Silence Is Data: Censored Gaussian Processes for Event-Triggered Low-Power Sensor Networks

Battery-powered sensor networks do not die mainly from sending data. They die from listening. A node's receiver draws power comparable to its transmitter while it waits for packets that usually never come, and this "idle listening" dominates energy use when traffic is sparse. Hardware and protocol fixes reduce the waiting but leave a second waste in place. One fix is a separate microwatt wake-up radio that switches the main radio on only when called. Another is a protocol in which nodes briefly sample the channel for an extra-long preamble. Either way, the network still wakes, polls, and transmits on a schedule that knows nothing about the physical field being measured. Scheduling by the field instead means sending only when a reading changes meaningfully, and waking a node only when the network is genuinely unsure about it. That kind of scheduling creates a statistical problem that existing tools handle poorly.

The idea of this note is simple. Suppose a node reports only when its reading moves more than a set amount δ from its last report. Then its silence is not missing data. Silence tells the gateway that the node's value stayed inside a known band. A Gaussian process (GP) is a probabilistic model that places a distribution over smooth functions and gives calibrated error bars. A GP at the gateway can treat each silent interval as an interval-censored observation and pool those bands across space through its covariance. It can then use its posterior uncertainty to decide which node to wake next, and when. The scalability problem is real. The gateway must keep a spatiotemporal GP updated online over many nodes and long horizons, with irregular, event-driven, mostly censored observations, a regime that the streaming GPs in the file do not handle.

Core Claim

A gateway-side sparse spatiotemporal GP that conditions on silence as interval-censored evidence can do two things that per-node event triggering and uncensored GP interpolation cannot. It can reconstruct the field at unsensed locations to a target error with substantially fewer uplinks and wake-ups. It can also set each node's report threshold and next rendezvous time from posterior uncertainty, so nodes need no periodic listening at all between rendezvous. The microcontroller on each node only compares its reading against its last report, so no GP algebra ever runs on the 16-bit node. This resolves the tension between GP inversion cost and embedded budgets in the simplest way: the model lives where the power is.

Concretely, suppose node $i$ at location $x_i$ last reported value $y_i$ and stays silent through sensing instants $t \in S_i$. Its silence contributes the likelihood

$$p(\text{silence}_i \mid f) = \prod_{t \in S_i} \left[\Phi\!\left(\frac{y_i + \delta_i - f(x_i,t)}{\sigma_n}\right) - \Phi\!\left(\frac{y_i - \delta_i - f(x_i,t)}{\sigma_n}\right)\right],$$

Here $f$ is the latent field, $\Phi$ is the standard normal cumulative distribution, $\sigma_n$ is the sensor noise, and $\delta_i$ is the node's report threshold. Reports contribute ordinary Gaussian terms. Spatial correlation is what makes this more than a per-node trick. Bands from several silent neighbors jointly constrain the field between them, and one neighbor's report updates belief about silent nodes nearby. Because of this pooling, larger thresholds, and therefore fewer transmissions, should be affordable at the same field-level error.

Relation to Prior Work

The closest work splits into three roles. Two papers motivate the gap. HOSNA uses a base-station LSTM forecaster plus genetic-algorithm/particle-swarm scheduling to proactively wake agricultural nodes. It shows that field-aware activation beats LEACH and random duty cycling, but it acts on point forecasts with no uncertainty, so it cannot tell when a sleeping region has become unknown. Variational Gaussian process for optimal sensor placement (VGPosp) uses a sparse variational GP with flow preconditioning and Markov-chain Monte Carlo search to choose where to install sensors in a 3D computational fluid dynamics room. It solves placement offline and never models when nodes talk.

Two further papers supply the radio substrate and its constraints. On-Demand TDMA and KRATOS eliminate idle listening in hardware with a 1.83 µW wake-up receiver. However, every beacon wakes the whole cluster and every node reports every round. LPWC avoids extra hardware by having nodes periodically run LoRa channel activity detection (a short check for a preamble) against long gateway preambles. Its fixed check interval (about 0.45 s) is field-agnostic, and it suffers overhearing. Our scheduler targets exactly these two gaps: blind, cluster-wide wake-ups and field-agnostic check intervals.

The modeling side has several sources. Recurrent Memory for Online Interdomain Gaussian Processes (OHSVGP) keeps a fixed-size polynomial memory updated by linear differential equations, but it assumes regular streams with fixed hyperparameters. Its continuous-time updates can in principle be integrated across arbitrary gaps, which we will test, and we treat its fixed-hyperparameter limitation as a constraint. Leveraging Locality and Robustness to Achieve Massively Scalable Gaussian Process Regression (GPnn) contributes two things: hyperparameters fitted on a subset, and local neighborhood prediction with post-hoc variance calibration. We reuse both for the spatial component. Approximation-Aware Bayesian Optimization shows that utility-weighted variational objectives move scarce inducing points toward decision-relevant regions. Monitoring networks must catch anomalies anywhere, so we deliberately do not concentrate the approximation budget that way, and the experiments test whether that choice is right. Sleep Control for Tracking in Sensor Networks poses sleep control as a partially observable decision problem with parametric target dynamics. We replace its known dynamics with a learned nonparametric field and give up its policy-optimality analysis in exchange for scalability.

Research Program

The work has three components.

The first is the inference engine. It is a spatiotemporal GP with a separable kernel: a Matérn temporal part in state-space form, which handles irregular timestamps natively, and a spatial part over inducing locations. Censored terms enter through a variational posterior whose expected log-likelihood under the band likelihood above has a closed form via Gaussian integrals. Each report or rendezvous triggers an update whose cost depends on the number of inducing points, not on history length. We will compare this against an OHSVGP variant integrated across irregular gaps and against a GPnn-style local model refit per query. The aim is to establish which scalable family tolerates heavy censoring without overconfident variances.

