Seed-Stable Waveform Prototypes for Explaining Non-Line-of-Sight UWB Signals

Indoor radio positioning with ultra-wideband (UWB) radios works by timing how long a pulse takes to travel from a beacon to a receiver. When a wall, pillar or person blocks the direct path, which is called non-line-of-sight (NLOS) propagation, the pulse arrives late along a reflected path. The system then overestimates the distance. Neural classifiers that read the raw received waveform (the channel impulse response, or CIR, meaning the recorded echo profile of one pulse) can now flag such signals with high accuracy. They cannot say why they flagged a signal, though, and RF engineers deploying them in a new building have no way to check whether the model has learned physics or clutter.

A natural fix is a prototype network: a classifier that compares each input to a small set of learned "typical examples" and decides by weighted similarity. Its prototypes can be decoded back into waveforms that an engineer can inspect. This note argues that such explanations are only worth something if they are reproducible. Prototypes are learned centroids, and centroid-based learning is known to depend on where the centroids start. If two training runs with different random seeds yield different prototype waveforms and assign the same test signal to different "reasons," the explanation reflects the seed rather than the data. We propose a prototype autoencoder for 1D CIR waveforms that combines three things. Its compression step is told to preserve the weak first-arriving pulse. Its prototypes are seeded deterministically from the data. Its success criterion includes seed-to-seed agreement of explanations alongside accuracy.

Core Claim

On small-sample UWB NLOS identification, a decodable prototype classifier can match the accuracy of the best black-box CIR network. Our claim is that the extra ingredients make it do so while producing explanations that agree across random seeds far more often than a standard randomly initialized prototype model. Those ingredients are a reconstruction loss weighted toward the first-path region and deterministic, balance-regularized prototype seeding. In short, the accuracy-interpretability trade-off for RF waveforms is mainly a stability problem, and stability can be engineered.

The claim is new along two axes the literature leaves separate. Decodable latent prototypes exist for images. ProtoVAE places class prototypes in a variational autoencoder's latent space, decodes them to pixels, and reports 84.6% on CIFAR-10 against 83.9% for a black-box CNN. It initializes prototypes randomly and never measures whether they recur across runs. VAESim discovers prototypes without labels using exponential-moving-average updates. Its authors report that results depend strongly on the initial prototype matrix, with redundant or dead prototypes when it fails to span the data. That report is the direct motivation for our stability focus. NLOS classifiers exist for waveforms but are opaque or coarse. Lab to Multiscene Generalization trains a three-layer network on raw CIR with anechoic-chamber pretraining and added Gaussian noise. Nonparametric Obstruction Detection classifies six hand-crafted waveform statistics with a least-squares support vector machine. Neither offers case-based explanations, and neither touches initialization.

Research Program

Encoder. The model is a 1D convolutional autoencoder. It compresses a CIR of a few hundred taps into a latent vector of around 8 to 16 dimensions.

First-path weighting. Generic reconstruction loss is dominated by the strong late reflections and can blur the weak first peak that actually carries the true arrival time. This is the risk flagged for plain compression autoencoders such as the rate-distortion sensor autoencoder, and more generally by the data-processing-inequality analysis. We therefore weight reconstruction error by a window $w_t$ centered on the detected first-path tap:

$$\mathcal{L}_{rec} = \sum_t (1+\alpha w_t)\,(x_t-\hat{x}_t)^2 .$$

Here $x_t$ and $\hat{x}_t$ are the input and reconstructed CIR at tap $t$, and $\alpha \ge 0$ sets how much the first-path window matters.

Prototype head. On top of the latent space sits a ProtoVAE-style head with $K$ prototypes per class. Each input's similarity to every prototype feeds a linear classifier, and every prototype can be decoded into a synthetic CIR.

Seeding. We depart from ProtoVAE in how prototypes start. We first pretrain the autoencoder alone. We then place each class's prototypes on that class's latents using farthest-leap selection, a deterministic linear-time seeding rule from Mishra et al.. That rule starts at the class mean and repeatedly jumps to the farthest remaining point. It is cheap in a 16-dimensional latent, which sidesteps the high-dimensional cost that rule suffers on raw data.

Balancing during training. During joint fine-tuning we add an exclusive-lasso penalty on soft prototype assignments. This penalty equals the sum of squared cluster sizes, following Balanced Clustering via Exclusive Lasso, and discourages the lopsided collapse in which one prototype absorbs a whole class. That paper also warns that too strong a penalty forces artificial uniformity onto skewed data. We therefore tune its weight over the range it reports, $10^{-2}$ to $10^2$. We also report assignment-size histograms, so that a forced split of a genuinely dominant mode is visible.

Stability metric. We train $S=10$ seeds per configuration. Prototypes of two runs are matched by Hungarian assignment on the normalized cross-correlation of their decoded waveforms. We report two quantities. The first is prototype recurrence, the mean matched correlation. The second is explanation agreement, the fraction of test CIRs whose most-similar prototype maps to the same matched prototype in both runs, averaged over run pairs. No reported result was found on Lacuna for either quantity on any prototype model, so we compute them for all baselines ourselves.

