When a building or bridge is watched by a wireless sensor network, each sensor checks itself against its neighbors. If one accelerometer stops agreeing with the sensors around it, the network flags that sensor as faulty. This neighborhood consensus assumes the building's way of vibrating is fixed. The ground under a structure does not cooperate. Soil is compliant: it lowers the structure's natural frequencies, and it adds a rocking motion at the base. During strong shaking it can soften within seconds. When that happens, the pattern of agreement between sensors changes across the whole structure at once, and a consensus checker built for a fixed base can read a real foundation event as a burst of sensor faults. It can also miss a genuine fault hidden inside the shift.
We propose to let a soil-structure interaction (SSI) model decide who talks to whom. SSI modeling means treating the supporting ground as a spring-and-damper medium rather than a rigid clamp. From the current estimate of a small set of foundation parameters (translational stiffness, rocking stiffness, radiation damping), the model predicts which sensors should be correlated. The network forms its checking neighborhoods from that prediction and re-forms them when the foundation estimate drifts. A change in agreement is then tested against two hypotheses. The first is that it fits a few-parameter change in foundation impedance, meaning the soil's effective stiffness and damping as the structure sees them. That signals a foundation event, reported as such. The second is that it cannot be explained by any such change, which points to a local sensor fault.
The two-sentence version: soil changes move sensor correlations in a small, structured family of directions, while broken sensors do not. So neighborhoods for in-network fault checking should be built from, and rebuilt by, an SSI model instead of being fixed or re-learned blindly from data.
Core Claim
Consider a network with a limited communication budget. Physics-conditioned, dynamically re-formed neighborhoods will keep sensor-fault detection accurate through foundation-softening events, where both static neighborhoods and purely data-driven re-clustering fail. The same messages will also yield a running estimate of the foundation parameters that approaches what a centralized Bayesian update would give.
The mechanism rests on one piece of structure. Let $\theta$ be the vector of foundation impedance parameters. Let $\Phi(\theta)$ be the model's mode shapes, sampled at the sensor locations. Let $\Lambda$ be the modal response power over the analysis window, and let $\sigma^2$ be the sensor noise level. The predicted covariance of sensor readings is then
$$\Sigma(\theta) = \Phi(\theta)\,\Lambda\,\Phi(\theta)^\top + \sigma^2 I.$$
A foundation change $\delta\theta$ can only move $\Sigma$ along the few directions $\partial\Sigma/\partial\theta_k$, with one direction per parameter $k$. A sensor fault moves the row and column belonging to a single node, whether that is a debonded mount, electrical drift, or a dropout. The detector projects the observed change onto the span of the foundation directions. Whatever remains after that projection and is concentrated on one node is attributed to a fault.
Relation to Prior Work
The closest work is Dependable Structural Health Monitoring Using Wireless Sensor Networks (DependSHM). It groups sensors into fixed physical neighborhoods, uses a mutual-information test to decide whether a sensor still agrees with its neighbors, and uses local Kalman filters to rebuild corrupted signals. It reports about 98% fault detection on a 100-node simulated tower and a 10-story laboratory frame. It supplies our in-network machinery and our main comparison. Its neighborhoods are static, however, and its structural model has a rigid base, so it has no hypothesis available for "the soil moved."
Evidence that the gap matters comes from SSI mechanics. Shiuly et al. show that compliant soil shifts frequencies and mode shapes, and Faraonis et al. report that ignoring SSI leads monitoring algorithms to misread modal signatures. The Millikan Library study, Estimation of Soil–Structure Model Parameters for the Millikan Library Building Using a Sequential Bayesian Finite Element Model Updating Technique, found that identified soil stiffness and viscosity differ notably from NIST recommendations. That finding constrains our design: the foundation parameters must be estimated online rather than taken from code values. The same study also supplies the method we reuse. It separates fixed-base parameters from impedance parameters in two steps, and it uses an information-theoretic analysis to decide which parameters the available channels can identify. We run that analysis per candidate neighborhood to decide which sensors each neighborhood needs. The study's finding that sensors in an adjacent tunnel are better treated as unknown inputs directly informs which channels we treat as foundation anchors.
