Constellation-Grouped Causal Normalization as an Adaptive Weighting Signal for Multi-GNSS Positioning

Precise satellite positioning with several constellations at once (the US GPS, China's BeiDou, Europe's Galileo, Japan's QZSS) depends on external correction streams. These streams carry satellite orbit and clock corrections produced by different analysis centers, and their quality differs a lot by constellation and by producer. BeiDou cross-track orbit errors are roughly six times larger than GPS errors within a single stream. Switching from one analysis center to another changes the error distribution abruptly. Learned models that decide how much to trust each satellite are now common. They almost always pass their inputs through normalization, which rescales each signal to zero mean and unit variance before the network sees it. We argue this creates a specific failure. The normalization step divides out the absolute size of each satellite's post-correction residuals, and that size is exactly the evidence of which constellation's corrections are currently poor. Standard windowed normalization also computes its statistics using future samples, which a real-time receiver does not have.

We propose to stop throwing that scale away. The input normalizer computes its statistics causally, using only past data, and computes them separately for each constellation and orbit type. The removed scale then goes directly into the positioning filter as that group's measurement variance. The network only learns a correction to that variance from the scale-free shape of the signals. The normalization statistics therefore become the adaptive parameter. When a user switches correction streams, the model re-adapts without backpropagation: the causal statistics re-estimate themselves, and the per-group gain and bias inside the network are re-matched to stored reference moments. The core claim can be tested directly. Under a switch to an analysis center never seen in training, this model should recover pre-switch accuracy faster than both a learned weighting model with standard normalization and a classical adaptive filter that re-estimates per-constellation variances with no learning.

What the literature establishes, and where it stops

The motivating evidence is the constellation- and producer-dependent quality of corrections. Assessment of Multiple GNSS Real-Time SSR Products from Different Analysis Centers compares real-time orbit and clock correction streams across producers. It reports, for example, GPS orbits better than 8 cm RMS, and BeiDou and Galileo orbits from the CNES stream at about 14.54 cm and 4.42 cm respectively. Validation and Assessment of Multi-GNSS Real-Time Precise Point Positioning in Simulated Kinematic Mode Using IGS Real-Time Service traces these differences to uneven tracking-network coverage, with BeiDou along/cross-track errors of about 17 and 19 cm against about 3 cm for GPS. Precise Orbit and Clock Products of Galileo, BDS and QZSS from MGEX Since 2018 shows that satellites in geostationary and inclined-geosynchronous orbits are worse than satellites in medium Earth orbit. Radial errors reach about 20 cm for the QZSS geostationary satellite, because stationary geometry makes orbit and clock errors hard to separate. That result is why we group by orbit type as well as by constellation. Performance Evaluation of Precise Point Positioning for BeiDou-3 B1c/B2a Signals documents a persistent vertical deficit, with GPS static positioning 46.3% better vertically than BeiDou-3. This tells us to report the vertical axis separately.

Two learned approaches already touch this problem, and both leave the scale question alone. Deep Reinforcement Learning with Robust Spatial-Temporal Representation for Improving GNSS Positioning Correction models constellations as nodes in a graph network, uses a Transformer over time, and has a policy output a position correction. It uses default Transformer normalization. It never treats correction-stream quality as an input or examines what its normalization removes. Deep Learning-Based Diffraction Identification and Uncertainty-Aware Adaptive Weighting maps classifier entropy to measurement variance. It shows that soft reweighting beats hard satellite exclusion: vertical RMS falls from 1.1932 m to 0.2920 m and the ambiguity fixing rate rises to 99.9%. Its inputs are rover-only features, though, with no correction-quality or per-constellation scale. It supports our choice to output continuous variances rather than accept/reject decisions, and we do not claim soft weighting itself as new.

