Knowing When You Don't Know: Evidential Trust in Each GNSS Constellation During Ionospheric Storms

A receiver that tracks GPS, Galileo and BeiDou together usually positions faster and more accurately than one that tracks GPS alone. During a geomagnetic storm, that advantage shrinks. Across 305 Australian stations in the May 2024 storm, the triple-constellation precise positioning solution took 87.2% longer to converge vertically, compared with 56.9% longer for GPS alone, and its vertical accuracy fell by 18.5% versus 11.8% (Wang et al.). Each extra constellation adds geometry, and in a storm it can also add corrupted signals. Current processing has two crude ways to react. It can apply fixed weights that ignore the storm, or it can wait until enough residuals pile up to show that something is wrong. During that wait, the filter has no principled way to say "I have no evidence yet about whether Galileo is healthy." The only honest answer in that window is "I don't know."

We propose to give each constellation an explicit trust state based on Dempster–Shafer belief functions. This is a generalization of probability that can put weight on "nominal or degraded, undetermined" separately from "nominal" and "degraded." That trust state is built from three sources. The first is the constellation's consistency with the others. The second is its own carrier-phase behavior. The third is coarse, categorical space-weather advisories ("minor", "strong", "severe" storm), which we treat as vague linguistic evidence rather than precise numbers. The core claim is testable. Keeping ignorance separate from evidence of degradation lets the positioning engine respond differently to the two. When it is ignorant, it keeps the constellation in the solution and widens its safety bound. When it has evidence of degradation, it down-weights the constellation. This separation should recover most of the storm-induced convergence loss without increasing the rate at which the true error exceeds the stated bound. Collapsing the same evidence to an ordinary probability should not achieve this.

Core Claim

The two failure modes call for opposite actions, so they should be represented separately. A constellation that conflicts with the others should lose weight in the position estimate. A constellation about which nothing is yet known should keep its weight, because removing geometry slows convergence, but the protection level should grow. A protection level is the receiver's stated bound on its own error. A single scalar reliability score cannot express both responses at once.

For each constellation $c$ we use a two-state frame $\Theta_c = \{N, D\}$ (nominal, degraded). The mass function $m_c$ distributes unit belief over $\{N\}$, $\{D\}$ and $\Theta_c$, and $m_c(\Theta_c)$ is the explicit ignorance. The fused mass drives two separate controls:

$$\sigma_{c}^{2} \leftarrow \sigma_{c}^{2}\,\bigl(1 + \alpha\, m_c(\{D\})\bigr), \qquad PL \leftarrow PL_0 + \beta \sum_c m_c(\Theta_c)\, \Delta_c .$$

Here $\sigma_c^2$ is the observation variance of constellation $c$ in the filter, and $PL_0$ is the nominal protection level. The term $\Delta_c$ is the bound increase obtained by excluding constellation $c$, which is the standard fault-subset quantity computed in integrity monitoring. The parameters $\alpha$ and $\beta$ are tuned on calm days only. Keeping the frame per-constellation is deliberate. Belief-function methods scale exponentially with frame size, so a joint frame over four constellations would be expensive to compute. Four two-state frames keep the computation trivial and still let constellations be judged independently.

How the Evidence Is Built

The first source is cross-constellation consistency. We solve with constellation $c$ left out, then examine the residuals of $c$'s observations against that solution. Large, persistent residuals put mass on $\{D\}$. Small residuals put mass on $\{N\}$. Too few observations, such as just after a reset or with low satellite counts, put mass on $\Theta_c$. This is the multi-constellation structure working for us: tight combination (Liu et al.) provides an independent reference for each constellation.

The second source is the constellation's own carrier-phase health, meaning its rate of geometry-free cycle slips and of ionospheric rate changes. We reuse the adaptive detection thresholds of Tang et al.. This source turns slip statistics into mass instead of discarding them after a binary decision.

