Ambient-Aware Dither for One-Bit Hydrophone Arrays

Low-power underwater sensors could run for months if each hydrophone fed a one-bit converter, a comparator that records only whether the pressure is above or below a threshold, instead of a power-hungry 16-bit converter. The obstacle is that a lone comparator throws away almost everything about a weak signal. Stochastic resonance helps here. The term names the effect where adding the right amount of random noise before a threshold makes a sub-threshold signal show up in the pattern of threshold crossings. In the ocean, noise is not something the designer has to add. It is already there, and it is loud, impulsive (snapping shrimp, surface impacts), and correlated across nearby hydrophones (distant shipping, surface waves). Whether a one-bit array should add noise, spread its thresholds, or do neither therefore depends on ambient conditions that drift over hours.

We propose a one-bit hydrophone array that estimates the ambient noise field from its own output bits and, online and without labels, sets each element's threshold offset and injected-noise level to maximize information about the target signal. The central scientific claim runs against the textbook result. Theory for arrays of threshold units says that once noise is strong enough, every unit should share the same threshold. We argue that under spatially correlated ocean noise this prescription is wrong, and that spreading the thresholds deterministically beats both identical thresholds and added random noise. The practical claim is that at a matched output bit rate, this controller closes most of the gap to full-precision processing on real ship-radiated recordings.

Why this is the right intersection, and what is already done

Stochastic resonance only helps when there is an information bottleneck, meaning a nonlinearity the signal must pass through. Applying it as software preprocessing to recordings already digitized at high resolution cannot add information, because a noisy transform of stored data carries no more than the data itself. We therefore avoid the tempting idea of an "SR front-end before a deep classifier." The Gap 5 framing in our research file points toward that idea, and the hyperspectral contrast gains of 3D dynamic stochastic resonance do not transfer to it. A one-bit converter is a genuine bottleneck, so it is the setting where the physics has real leverage.

Three lines of prior work bear on this. The only direct underwater crossover is Mixing and Normalizing Stochastic Resonance for beacon detection. It down-mixes a 37.5 kHz locator pulse into the low-frequency range where a bistable (double-well) system works, then tunes the well by hand, and it was tested only in an anechoic tank. That work motivates ours by showing that SR ideas survive contact with hydrophone hardware. It differs from ours in using a single tone, a single sensor, a dynamic well rather than a quantizer, and fixed parameters.

Two papers supply the one-bit method. Noise-Boosted MUSIC adds Gaussian noise before sign quantization and averages L = 500 realizations, which smooths the quantizer into a differentiable response and lets direction finding approach unquantized accuracy. Its noise level is chosen by offline grid search under white Gaussian assumptions. One-bit MMSE under Gaussian-mixture priors shows that non-Gaussian inputs suffer more one-bit distortion than Gaussian inputs of equal variance, and confirms SR gains at low SNR. That result constrains us: heavy-tailed ocean noise cannot be treated as its Gaussian-equivalent variance.

Two further papers supply the adaptation and the theory we contest. Activity-driven adaptive stochastic resonance shows that a neuron can tune its internal noise variance from its own firing statistics. We adopt that label-free, output-statistics style of adaptation, but control both threshold spread and noise level and account for cross-element correlation. Suprathreshold SR in threshold arrays proves that optimal thresholds collapse to a single value beyond a noise level of about σ ≈ 0.91 for Gaussian signal and noise. That proof assumes independent noise across elements, and this is exactly the assumption the ocean violates. Lévy-noise SR shows that the tail index of heavy-tailed noise changes resonance behavior. This motivates fitting a heavy-tailed noise model rather than a variance alone.

Finally, the stochastic Gaussian classifier for ship recognition is the nearest underwater work that uses "stochasticity" at all. Its randomness lives in the classifier weights, and its inputs come from full-precision audio. It serves as our full-precision reference, not as a competitor for the same idea.

