Data Transmission). The second is deterministic multi-coset "power spectrum blind sampling" (Vibration-based structural performance assessment via output only sub-Nyquist/ compressive wireless sensor data). Both structures were steel or prestressed concrete. The 2017 study also reports a warning that motivates this work: the regions between peaks are the first to degrade as compression increases.

Huang et al. place a different layer on top (Energy-Efficient Heterogeneous Wireless Sensor Deployment with Multiple Objectives for Structural Health Monitoring). They use continuous particle swarm optimization to place wireless accelerometers and strain gauges on a linear-elastic bridge model. Their objective trades modal clarity against a cluster-tree energy model. Every placed sensor is assumed to stream full data.

Neither line lets the sampling rate vary per node, and neither lets material damping shape the objective. Our proposal differs from Huang in its decision variables, which are position plus rate rather than position alone. It differs from Gkoktsi in treating compression as a spatially allocated resource rather than a global setting.

The physics suggests a testable scaling. A mode with damping ratio $\zeta_r$ at frequency $f_r$ has a half-power bandwidth of $2\zeta_r f_r$. Over a record of length $T$, its peak therefore occupies roughly $2\zeta_r f_r T$ frequency bins. Compressive recovery needs a number of measurements that grows roughly linearly with the number of occupied coefficients, up to a logarithmic factor. So the minimum sampling fraction at node $i$ should scale approximately with $\sum_r \zeta_r f_r T$, summed over the modes that node observes with significant amplitude. Evidence that damping regimes differ strongly by material comes from studies showing large, mode-dependent frequency shifts in viscoelastic pipes under fluid coupling (Analysis of Vibration Characteristics of Viscoelastic Slurry Pipe Considering Fluid–Structure Interaction Effects). Further evidence comes from altered energy dissipation in fiber-reinforced rubberized concrete (Enhanced Dynamic Compressive Behavior of Rubberized Concrete with Steel–Glass Fibers). If the scaling holds, it tells us directly where samples are worth their energy.

Research Program

We first build multi-material finite element models with identified damping. The four-span continuous bridge model of Huang et al. provides the network geometry. We then assign some spans or decks viscoelastic or rubberized-concrete damping levels, and the girders steel damping levels. A second model is the polyethylene slurry pipe with fluid coupling from Hu et al. The ambient response comes from forcing each model with white noise, and also with colored or harmonic noise so that we do not flatter the method. The simulated signals are then subsampled under both compression schemes and passed through frequency-domain decomposition, a standard output-only modal identification method.

From these runs we fit a per-node surrogate. It predicts identification error in frequency, damping ratio, and mode-shape agreement as a function of sampling fraction, local modal amplitude, and local damping. Mode-shape agreement is measured by the modal assurance criterion (MAC), a 0-to-1 similarity score between identified and reference mode shapes. The surrogate turns the expensive simulate-and-identify loop into a cheap objective.

The optimization problem chooses a sensor set $S$ and per-node sampling fractions $\rho_i \in [\rho_{\min}, 1]$. It minimizes the predicted weighted modal error $E(S,\rho)$ subject to

$$\sum_{i \in S} \big(a\,\rho_i + b\,\rho_i\, d_{i,\mathrm{CH}(i)}^{\,n}\big) + E_{\mathrm{CH}}(S) \le B,$$

where $a$ is the per-sample acquisition cost, $b\,d^n$ is the radio cost of reaching node $i$'s cluster head at distance $d$ with path-loss exponent $n \ge 2$, $E_{\mathrm{CH}}$ is the cluster-head relay cost, and $B$ is the energy budget. The choice of cluster heads couples the nodes: raising one node's rate increases its cluster head's relay load. That coupling is why we solve the problem with a multi-objective swarm whose external archive of non-dominated solutions is clustered to keep it diverse, following Chenye Qiu et al.. The discrete placement variables are handled by relaxation and rounding.