The second is the scheduler. After each update, the gateway computes, for every node and for held-out field locations, the time at which posterior standard deviation will cross a tolerance τ. That crossing time becomes the node's next rendezvous. The gateway delivers it as an addressed wake-up pattern on KRATOS-style hardware or piggybacks it on a downlink in LPWC. Between rendezvous, the node does no channel checks; its only radio activity is event-triggered uplink. Thresholds $\delta_i$ are retuned at rendezvous so that the expected uplink count is minimized subject to the predicted field error staying under τ.

The third is energy accounting. The power figures are those of the SX1276 radio and the KRATOS node, and channel behavior is simulated in LoRaSim, so that every modeling gain is translated into joules and battery days rather than message counts.

Baselines

The naive reference point has two parts. The first is On-Demand TDMA polling of every node at a fixed period, as in On-Demand TDMA. It reports over three years of life on a 1200 mAh battery at 1-minute polling and 100% packet delivery on its 11-node testbed. We run it with identical node positions, sensing period, and energy model, and treat its full-data reconstruction as the accuracy ceiling. The second is LPWC with its analytical optimal check interval. It reports roughly 324 days on a 3000 mAh battery in LoRaSim, and we re-simulate it under our traffic so that overhearing and preamble airtime are included.

The strongest existing predictive method in the file is HOSNA's LSTM-driven activation. Its paper reports 117.8 J total energy and 94.0% data accuracy over 1,000 rounds in its agricultural simulation. Those numbers come from a different simulator and are not directly comparable. We will therefore reimplement an LSTM forecaster that triggers wake-ups when its forecast changes by more than δ. It gets the same training data and a hyperparameter search budget equal to ours. To beat it, our method must reach equal or lower field error at fewer joules, and must detect injected plume onsets with shorter latency.

The baseline we consider most dangerous is send-on-delta with zero-order hold. Each node reports when its reading moves δ, and the gateway holds the last value at sensors and interpolates by standard GP regression on held values at unsensed locations. This shares our node-side behavior exactly.

The key ablation is our full model with the censored likelihood replaced by treating silent intervals as missing. A second ablation fits independent per-node temporal GPs with censoring but no spatial coupling. Together these two isolate the spatial-pooling contribution.

For the field-reconstruction benchmark, VGPosp reports that assimilating 7 well-placed sensors reduces Clarence Centre mean squared error from 0.17 to 0.0005. We use 0.0005 as the full-information target. Beating the baselines means matching a given field error with clearly fewer uplinks plus wake-ups than send-on-delta with hold and than the uncensored ablation. For the GP-with-censoring setting in wireless scheduling itself, there is no reported result on Lacuna.

Why not just…?

Why not just use send-on-delta with last-value hold? This is the real threat. At sensor locations it already bounds error by δ with no model at all. Its weakness is between sensors and across neighbors: it cannot let one node's report revise belief about a silent neighbor, and it cannot choose δ or rendezvous times from field uncertainty. Whether that matters is empirical. If, at matched error at held-out Clarence Centre locations, the censored GP saves fewer than a modest fraction of uplinks over this baseline, the proposal's value reduces to scheduling alone, and we will report it that way.

Why not just use an LSTM forecaster as in HOSNA? HOSNA's forecaster gives point predictions. The scheduler needs to know when uncertainty about a sleeping region has grown, which a deterministic forecast cannot express. Our injected-anomaly experiments test this directly: events placed in regions where nodes have been silent for a long time should be missed or caught late by forecast-triggered wake-ups. A deep ensemble would partially close this gap. We include one as a secondary check, though it cannot ingest censored silence without the same likelihood machinery.

Why not just rely on wake-up receivers? KRATOS already makes waiting cheap, so on that hardware our gain is not in idle power. It is in avoiding unnecessary wake-ups and transmissions, at 240 mW per transmission, for nodes whose values the model already knows. The wake-up receiver's −50 dBm sensitivity also forces cluster heads. Fewer rounds reduce their load, which we measure separately.

Why not just use a per-node Kalman filter with censoring? This is a reasonable control-style alternative, and it is close to our per-node ablation. If spatial coupling adds nothing over it, then GP scalability is irrelevant to the problem. The ablation is designed to settle exactly that question.

Proposed Experiments

The primary benchmark is the Clarence Centre simulation (10,000 locations, 3,500 timesteps). We place 7 to 50 virtual nodes, both at VGPosp-style locations and at random, run each policy, and plot held-out field mean squared error against total energy using SX1276 and KRATOS power figures. Temporal stress tests use the Solar series and ERA5 subsets from OHSVGP, with artificial node grids. LoRaSim supplies collision, overhearing, and airtime-limit effects.

Three checks decide the outcome. First, the censored model must dominate both send-on-delta with hold and the uncensored ablation on the error-versus-energy curve at held-out locations. Second, posterior coverage of 90% intervals must stay near nominal under heavy censoring; we use GPnn's calibration step where needed, and we check whether miscalibration grows with silence length. Third, on injected localized events, detection latency must not exceed that of fixed polling at equal energy. Additional measurements are gateway update time per event as node count grows from tens to thousands, and the sensitivity of all three results to fixed versus periodically refit hyperparameters. The decisive negative outcome would be a model that saves energy only by becoming overconfident in silent regions. The coverage and anomaly-latency checks exist to expose exactly that failure.