Censored Path-Loss Generators for Wi-Fi Fingerprint Augmentation

Wi-Fi fingerprinting locates a phone indoors by matching the vector of signal strengths it hears from hundreds of access points against a database of vectors recorded at known spots. On UJIIndoorLoc, the standard three-building, multi-floor campus benchmark, the main limit is now the database rather than the model. Whole wings are unsurveyed, a few users and phones dominate the records, and the validation set was recorded months later on different devices. The best existing augmentation for this benchmark is a multi-output Gaussian process (a statistical interpolator that predicts all access points' signals jointly as a smooth function of position). It cut 3D error to 8.42 m. It also has three known flaws. It knows nothing about how radio signal decays with distance and floors. It samples positions inside rectangular bounding boxes that extend past the real building outline. And it produces a continuous signal value for every access point at every point, while real phones simply fail to hear anything weaker than their sensitivity floor.

We propose a fingerprint generator built around a small physical model fitted blind, without any floor plan or access point coordinates. The model has a hidden location, transmit power, and decay rate for each access point, a loss for each floor slab crossed, a gain offset and detection threshold for each phone model, and a learned spatially correlated residual for everything the physics misses. Synthetic fingerprints are drawn at positions inside a building outline estimated from the data, for phone models the database underrepresents. The generator then applies the phone's detection threshold, so it outputs the same "not heard" pattern that real data contains. The bet is that this structure helps most where interpolation is weakest: in unsurveyed regions and on unseen devices.

Core Claim

A generator whose mean is a censored, device-aware path-loss model will improve downstream localization on UJIIndoorLoc more than the purely statistical multi-output Gaussian process augmentation of Multi-Output Gaussian Process-Based Data Augmentation for Multi-Building and Multi-Floor Indoor Localization and Multi-Dimensional Wi-Fi RSSI Data Augmentation Based on Multi-Output Gaussian Process. We expect the largest gains on held-out spatial regions and held-out phones, and little or no gain on densely surveyed areas. Our claim rests on three ingredients: the physical mean, the censoring of weak signals, and the device terms. A deep generator does not need to carry it, and we treat the choice of residual model as an ablation.

The received strength from access point $j$ at position $p$ (planar coordinates plus floor $f$) on phone model $d$ is modeled as

$$ r_{jd}(p) = P_j - 10\,n_j \log_{10}\lVert p - a_j \rVert - L\,|f - f_j| + g_d + s_j(p) + \epsilon . $$

Here $P_j$ is the reference power, $n_j$ the path-loss exponent, $a_j$ and $f_j$ the hidden access point position and floor, $L$ the per-slab attenuation, $g_d$ the phone gain offset, $s_j$ a spatially correlated shadowing residual, and $\epsilon$ measurement noise. A reading is recorded only if $r_{jd} > \tau_d$, the phone's detection threshold. Otherwise it is stored as "not detected," which UJIIndoorLoc codes as 100. We fit all parameters by maximizing a censored (Tobit-style) likelihood. Detected readings contribute a density term. Non-detections contribute the probability of falling below $\tau_d$, so they carry information instead of being imputed to an arbitrary constant.

Research Program

The first component is blind physical calibration. UJIIndoorLoc withholds access point coordinates and floor plans. Two of the file's sources establish this constraint: the UJIIndoorLoc case study documents the missing information, and RayLoc shows that full ray tracing needs an exact 3D twin of the building, which rules out forward simulation here. Work on cooperative RSS localization with unknown transmit power, path loss exponent, and anchor location supplies the method precedent: these quantities can be estimated jointly from signal strength alone. We fit per-access-point parameters building by building, initializing $a_j$ at the strength-weighted centroid of its detections. One known ambiguity is that transmit power and distance can trade off against each other. We handle it with weak priors on $n_j$ and by sharing $L$ and the phone terms across all access points. We report how stable the fitted parameters are across random restarts and bootstrap resamples, and we treat unstable access points as a finding rather than hiding them.

The second component is the residual and the sampling region. The residual $s_j$ is modeled two ways. One is a multi-output Gaussian process on the physics residuals, which is the existing MOGP machinery with the Matérn 5/2 kernel reported optimal in the prior work, now applied to a physically detrended target. The other is a small conditional diffusion model, a generator that learns to turn noise into samples by gradual denoising, trained on residual vectors together with their detection masks. The residual is the term most likely to encode site-specific multipath, so it is where a flexible generator could matter. Synthetic positions are drawn uniformly inside a per-floor footprint: an alpha shape (a tight, possibly concave outline) fitted to the training coordinates. This directly fixes the out-of-building sampling that the MOGP authors acknowledge.

The third component is targeted sampling. Augmentation is spent where the data is thin. Positions are oversampled in footprint cells with few real records. Phone identities are drawn from a distribution flattened toward rare phones, with $g_d$ and $\tau_d$ taken from the fitted values for known phones. For a "new phone" condition, we sample offsets from the fitted spread across phones. This is a physical version of the device imbalance that the UJIIndoorLoc case study identifies as a core weakness.

The design also has a constraint from the generative-augmentation literature. Optimizing Synthetic Data for Enhanced Pancreatic Tumor Segmentation found that synthetic samples misaligned with their labels hurt downstream models badly: Dice score fell from 48.48% to 32.49% under boundary noise. For fingerprints, the analogue is a synthetic vector paired with the wrong position. We therefore filter synthetic samples whose likelihood under a held-out real-data model falls below a threshold, and we report results with and without this filter.

