Clock-Aware Modal Identification from Band-Limited Underwater Acoustic Telemetry

Offshore structures such as wind-turbine monopiles, jackets and subsea pipelines vibrate in a few characteristic patterns called modes. Each mode has a natural frequency, a damping ratio, and a spatial shape. Engineers track these from ambient vibration to judge foundation health. Operational modal analysis, which extracts modes from ambient vibration without a controlled hammer or shaker, currently assumes that every sensor is cabled and that all sensors share one clock. Wireless sensors on submerged members cannot meet either assumption. They must communicate by sound through water, which carries perhaps a kilobit per second and loses packets to collisions. Each node also keeps time on its own drifting crystal. This note proposes that nodes should not send raw time series. Each node sends only the handful of Fourier coefficients that sit inside narrow bands around each mode. The base station then infers the modal parameters together with each node's unknown clock error. The key point is that a clock offset at a sensor appears in the data as a phase rotation that grows linearly with frequency. Ordinary modes of a lightly damped structure have nearly real shapes, so that rotation can be estimated from the vibration data itself rather than from extra synchronization hardware.

The practical result would be mode shapes and damping ratios from retrofit acoustic sensor networks, at roughly two orders of magnitude less transmitted data than raw streaming, and with clock resynchronization needed far less often. Resynchronization costs energy and channel time on a battery-powered node. The proposal treats the underwater channel as the operating regime, modal analysis as the core mechanism, and environmental noise (rotor harmonics and hydrodynamic noise) as a constraint on which frequency bins can be trusted.

Core Claim

Bayesian frequency-domain modal identification works on Fourier coefficients in a selected band rather than on full time histories. The Bayesian Spectral Decomposition used on the Biqing Bay 5.5 MW monopile in Operational Modal Analysis of a Monopile Offshore Wind Turbine via Bayesian Spectral Decomposition is one example. Because the method only needs those coefficients, the data a node must send is already small. What breaks over acoustic links is phase coherence between nodes, and the cause is clocks, not the propagation delay itself. Propagation delay only makes data arrive late, and modal analysis of 20-minute segments is not real-time.

The model is as follows. Node $i$'s clock has an offset $\tau_i$ and a fractional skew $\epsilon_i$. Its reported Fourier coefficient at frequency $\omega_k$, for segment $s$ starting at nominal time $t_s$, is then the true coefficient multiplied by a phase factor:

$$\tilde{F}_i(\omega_k, s) = F_i(\omega_k, s)\, e^{-j\omega_k(\tau_i + \epsilon_i t_s)}.$$

Here $F_i$ is the coefficient a perfectly synchronized node would report, and $j$ is the imaginary unit. For classically damped modes, the cross-node phase of $F$ near resonance is close to 0 or $\pi$. The factor $e^{-j\omega_k(\tau_i+\epsilon_i t_s)}$ is therefore identifiable from three sources. The first is its dependence on frequency across bins and modes. The second is its drift across segments. The third is a weak prior from two-way acoustic ranging handshakes, whose timing errors are inflated by multipath.

We claim that a Bayesian model with modal parameters plus per-node offset and skew as unknowns will recover mode shapes and damping close to the synchronized-cable result. It will do so where naive "resync then run standard modal analysis" degrades, and it will transmit only in-band coefficients.

How large is the effect? A timing error $\delta$ produces a phase error $2\pi f \delta$ at a mode of frequency $f$. Hydrodynamic multipath spread is often tens of milliseconds. A typical 20 ppm crystal drifts about 70 ms per hour. At the 0.284 Hz first bending mode, a 50 ms error is only about 5°, which is harmless. At a 5 Hz jacket or higher bending mode, the same error is about 90°, which destroys mode shapes. The proposal is therefore only interesting for higher modes and for nodes that resynchronize rarely. We design the experiments to map exactly where that boundary lies.

Positioning Against Prior Work

The modal-analysis work in the file supplies the estimator and the target accuracy. Operational Modal Analysis of a Monopile Offshore Wind Turbine via Bayesian Spectral Decomposition identifies frequencies, damping and their coefficients of variation from cabled, synchronized tower accelerometers. It also reports that rotor harmonics hid the second bending mode during rated operation. Our likelihood keeps that paper's band-wise spectral structure. It adds clock nuisance parameters and removes bins contaminated by harmonics.