The third source is categorical space-weather advisories, and this is where linguistic uncertainty enters. A storm level such as "G3, strong" is a word with fuzzy boundaries. It is also issued for the whole planet, while its effect on a given station and constellation varies widely. A cloud model represents such a concept with three numbers: a center (expectation), a boundary width (entropy), and the spread of that width across contexts (hyper-entropy). The hyper-entropy captures the regional variability of the advisory. Cloud models were turned into belief masses for expert ratings by Zhou et al. and for vibration features by Li et al.. We follow the same route, but with a rule tied to how much the advisory is worth locally. Hyper-entropy is fitted from how much the storm-time degradation of a given advisory level varied across stations in the training regions, and that variability becomes mass on $\Theta_c$ rather than on $\{D\}$. A high advisory level therefore mainly says "be ready." Mostly it raises ignorance and the protection level, and it does not on its own cut a constellation's weight.

We deliberately avoid parsing free text with a language model. QUITE shows that models translating words of estimative probability ("likely", "unlikely") into numbers are brittle, with GPT-4 zero-shot reaching 36.7%. Advisory categories are already a small controlled vocabulary, so the uncertainty to model is semantic and regional, not syntactic.

Combination Under Conflict

Dempster's rule multiplies masses across sources and renormalizes away their disagreement, $K$. When sources strongly disagree, this produces confident nonsense, as A Defect in Dempster-Shafer Theory and Can Evidence be Combined in the Dempster-Shafer Theory? argue. Storms are exactly when the advisory may say "severe" while a constellation's residuals look clean. These papers therefore constrain the design. We use Dempster's rule only when $K$ is small. Above a threshold, we use proportional conflict redistribution from Dezert–Smarandache theory. As an alternative, we transfer the conflict to $\Theta_c$, which treats disagreement as a reason for caution rather than as noise to discard. Which of the two works better is an empirical question within the program.

Evidential fusion of numeric and textual sources has been done in clinical prediction. Ruan et al. fused EHR tables and notes with Dempster's rule and report an 18.08% lower Brier score. That work motivates the approach, but it outputs a class label, not a variance and a bound inside a physical estimator, and it does not need to handle high-conflict states.

Research Program

We will build the evidence layer on a PPP-RTK pipeline, meaning precise point positioning sped up by network corrections. The pipeline will follow the server- and user-side design of Wang et al. and use GFZ real-time orbits and clocks and IGS differential code biases. We will also attach the same layer to an advanced receiver autonomous integrity monitoring (ARAIM) bound computation of the kind studied in the VPL optimization paper. That paper re-allocates integrity risk across fault subsets with particle swarm optimization over fixed Gaussian fault rates. Our layer makes the per-constellation inputs to that allocation time-varying and ignorance-aware, so the two methods compose rather than compete.

The analysis has three parts. First, we will measure how mass on $\{D\}$ and on $\Theta_c$ evolves per constellation across the storm onset. Second, we will measure whether ignorance mass rises before residuals reveal degradation. Third, we will measure how much of the triple-constellation penalty is removed by down-weighting compared with protection-level inflation.

Baselines

The naive reference is the unmodified triple-constellation PPP-RTK with standard elevation-dependent weighting. On the 305-station data, Wang et al. report storm-time increases of 55.6% in horizontal and 87.2% in vertical convergence time, and declines of 10.4% horizontal and 18.5% vertical accuracy. GPS-only rises by 56.9% in vertical convergence time and loses 11.8% in vertical accuracy. Our pipeline will not exactly match theirs, so we will first reproduce these relative numbers within our pipeline before reporting any gain.

The strongest existing competitor is an innovation-based adaptive filter with per-constellation variance factors. This method re-estimates each constellation's noise from its recent residuals, using the same cycle-slip thresholds (Tang et al.) and the same leave-one-constellation-out residuals that feed our first source. It uses the physical evidence our method uses, without the ignorance channel, which makes it the fairest competitor. We found no reported storm-time result for this specific baseline on Lacuna, so we will run it ourselves. The second competitor is scalar advisory scaling, where the planetary index or storm level multiplies all variances and the protection level through a lookup curve fitted on the same training stations.