Physical faithfulness. Separately, we test whether decoded prototypes are physically meaningful. We check whether the first-path delay and excess-delay spread of each decoded prototype match those of its assigned real signals.

Baselines

Naive reference points. The do-nothing reference is the majority-class predictor. The weaker baseline reported in Lab to Multiscene Generalization is the same network without transfer or noise injection, which scores 84.4% in the office corridor and 81.1% in the underground parking lot with $N=50$ target samples. These anchor what "learning anything" means at this data size.

Strongest black-box competitor. This is the lab-pretrained, noise-augmented CIR network from the same paper. It reports 98.8% NLOS identification accuracy in the corridor and 97.1% in the parking lot at $N=50$. We reimplement it and train every model, ours included, on the identical anechoic pretraining set, identical noise augmentation, identical 50-sample target splits, and an equal hyperparameter-search budget. The purpose is to reproduce its numbers before comparing. Beating it on accuracy is not required. The proposal succeeds if it stays within about one percentage point of the reproduced accuracy while giving explanations. We also give this network post-hoc saliency explanations, so its explanation agreement across seeds can be measured on equal footing.

Strongest interpretable competitor. This is the six-feature least-squares SVM from Nonparametric Obstruction Detection, which reports 91% accuracy (false-alarm rate 0.08, miss rate 0.09) on its 1,024-measurement office campaign. We run it on every dataset with the same splits.

Prototype baselines. The standard prototype baseline is ProtoVAE adapted to 1D convolutions with its original random initialization and orthonormality loss. A two-stage baseline, detailed in the next section, is the most dangerous. Three ablations each remove one ingredient: first-path weighting ($\alpha=0$), deterministic seeding (random or k-means++ seeding instead), and the balance penalty. Beating ProtoVAE means a clear gain in explanation agreement, for instance from a level where seeds disagree on a large share of signals to one where most agree, with no accuracy loss.

Why not just…?

Why not just train the black-box CIR network and add saliency maps? Its accuracy is excellent (98.8% and 97.1%) and hard to beat. ProtoVAE, however, measured post-hoc relevance to be much less faithful than its built-in prototypes, with an average-drop score of 3.4 for ProtoPNet versus 0.4 for ProtoVAE on MNIST. A saliency heatmap over taps also cannot say "this looks like a metal-reflector echo." We will measure saliency's seed-to-seed agreement directly. If it turns out as stable as our prototypes and equally informative to engineers, the case weakens, and we will say so.

Why not just use the hand-crafted features, which are already interpretable? Energy, rise time and kurtosis are readable, and on the MIT campaign they reach 91%. They are summary scalars, though. They cannot separate two obstruction types with similar energy but different echo structure, and they sit well below what raw-waveform networks reach in harder sites. The test is whether our prototypes split NLOS signals into subtypes that the six features merge. The chamber metal reflector versus concrete pillars in the parking lot is one such case, checked by clustering purity against known site and material labels.

Why not just run k-means++ on a frozen autoencoder latent and classify by nearest centroid? This is the real threat. It is deterministic up to seeding, simple, and possibly just as stable. We include it as a named baseline with identical encoder and first-path weighting. If it matches our model on accuracy and agreement, then joint fine-tuning adds nothing and the contribution reduces to the first-path-aware latent plus stability evaluation. We would report exactly that. We expect it to lose accuracy, because a centroid classifier cannot reshape the latent to sharpen class boundaries. The experiment settles the question either way.

Why not skip identification and use robust convex localization or inertial fusion? Semidefinite-programming localization without NLOS identification and UWB-inertial fusion address positioning, not explanation. The UTIL dataset reports roughly 10 cm error for error-state Kalman filtering in clean layouts but significant deterioration in its cluttered constellation, with no number given. We treat localization as a secondary check, not the core claim.

Proposed Experiments

The primary experiments use the corridor and parking-lot CIR data of Hua et al. at $N \in \{20, 50, 200\}$ target samples, plus the MIT office campaign. The decisive check is a two-by-two grid of seeding (random versus farthest-leap) crossed with first-path weighting (off versus on). For each cell we measure accuracy, prototype recurrence and explanation agreement over ten seeds. The hypothesis fails if deterministic seeding does not raise explanation agreement at $N=50$, or if our model trails the reproduced black-box accuracy by more than about one point.

The second check targets physical faithfulness. For decoded prototypes with $\alpha>0$ versus $\alpha=0$, we measure the error between each prototype's first-path delay and the median first-path delay of its assigned real CIRs, and whether a linear probe on the latent recovers first-path delay.

The third, exploratory check applies only if raw waveforms and motion-capture ground truth can be paired in the UTIL identification recordings. There we use similarity to line-of-sight prototypes to scale the measurement variance of each beacon in an extended Kalman filter, compared with hard discarding and top-three ranking. It tests whether the explanation signal is also operationally useful. The report states in advance that this experiment depends on data availability and does not bear on the core claim.