Estimating the foundation parameters of offshore wind turbines through Bayesian model updating shows two things that constrain our design. A deterministic fit of foundation stiffness to identified modes is non-convex, with errors up to 300%. A sampling-based Bayesian update instead recovers translational stiffness to within 2.83%. That paper assumes perfectly synchronized simulated sensors with 3% white noise. We use its 15-MW setup as our simulation testbed and its stiffness error as the downstream test of whether our modes are good enough.

The networking work motivates the problem but treats payloads as generic. Impact and Analysis of Space-Time Coupling on Slotted MAC in UANs shows that packets in slotted underwater networks collide across slots depending on node geometry. It also shows that shorter guard intervals buy 35% more peak throughput at the price of more collisions. AoI-MDP embeds observation delay into decision-making and reduces time-averaged information age from about 2.70 to 1.97. Neither paper asks what the payload must preserve for modal analysis, which is cross-node phase. We use the first paper's simulator to generate realistic loss and collision patterns. We argue that its throughput metric is the wrong objective for this application.

The closest structural-monitoring precedent is Damage Identification Algorithm of Beam Bridge Structure Based on Curvature Mode and Strain Mode. There, terrestrial wireless nodes compute local FFTs and send compressed features. That design assumes radio links and a common clock, and it sends peak magnitudes that discard inter-node phase. We keep its idea of compressing at the node but send complex in-band coefficients instead.

Two results warn against assuming that all modes are real. The fluid-structure studies Acoustic Modal Characteristics of Pump Tower Structures Based on Fluid–Structure Coupling Effects and Analysis of Vibration Characteristics of Viscoelastic Slurry Pipe Considering Fluid–Structure Interaction Effects report that fluid loading increases damping and breaks mode symmetry. Non-proportional damping would make mode shapes genuinely complex, which partially contradicts our identifiability argument. We address this head-on. The model estimates mode complexity as a parameter that the ranging prior anchors. The experiments then measure how much complexity the method tolerates before offset and true complexity can no longer be told apart.

Research Program

Each node computes FFTs over 20-minute segments. It retains complex coefficients in a band of fixed relative width around each target mode. At 1/1200 Hz frequency resolution, a 0.05 Hz band is about 60 bins. Sixteen-bit real and imaginary parts then come to about 2 kbit per mode per channel per segment. Raw streaming, even decimated to 12.5 Hz, needs about 240 kbit per channel per segment. That is roughly a hundredfold compression.

The node excludes bins near the 1P and 3P rotor harmonics. It locates them from rotor speed, which is either logged by the turbine's supervisory control system or estimated as the sharpest stationary spectral line. Excluding these bins is how the method handles the mode masking reported for the Biqing Bay turbine. Nodes also send two-way ranging timestamps whenever the protocol already exchanges them.

At the base station, a Bayesian model has three groups of unknowns. The first group is the modal parameters: frequency, damping, shape and modal force spectrum. The second is per-channel colored noise spectra that represent hydrodynamic noise. The third is the per-node offset and skew, constrained by the ranging prior.

Inference proceeds in two stages. Coordinate ascent gives a most-probable estimate. A Laplace approximation then gives posterior uncertainty. If the stiffness posterior turns out to be multimodal, we switch to transitional Markov chain Monte Carlo, the sampler used by Simpson et al. Missing packets simply remove segments or bins from the likelihood. This is the main way in which collisions from space-time coupling enter the analysis.

We will also derive when the clock parameters are identifiable. The offset is identifiable only modulo half a period for a single mode, but becomes unique when a second mode or the ranging prior is available. Skew is identifiable from the phase trend across segments.

Baselines

The oracle reference is Bayesian spectral decomposition on perfectly synchronized, losslessly transmitted data. This is the configuration of the Biqing Bay study, which reports mode-shape agreement (Modal Assurance Criterion) exceeding 0.98, frequency coefficient of variation below 0.002, and damping coefficient of variation below 0.09 (Operational Modal Analysis of a Monopile Offshore Wind Turbine via Bayesian Spectral Decomposition). We treat it as an upper bound rather than a competitor. It is run on the same segments with the same bands. Our method succeeds if it stays within 0.02 of the oracle's mode-shape agreement and within the oracle's posterior spread on damping.