For integrity, the reference is ARAIM with particle swarm risk allocation. That paper reports average vertical protection levels lower by up to about 1 m in multi-constellation setups and availability rising from 90.77% to 92.44%, on its own simulated and real data, not on the Australian storm. We will run this allocation on the storm data with and without our time-varying inputs.

The key ablation replaces each fused mass function by its pignistic probability, the standard way to collapse a belief function into an ordinary probability, so that weights and bounds come from a single number. A second ablation removes the advisory source.

All methods will use the same stations, corrections and calm-day tuning budget. Every tunable parameter, $\alpha$ and $\beta$ included, gets an equal number of grid points fitted on May 1–10. Testing uses May 10–13, with stations split by region so that cloud parameters are never fitted on test stations. Beating the strongest baseline means two things together. Storm-time vertical convergence must be at least 25% shorter than under the adaptive filter. The misleading-information rate, meaning epochs where true error exceeds the stated bound, must not be higher, with confidence intervals obtained by bootstrapping over stations.

Why not just…?

Why not just use an adaptive Kalman filter that inflates noise from residuals? This is the most serious threat. It responds to evidence of degradation exactly as our $\{D\}$ mass does. What it cannot do is act before residuals accumulate, or distinguish a silent constellation from a healthy one. The decisive measurement is the first 15–30 minutes after storm onset, plus epochs following filter resets. If the adaptive filter matches us there on both convergence and bound validity, the ignorance channel adds nothing and we will report that.

Why not just multiply the covariance by a scalar storm index? Luo et al. found that compressing structured language feedback into scalars discards information, but that evidence comes from language-model training, not navigation. We do not rely on it. The concrete argument is that a planetary index affects all constellations and stations equally, so it cannot down-weight BeiDou in one region while leaving GPS alone. It also forces one number to control both weighting and bounding. The scalar-scaling baseline and the no-advisory ablation together measure how much the advisory contributes at all. If both match the full method, the linguistic component is cosmetic and the claim narrows to evidential fusion of physical evidence.

Why not just drop the extra constellations during storms? GPS-only degrades less in relative terms, but Wang et al. found that the triple-constellation solution still outperformed GPS-only in convergence and accuracy, so wholesale exclusion throws away geometry. We include storm-triggered exclusion as an extra baseline. Our method should beat it on accuracy and convergence, and if it does not, selective trust is unnecessary.

Why not have a language model read the bulletins directly? The QUITE results show that turning vague probability words into numbers is where such models are weakest, and advisory categories already form a closed vocabulary. We do not test this alternative.

Proposed Experiments

The primary experiment uses the 305-station multi-GNSS dataset (May 1–13, 2024), with calm days for tuning and storm days for testing. We will report horizontal and vertical convergence time, positioning accuracy, wide-lane and narrow-lane delay stability, protection levels, availability at the alert limits used in the VPL optimization paper, and misleading-information rate.

Three checks decide the claim. First, the pignistic ablation must be measurably worse than the full method in the onset window, which isolates the value of representing ignorance. Second, mass on $\Theta_c$ must rise before mass on $\{D\}$ for constellations that later degrade, which tests whether the early-warning mechanism exists. Third, bound validity must hold at every station grouping, or the convergence gains are bought with unsafe bounds.

A secondary transfer test uses the multi-constellation raw measurements in the Google Smartphone Decimeter Challenge 2021. There the coarse categorical source is an environment label (open sky, suburban, urban canyon), and the competitor is neural pseudorange correction. That method improved horizontal score by 59% and 10% over weighted least squares in its two scenarios, using GPS only (Liu et al.). The question is whether per-constellation ignorance helps when reflected-signal errors, rather than the ionosphere, drive conflict between constellations. Treating the neural correction as one more source of $\{N\}/\{D\}$ evidence is a natural extension.