Core Claim

Consider an array of K one-bit elements. Element k outputs $b_k(t) = \mathrm{sign}(x_k(t) + d_k(t) - \theta_k)$, where $x_k$ is the hydrophone pressure, $d_k$ is injected noise of standard deviation $\sigma_d$, and $\theta_k$ is that element's threshold. For a weak signal amplitude $s$ and independent elements, the information the bits carry about $s$ is

$$J(s) = \sum_{k=1}^{K} \frac{f_k(\theta_k - s)^2}{F_k(\theta_k - s)\,[1 - F_k(\theta_k - s)]}$$

where $F_k$ and $f_k$ are the cumulative distribution and density of the effective noise (ambient plus injected) at element k. Correlated ambient noise makes elements with identical thresholds flip together, so their bits become redundant and the true information falls below the sum. Deterministic threshold spread removes that redundancy without widening $f_k$. Injected noise also removes it, but at the cost of flattening $f_k$.

We therefore predict three regimes. In quiet, near-independent noise, added noise helps, which is the classical stochastic resonance regime. In loud, independent noise, neither added noise nor threshold spread helps, and the identical-threshold result holds. In correlated noise, threshold spread helps and added noise hurts. A controller that tells these regimes apart from the bits alone should dominate every fixed design.

Research Program

The controller has three parts, all computable from comparator outputs.

The first part estimates the per-element marginal. The fraction of ones as a function of a slow known threshold sweep traces out $F_k$. We fit a Gaussian scale mixture to it so that impulsive tails are captured.

The second part estimates spatial correlation from bit agreement. For zero-mean Gaussian inputs, the arcsine law relates input correlation to bit correlation through $\rho_{ij} = \sin(\tfrac{\pi}{2}\mathbb{E}[b_i b_j])$. We correct this relation for the fitted mixture by table lookup.

The third part is a slow optimizer. It chooses the threshold spread and $\sigma_d$ to maximize a correlated-noise version of $J$, computed from the fitted model with a Gaussian-copula approximation for pairwise redundancy. The update is gradient-like and rate-limited, in the spirit of the activity-driven rule.

Downstream, the decimated bit streams are either reconstructed to a waveform by Bussgang linear estimation, following the linear estimators analyzed in the MMSE paper, or fed directly to an unchanged classifier.

We also run a theory arm. We extend the threshold-bifurcation analysis of the SSR paper to equicorrelated Gaussian noise and to Student-t noise, using the same simulated-annealing search. This produces a phase diagram showing where the optimal thresholds collapse to one value as a function of correlation and tail index.

Data and emulation

Our recordings are high resolution, so the one-bit bottleneck is emulated exactly by applying comparators in software to the raw waveform. This is a legitimate test of a hardware bottleneck, unlike SR applied as post-processing.

For ship recognition we use DeepShip (about 47 hours, four vessel classes) and ShipsEar (11 vessel classes plus a background class). DeepShip and ShipsEar are single-hydrophone datasets, so we build virtual arrays. Each element receives the same ship signal with small inter-element delays, plus ambient noise. The noise is drawn from ShipsEar background recordings and mixed across elements to reach a controlled correlation ρ ∈ {0, 0.3, 0.6, 0.9}, at SNRs down to −10 dB. We report the injected impulsive fraction separately.

For communication-signal detection and modulation recognition we pass synthetic PSK, FSK, DSSS, and OFDM bursts through measured Watermark NCS1 and AUVFest07 channel responses. We use the Danjiangkou multi-range recordings if they are obtainable. We state plainly that spatial correlation in these experiments is synthetic. The file contains no multichannel field dataset, so a field array trial is the natural follow-up and falls outside this proposal's evidence.