The strongest practical competitor resynchronizes clocks by two-way ranging before every segment, corrects timestamps, and then runs the same Bayesian spectral decomposition on the same in-band coefficients. This is the obvious engineering answer, and for low modes it may be adequate. It receives the same packets, the same ranging exchanges, and the same bins with harmonics excluded. We beat it if, at matched energy, we achieve higher mode-shape agreement and lower damping error. Matched energy means either the same number of ranging exchanges, or fewer exchanges at equal accuracy.

The naive reference transmits only per-node peak frequencies and magnitudes, in the style of Damage Identification Algorithm of Beam Bridge Structure Based on Curvature Mode and Strain Mode. We expect it to recover frequencies but not mode-shape signs or damping. A second naive reference ignores clocks entirely and runs standard modal analysis on coefficients with uncorrected timestamps.

A bandwidth-matched alternative streams decimated raw data at the same bit budget. Over a network of ten or more nodes, this forces lower sampling rates or fewer segments.

Three ablations isolate the key ingredient. The first removes the clock nuisance parameters. The second removes the ranging prior. The third imposes strictly real mode shapes instead of estimating complexity.

The downstream comparison uses the Bayesian foundation update of Estimating the foundation parameters of offshore wind turbines through Bayesian model updating. Fed with synchronized modes, it reports 2.83% translational stiffness error. We measure how far each telemetry scheme pushes this error above 2.83%.

Why not just…?

Why not just cable the subsea sensors? Cables give synchronized data and remain the right answer for new builds that have conduits. The case for acoustic telemetry is retrofit. The file notes that contact sensing is intrusive and hard to scale (Acceleration-guided Acoustic Signal Denoising Framework), and that ocean deployments face fatigue and anchoring hazards (Barbie et al.). We do not claim our method beats cables. The oracle baseline is exactly the cable case, and we report the gap to it.

Why not just resynchronize clocks frequently and run standard modal analysis? This is the real threat. For the 0.284 Hz fundamental, our own error budget says it probably suffices. The proposal therefore stands or falls on two settings: modes above roughly 2 Hz, and duty cycles where ranging exchanges are rare or degraded by multipath. The main experiment sweeps mode frequency against resynchronization interval and multipath timing error. If the resync baseline matches us everywhere within realistic energy budgets, we will report that, and the contribution reduces to the compression scheme.

Why not just stream decimated raw data? For three or four nodes and only the fundamental, the bit budget fits. That changes when the network has ten or more nodes, when higher modes force higher sampling rates, and when collisions grow with packet size as in Impact and Analysis of Space-Time Coupling on Slotted MAC in UANs. The bandwidth-matched baseline measures this directly. Streaming also does not remove the clock problem.

Proposed Experiments

The first testbed is the IEA 15-MW simulation of Simpson et al.: parked operation, turbulent wind, waves following the JONSWAP sea-state spectrum, and 3% noise. We add virtual nodes on submerged members. Each node gets a clock with an offset and a random-walk skew. Packet loss traces come from the NS3 slotted-MAC configuration of Wang et al., under several guard coefficients. Here the true modes and stiffness are known.

The second testbed is the Biqing Bay monopile record from the BSD study, if its owners grant access. On that record we impose synthetic clock errors and packet loss on real data, under both rated and idle conditions. Real data matters because it tests harmonic exclusion and real hydrodynamic noise. Lacuna lists no public release of this record. If access is refused, we will say so and rely on the simulation alone.

The metrics are mode-shape agreement with the oracle, frequency error and damping error, posterior coverage of the true values, estimated against true clock offset, bits transmitted per segment, and the downstream translational-stiffness error.

The decisive check is a phase diagram over two axes: mode frequency from 0.28 Hz to 10 Hz, and resynchronization interval from 20 minutes to 24 hours. Each cell shows where the resync baseline falls below 0.9 mode-shape agreement and whether our method holds above it. A second check sweeps injected mode complexity to find where offset and complexity become confounded.

A third check reruns the downstream foundation update with modes from each scheme. If our modes do not move stiffness error measurably closer to the synchronized 2.83% than the resync baseline does, the method has no practical value, even if it wins on mode-shape agreement.