On the networking side, the idea that known physics can precompute topology comes from LEO satellite work. Wang et al. build predictive graphs from orbital mechanics so the network need not rediscover links reactively, and Jia et al. discretize changing topologies into time-expanded snapshots. We carry that philosophy over: soil mechanics plays the role of orbital mechanics, and the "topology" is the logical checking graph rather than radio links. Learned distributed controllers such as SkyLink, which uses per-node bandits for satellite link management, and Distributed Link Sparsification for Scalable Scheduling Using Graph Neural Networks, which uses a graph network to let wireless links self-silence, are relevant mainly as a warning. Both optimize communication objectives such as queues, delay, and interference with no notion of which data is physically informative. We deliberately do not learn the topology policy end-to-end: fault events are rare, and a learned policy would have little to train on.
Research Program
The system has three parts.
The first is a compact SSI model per structure. It combines a fixed-base finite element model with a lumped foundation impedance (translational and rocking springs and dashpots) and is calibrated with the sequential Bayesian procedure of the Millikan study. From this model we precompute $\Phi(\theta)$ and its sensitivities $\partial\Phi/\partial\theta$ over a grid of plausible foundation states. The result is a lookup table that low-power nodes can store, playing the same role as a precomputed orbital schedule.
The second is the topology rule. For the current $\theta$, each neighborhood is chosen to maximize the predicted mutual information between a member and the rest of the group, subject to a per-node message budget. Every neighborhood also includes one or two foundation-anchor channels, such as basement or pile-cap sensors, which carry information shared across neighborhoods. When the consensus estimate of $\theta$ moves by more than a threshold, the grid lookup can change the recommended neighborhood. If it does, nodes switch to the new neighborhoods at the next window boundary. Rewiring is therefore triggered by physics, not by a timer or by packet statistics.
The third is the two-hypothesis test described in the core claim, run inside each neighborhood. Each neighborhood's estimate of $\delta\theta$ is averaged across neighborhoods by gossip, a scheme in which nodes repeatedly exchange and average values with their neighbors. Because a real foundation change is global, the estimates should agree across neighborhoods. A fault is local, so it shows up as disagreement.
The analysis part of the program asks when the foundation directions and fault directions are separable. That depends on sensor layout and the number of identifiable parameters, and we will state the condition as a rank requirement on the stacked sensitivity matrix, per neighborhood.
Baselines
The naive reference is DependSHM-style static neighborhoods on a fixed-base model. Neighborhoods are set once by geometric proximity and never change. It is run on exactly the same injected faults, foundation events, and message budget. It shows what happens if soil is ignored entirely. Beating it means a clearly lower false-fault rate during softening windows at equal budget.
The strongest existing distributed method is DependSHM itself, with its mutual-information check and Kalman reconstruction, reimplemented from its paper. We give it one advantage it lacks in the original: its fixed neighborhoods are built from an SSI-aware model at the nominal foundation state, so that any difference we find comes from rewiring and the foundation hypothesis rather than from a better initial model. Its reported fault detection accuracy is about 98% on the Guangzhou tower simulation, against 75–85% for NFMC and 28% for SPEM at scale. Those numbers are under a fixed base with no foundation events. We will first reproduce DependSHM's accuracy in that stationary setting. Beating it means matching it there while keeping a detection F1 score at least 10 points higher during foundation-softening windows.
The strongest data-driven competitor is sliding-window re-clustering. Neighborhoods are re-formed every window from empirical cross-correlations, with no SSI model. Its window length and re-clustering threshold get the same tuning budget, on the same validation split, as our grid resolution and rewiring threshold.
The centralized upper bound is full sequential Bayesian updating on all channels, following the Millikan study, with lossless transport. No fault-detection or online-timing result for it is reported on Lacuna. It is included to show how close our distributed parameter estimate and fault calls come to centralized processing, and at what message cost.