Baselines

The naive reference is the same downstream localizer trained on the real UJIIndoorLoc training split only, with the cleaning that the UJIIndoorLoc case study showed matters. That means recoding non-detections to −105 or −110 dBm and dropping records with no detected access point, which it reports lowers RNN 3D error from 9.21 m to 8.54 m. This prevents gains that come purely from cleaning from being credited to augmentation.

The strongest augmentation competitor is the single-building multi-output Gaussian process. It uses a Matérn 5/2 kernel with variance 1 and length scale 10, and doubles the data (augmentation ratio 1). Multi-Dimensional Wi-Fi RSSI Data Augmentation Based on Multi-Output Gaussian Process reports 3D localization error of 8.42 m with a hierarchical RNN versus 8.62 m without augmentation. Multi-Output Gaussian Process-Based Data Augmentation reports 8.42 m and a 93.84% floor hit rate. We re-implement it in GPy with the same kernel and ratio, and feed it to the same localizer.

The strongest non-augmentation result in the file is Linked-DNN with hierarchical stage-wise training, at 8.19 m 3D error and a 100% building hit rate, from Hierarchical Stage-Wise Training of Linked Deep Neural Networks. We use it as a second downstream localizer, so augmentation is tested on both the hierarchical RNN and Linked-DNN.

A noise-based reference uses truncated Gaussian noise injection inside a Mean Teacher semi-supervised setup, from Mean-teacher-based semi-supervised learning framework, which reports a 3.709% to 7.748% relative improvement in mean 3D error. A black-box generator reference is a conditional diffusion model trained directly on raw fingerprints conditioned on position and floor, with no physics.

All methods share the same real data, the same synthetic-sample budget, the same localizer architecture and training schedule, and the same number of hyperparameter trials, tuned on a split carved from the training set, never on validation. Results are averaged over five seeds.

The ablations isolate the key ingredients. We remove censoring, so the physical model produces continuous values and missing entries are coded −105. We remove the phone terms. We replace the alpha-shape footprint with a bounding box. We remove the physical mean, which reduces the GP-residual variant to plain MOGP restricted to the footprint.

Beating the MOGP means a statistically significant reduction in 3D error on the official validation split across seeds, for both localizers. It also requires a clearly larger reduction on the held-out-region protocol described below, because that is where the claim says physics should matter. Matching MOGP on the official split while winning only on held-out regions would still support the core claim. Losing on held-out regions would refute it.

Why not just…?

Why not just restrict the MOGP samples to the building outline? This fixes one of the three flaws at almost no cost, and it is exactly our "physics removed" ablation. If footprint-masked MOGP matches the full model on held-out regions, then the physical mean and censoring are not pulling their weight, and we will report that. Our expectation runs the other way for a specific reason. A stationary Gaussian process reverts to its mean far from data, so inside a large unsurveyed wing it predicts a flat average fingerprint. A path-loss mean keeps predicting decay toward or away from each fitted access point.

Why not just a conditional diffusion or GAN on raw fingerprints? Physics-Informed Generative Modeling of Wireless Channels found black-box generators such as AmbientGAN hallucinated noise and lost the sparse structure of real channels. Physics-Informed Diffusion Models found that data-only diffusion overfit and violated physical residuals by about two orders of magnitude. Neither study used fingerprint data, so this is suggestive, not decisive. The raw-diffusion baseline and the diffusion-residual variant settle it on this benchmark.

Why not just add a path-loss penalty to a diffusion model, in the style of PIDM? A penalty term needs the same hidden access point parameters our model fits, so the blind calibration step is unavoidable either way. Once those parameters exist, sampling the physical mean directly is cheaper and more interpretable than penalizing a generator toward it, and PIDM reports roughly 23% extra training cost for its mean-estimation variant. If the diffusion-residual variant beats the GP-residual one, a learned generator is earning its place on top of the physics.

Why not just collect unlabeled traces and do semi-supervised learning? On UJIIndoorLoc, Mean Teacher gains were modest at 3.7% to 7.7%, and its authors attribute this to coverage gaps that perturbing existing samples cannot fill. Semi-supervised learning is complementary, not competing, so we also report our augmentation combined with Mean Teacher.

Proposed Experiments

All experiments use UJIIndoorLoc, with the official training split for fitting and the official later, different-device validation split for testing. We report mean 3D error, building and floor hit rates, and 75th/95th percentile error. We add a held-out-region protocol: for each building-floor, we remove one contiguous block of reference positions (about 15% of that floor's footprint) from training, train with each augmentation method, and evaluate only on real records inside the removed block. This is the decisive test. Our claim predicts the largest gap over MOGP here, with errors growing more slowly as distance into the void increases.

A device protocol holds out all training records from one or two phone models. It tests whether sampling $g_d$ and $\tau_d$ from the fitted spread reduces error on those phones, and we stratify validation error by phone.

Fidelity is measured independently of localization in three ways. First, a train-on-synthetic, test-on-real localizer. Second, the agreement between synthetic and real distributions of how many access points are detected per fingerprint, which is the direct test of whether censoring matters. Third, a two-sample classifier distinguishing real from synthetic fingerprints at matched positions.

The parameter-stability analysis reports bootstrap spread of fitted access point positions and of $L$. If calibration proves unidentifiable for most access points, the physical mean collapses toward a constant and the method should fail, cleanly, on the held-out-region test. We would publish that negative result.

As a portability check without access point ground truth, we repeat the official-split comparison on the SBUK-D dataset used by CAE-CNNLoc, whose reported 1.73 m error gives a much denser regime. In that denser regime we expect augmentation of any kind to help little.