Baselines

The naive reference point is full-rate streaming from Huang et al.'s placement, which is the accuracy ceiling. Next to it we run uniform compression at the 11% level of Gkoktsi et al. on the same placement. On the Bärenbohlstrasse overpass, Gkoktsi et al. (2019) report natural-frequency errors below 1% (also stated as below 1.3%) and MAC generally above 0.90 at 11% data. The same work reports battery life rising from about 1 year to over 8.5 years. Gkoktsi et al. (2017) report MAC consistently above 0.94 at an 89% reduction. These are our anchors for a lightly damped concrete deck. We found no reported result on Lacuna for damping-ratio identification error under compression, nor for compressed identification on viscoelastic or multi-material structures. We will therefore report both anchors ourselves.

The strongest existing method is Huang et al.'s swarm placement combined with uniform power spectrum blind sampling, with the uniform rate tuned to exactly the same energy budget $B$. For the network alone, Huang et al. report a 33.3% energy saving of the cluster tree over a flat topology. We keep their cluster-tree model so that the energy accounting is shared.

A strong non-swarm competitor is greedy allocation. It starts from the empty set. At each step it adds either a new sensor or a rate increment at an existing one, whichever gives the largest predicted error reduction per joule, and it recomputes cluster heads after every step. Every method gets the same surrogate, the same number of objective evaluations, and a matched hyperparameter tuning budget. For the swarm, tuning covers inertia and acceleration weights. For greedy, it covers the step size.

Two ablations isolate the key ingredient. In the first, the full joint optimizer is forced to use one uniform rate. In the second, the optimizer keeps variable rates but its surrogate assumes uniform steel-like damping everywhere, which makes it material-blind.

Our method beats the strongest baseline if two conditions hold at matched energy. It must reduce damping-ratio error on modes dominated by high-damping regions by a margin larger than the spread across 20 random seeds. It must also keep MAC at or above the uniform baseline's value on the lightly damped modes.

Why not just…?

Why not just use Effective Independence or a greedy method instead of a swarm? Effective Independence is the standard deterministic placement rule, which keeps the sensors that best preserve independence of the measured mode shapes. It assumes mode-shape vectors that can be orthogonalized. Huang et al. note that this fails for mixed accelerometer and strain arrays. Effective Independence also has no notion of sampling rate or radio energy. The greedy per-joule baseline removes those objections, and it is a real threat. If it matches the swarm, we will report greedy as the recommended solver. The claim about damping-aware allocation stands or falls independently of the solver.

Why not just model the structure as linear elastic with uniform damping? That is exactly the material-blind ablation. The slurry-pipe results of Hu et al., with frequency reductions of about 54% and odd and even modes dropping by different amounts, suggest that material and fluid effects are not small. What matters here, though, is whether ignoring them wastes samples. The ablation measures this directly at matched energy.

Why not just raise the uniform compression rate for everyone? At a fixed budget, raising every node's rate means spending samples on steel regions where peaks are already sparse. The uniform-rate ablation is the decisive test. If it matches variable allocation within seed noise, the core claim is false.

Why not just duty-cycle, or run an FFT on the node and send the top peaks? Duty-cycling risks missing non-stationary ambient events. Sending only peak locations discards the peak widths from which damping ratios are estimated, and damping is precisely what degrades in viscoelastic regions. We include on-node peak picking at matched energy as an extra comparison on damping error.

Proposed Experiments

The first experiment tests the scaling prediction directly. On each model, we sweep the damping ratio of a region and find the minimum sampling fraction that holds MAC above 0.90 and frequency error below 1%. We then check whether that minimum grows roughly linearly with $\zeta_r f_r T$.

The second is the matched-energy comparison of all baselines on the multi-material bridge and the pipe, reported as full Pareto fronts of error against energy.

The third probes damage. We introduce stiffness loss in one region, patterned on the crack-driven frequency drops reported in Experimental study on modal analysis and vibration response of tunnel structures under different damage conditions due to subway train loads. We then ask at what energy budget each method still detects that the frequency has shifted, and whether it can tell a genuine shift apart from artifacts of reconstruction between peaks.

If the overpass or wind-turbine records can be obtained, we will run a retrospective subsampling check on real data. Because these structures have a single dominant material, we predict little gain there. That prediction serves as a falsifiable negative control.