Three ablations isolate the ingredients. The first uses physics-derived neighborhoods that are never rewired. The second keeps rewiring but drops the foundation hypothesis from the test. The third removes the shared foundation anchors.
Why not just…?
Why not just re-cluster neighborhoods from recent data correlations? This is the real threat. If soil softens slowly compared with how fast faults develop, a sliding window will track the new correlations and do fine. It has two expected weaknesses. During shaking, softening happens within seconds and a window long enough to estimate correlations lags behind it. And a slowly drifting sensor gets absorbed into whatever neighborhood now matches it, because nothing tells the clustering which changes are physically possible. The decisive experiment varies softening speed and fault drift speed independently. If sliding-window re-clustering matches us across the grid, the physics prior is not earning its place and we would report that.
Why not just stream everything and run the centralized Bayesian update? Where bandwidth is ample, this is the right answer, and it is our upper bound. The case against it is budget. DependSHM reports that centralized processing consumed 60% more energy than its in-network scheme. Simpson et al. report that high-fidelity SSI simulation ran at 2272 seconds for 500 simulated seconds, which bounds how "online" the richest centralized models can be. Neither figure settles whether a lean centralized linear update would be affordable. We therefore report the full accuracy-versus-bytes curve and name the budget below which our scheme wins.
Why not treat foundation softening as just another damage class in a generic change detector? A generic detector can flag that something changed, but it cannot say whether one sensor or the whole foundation is responsible. Only that attribution determines whether an engineer replaces a node or inspects the piles. The second ablation, rewiring without the foundation hypothesis, measures exactly this loss.
Proposed Experiments
The first testbed is the Millikan Library 2002 Yorba Linda record: 34 real channels together with the published identified soil-structure model. On it we inject DependSHM-style faults (debonding, drift, dropout) into randomly chosen channels. We synthesize foundation-softening events by lowering the identified rocking and translational stiffness along ramps of varying speed, and generate responses through the calibrated model. A network emulator on top imposes packet loss and per-node message budgets.
The second testbed is a set of simulated frames on layered soil, modeled on the G+1, G+4, and G+9 configurations of Reliability and fragility analysis of multi-story buildings on layered soil deposits under near and far-field earthquakes. These are run in Seismo-VLAB, whose method for injecting seismic motion and whose absorbing boundary layers give realistic radiation damping. In this setting soil nonlinearity produces the softening rather than a scripted ramp, which tests whether a lumped impedance family is rich enough. If the Siemens Gamesa monopile data used by Simpson et al. can be obtained, it offers a nonlinear offshore case; we do not assume access.
We measure four things: fault detection F1 score, computed separately inside and outside foundation-event windows; false fault alarms per foundation event; error and delay in the distributed $\theta$ estimate relative to the centralized update; and bytes per node per window.
Two checks decide the proposal. First, the gap between our method and sliding-window re-clustering should open as softening speeds up and close as it slows. Second, detection performance should collapse exactly where the per-neighborhood rank condition fails. If the gain does not track softening speed, or does not follow the rank condition, the claimed mechanism is wrong even if the averages look good.
References
- Dependable Structural Health Monitoring Using Wireless Sensor Networks
- Reliability and fragility analysis of multi-story buildings on layered soil deposits under near and far-field earthquakes: a soil structure interaction study
- Implications of subsoil-foundation modelling on the dynamic characteristics of a monitored bridge
- Estimation of Soil–Structure Model Parameters for the Millikan Library Building Using a Sequential Bayesian Finite Element Model Updating Technique
- Tiansuan Constellation: An Open Research Platform
- Service Function Chain Dynamic Scheduling in Space-Air-Ground Integrated Networks
- SkyLink: Scalable and Resilient Link Management in LEO Satellite Networks
- Distributed Link Sparsification for Scalable Scheduling Using Graph Neural Networks
- Nonlinear Reduced Order Modelling of Soil Structure Interaction Effects via LSTM and Autoencoder Neural Networks
- Seismo-VLAB: An Open-Source Software for Soil–Structure Interaction Analyses