Two normalization papers supply our method. Does Normalization Choice Matter for Causal Large Time-Series Models? shows three things for univariate forecasting. Global normalization leaks future information. Strictly causal statistics generalize better. An inverse hyperbolic sine transform applied before normalization tames outliers. We reuse the causal-plus-inverse-sinh recipe, but on multichannel, grouped data where the statistics are exported rather than discarded. Revisiting Layer Normalization for Point Cloud Test Time Adaptation shows that under domain shift, layer normalization's centering and rescaling already stabilize features. The useful thing to adapt is the learned gain and bias applied afterward, which it matches to source moments without backpropagation. It reports +12.35% on ScanObjectNN-C at low memory. We reuse this adaptation step per constellation group. That paper also notes degradation with small batches. This is a real constraint for a single receiver, so we pool moments over a causal time window rather than over a batch.

Core Claim

The mechanism fits in one equation. For satellite $s$ in group $g(s)$, meaning constellation crossed with orbit type, the filter's measurement variance at epoch $t$ is

$$R_s(t) = \hat\sigma^2_{g(s)}(t)\,\exp\big(f_\theta(z_s(t))\big),$$

where $\hat\sigma^2_{g}(t)$ is the causal, exponentially weighted variance of post-correction residuals in group $g$, which is also the denominator the input normalizer divides by. $z_s(t)$ is the satellite's normalized feature history: residuals after inverse-sinh and causal normalization, signal-to-noise ratio, elevation, correction age, and correction-issue changes. $f_\theta$ is a small causal Transformer that outputs a log-variance correction. The forgetting factor $\lambda_g(t)$ of the running variance is itself adapted. When the filter's normalized innovation squared exceeds a chi-square bound, the factor shrinks multiplicatively; otherwise it relaxes additively. This is the multiplicative-decrease rule that Adaptive balancing of gradient and update computation times uses to tune inner-solver accuracy. We borrow it for a one-dimensional parameter where its per-step cost is negligible, which avoids the latency concern that iterative safeguards raise. This leaves two meta-parameters, the bound and the step sizes. We fix them once on validation stations and never retune them per stream, and we report sensitivity to them.

Research Program

We will build a standard float precise-point-positioning extended Kalman filter, meaning single-receiver positioning refined with external corrections, with states for position, clocks, troposphere and ambiguities. The only change is that the measurement covariance comes from the equation above. Training uses MGEX network stations with known coordinates, processed in simulated kinematic mode as in art_03d47f8d (Validation and Assessment of Multi-GNSS Real-Time Precise Point Positioning in Simulated Kinematic Mode Using IGS Real-Time Service). The loss is the negative log-likelihood of filter innovations plus position error against the known coordinates. Gradients flow through the filter update, which is short enough to unroll. The network is trained on corrections from a subset of analysis centers.

At test time there are no labels. When the stream changes, the per-group causal statistics re-estimate on their own. Each group's post-normalization gain and bias are then re-matched to that group's stored source moments, following art_89b5a9f6 (Revisiting Layer Normalization for Point Cloud Test Time Adaptation), with moments accumulated over a trailing window instead of a batch. The analysis has two parts. First, we measure how much of the switch-induced degradation the exported scale alone recovers, and how much the learned shape correction adds on top. Second, we check whether grouping by orbit type captures the geostationary and inclined-orbit penalty reported in art_5f62a945 (Precise Orbit and Clock Products of Galileo, BDS and QZSS from MGEX Since 2018: Comparison and PPP Validation).

Baselines

The naive reference is the conventional filter with elevation-dependent variance $\sigma_0^2/\sin^2(\theta)$ and fixed per-constellation prior ratios, as used in the precise-positioning evaluations in the file. For CLK93-corrected multi-GNSS simulated kinematic positioning, art_03d47f8d (Validation and Assessment of Multi-GNSS Real-Time Precise Point Positioning in Simulated Kinematic Mode Using IGS Real-Time Service) reports convergence to 20 cm in under 15 minutes, 3.0–4.0 cm horizontal accuracy and 5.0–7.0 cm 3D accuracy. We will reproduce this on our station set as the anchor. It gets the same stations, days and streams, and its prior ratios are tuned on the validation stations with the same budget as our meta-parameters.

The strongest non-learning competitor is a classical adaptive filter. It re-estimates one measurement variance per constellation group from its own recent innovations, a windowed variance-component estimate, with the window tuned on validation data. It gets identical inputs, grouping and switch events. It is the most dangerous baseline because our exported scale is essentially this estimator. We found no reported result on Lacuna for this estimator under analysis-center switching, so we will establish the number ourselves. Beating it means lower post-switch time-to-recover and lower vertical RMS on held-out analysis centers, with the difference significant across stations.

The strongest learned competitor is a weighting model with the same backbone and training, but with default layer normalization and global windowed input normalization. This is how the Transformer inside art_6c441fe0 (Deep Reinforcement Learning with Robust Spatial-Temporal Representation for Improving GNSS Positioning Correction) is normalized, as far as the file describes it. Its reported gains, 51.2% over a Kalman filter on GZGNSS and 6.5% on the Google Smartphone Decimeter Challenge, come from a different output design (direct position corrections by a policy). We will not claim to beat those numbers. Instead we will run its graph-Transformer representation as a second backbone variant on GZGNSS and the Smartphone Decimeter Challenge, under the same compute and tuning budget.

The ablations isolate the key ingredient. We will discard the exported scale, so the filter uses a fixed prior times the network output. We will pool the normalization across all constellations instead of grouping. We will replace causal with global statistics. We will remove the test-time gain/bias re-matching. And we will freeze the forgetting factor.

Why not just…?

Why not just use a classical per-constellation adaptive variance estimator? This is the real threat, and it may capture most of the constellation-level scale effect. Our design concedes that point by making that estimator the scale path. The question the proposal answers is whether the learned shape correction adds value beyond it: within-constellation structure such as individual geostationary satellites, midnight orbit discontinuities, correction-age effects, and occlusion. If the ablation that removes the network matches the full model on held-out analysis centers, we will report that the normalization insight reduces to classical estimation and that the learned layer is unnecessary for stream switching.

Why not just feed the analysis-center identity as an input? Conditioning works for producers seen in training and cannot help with an unseen one or a degraded day. Our held-out-analysis-center split tests exactly this. We will also include the ID-conditioned model as an extra baseline on seen centers, where it may well tie us.

Why not just fine-tune the network on the new stream? Fine-tuning needs ground-truth coordinates, which a moving receiver lacks in real time. It is also not instant. We will still run label-using fine-tuning on static stations as an upper bound, to show how much of the gap backprop-free adaptation closes.

Proposed Experiments

The decisive experiment is a stream-switch test. We will use real-time correction streams from the producers evaluated in art_6ee624fe (Assessment of Multiple GNSS Real-Time SSR Products from Different Analysis Centers) over MGEX stations. Models are trained with some producers and tested with others held out, and each switch is injected mid-arc. The metrics are time-to-recover to pre-switch 3D error, and horizontal and vertical RMS over the following hour. Calibration is measured as the average normalized estimation error squared, which should sit near the state dimension if the variances are honest. We will also report the ambiguity fixing rate, following art_e4a7e766 (Deep Learning-Based Diffraction Identification and Uncertainty-Aware Adaptive Weighting for GNSS Positioning in Occluded Environments).

A second test uses the 17-station BeiDou-3 dataset of art_378e3083 (Performance Evaluation of Precise Point Positioning for BeiDou-3 B1c/B2a Signals in the Global Range). It asks whether grouped scale reduces the vertical deficit relative to GPS, against that paper's kinematic BeiDou-3 figures of 2.6/6.0/8.5 cm north/east/up. A third test checks transfer to degraded environments, running the same models on GZGNSS and the Smartphone Decimeter Challenge. We will also run low-cost receiver traces in the regime of art_70d6410d (Assessment of the Real-Time and Rapid Precise Point Positioning Performance Using Geodetic and Low-Cost GNSS Receivers), where convergence to 0.1 m exceeds two hours, to see whether adaptive weighting shortens it.

The claim fails if the classical adaptive estimator matches the full model on held-out switches. It also fails if the scale-discarding ablation performs as well as the full model, since that would mean the network recovers scale from other features and the normalization dependency we hypothesize does not exist.