Baselines

The upper reference is full-precision processing: the same classifier on the same clips with no quantization. For DeepShip the strongest reported full-precision system is the SSAST-plus-CNN model with a stochastic Gaussian classifier, which reports 88.48% accuracy and 77.33% at −5 dB SNR, against about 68% for a standard ResNet (hybrid local-global representation learning). DART-MT reports 98.80% with 90% of labels (semi-supervised DART-MT). For modulation recognition, DFRT reports 95.84% at 15–18 km on Danjiangkou (dual-feature ResNet–Transformer). We will not reimplement these heavy models for every condition. Instead we fix one mid-sized spectrogram ResNet and retrain it identically on the output of each front-end. The published numbers anchor the full-precision ceiling, and our own ResNet provides the matched comparison. We found no reported result on Lacuna for any one-bit or quantized-input system on DeepShip, ShipsEar, or Danjiangkou.

The naive reference is plain one-bit sign quantization at zero threshold with no added noise.

The strongest existing method in this family is the noise-boosted quantizer of Noise-Boosted MUSIC. It uses Gaussian noise at an offline grid-searched level with identical thresholds, and we grant it an oracle grid search per condition. That is a stronger version than a deployable one.

A second strong competitor is the identical-threshold SSR array, which the SSR theory predicts is optimal at high noise.

A third competitor is a sigma-delta modulator, the standard low-power one-bit converter with feedback noise shaping, run at the same total output bit rate.

A fourth competitor is a uniform 2- to 4-bit converter at the same bit rate, obtained with fewer elements or a lower sample rate.

The key ablation isolates our main ingredient. It uses a deterministic, fixed, uniformly spread set of thresholds with no added noise and no adaptation.

Fairness rules apply to every comparison. All front-ends are matched on total bits per second. Every tunable baseline gets the same per-condition tuning budget on a held-out validation split. The classifier architecture, training schedule, and seeds are identical across front-ends.

Beating the baselines means two things across the correlated-noise and impulsive conditions. Our controller, running label-free and online, must exceed the oracle-tuned noise-boosted quantizer and the identical-threshold array by a margin larger than the seed-to-seed spread. It must also recover at least half of the accuracy gap between naive one-bit and full precision at −5 dB. If it only matches the oracle-tuned baseline, the contribution shrinks to "adaptation without labels." That is still useful, but it is a smaller result.

Why not just…?

Why not just use a high-resolution converter? On a mains-powered platform this is the right answer, and the full-precision baseline measures exactly what is given up. The one-bit and noise-boosted work (Noise-Boosted MUSIC, one-bit MMSE) is motivated by the power and cost of multi-bit converters on battery nodes. The 116.73M-parameter scale of DART-MT likewise shows that current full-precision pipelines are not built for such nodes. Our claim covers only the regime where the bit budget is fixed. We report the accuracy-per-bit curve so a reader can see where multi-bit conversion wins.

Why not just use a sigma-delta converter? This is the most serious threat. Sigma-delta is mature, one-bit, and shapes quantization noise out of band. It depends on a high-rate feedback clock per channel, and it gains nothing from array diversity. Our design is open-loop and parallel, and it exploits cross-element noise. The matched-bit-rate comparison settles the question empirically. If sigma-delta wins across conditions, we will report that the useful outcome is the correlated-noise phase diagram rather than the converter.

Why not just spread thresholds and skip noise entirely? Classical dither theory suggests that known deterministic offsets often beat random noise. Our own analysis predicts this outcome under correlated noise. The fixed-spread ablation is included precisely to test it. If fixed spread matches the adaptive controller everywhere, the adaptive part is unnecessary and the paper becomes a negative result about noise injection in the ocean. We would state that result plainly.

Why not just train the classifier on one-bit inputs with noise augmentation? Augmentation can teach a network to read degraded bits, but it cannot restore information the comparators never captured. The front-end sets an information ceiling that the downstream model cannot exceed. Every front-end is paired with a classifier retrained on its own output, so augmentation is already built into all arms. The comparison therefore isolates the front-end.

Decisive Checks

The first check is the regime test. On ShipsEar background segments sorted by level, the benefit of injected noise should be positive in the quietest tercile and near zero or negative in the loudest. The controller should reduce its added noise to zero in the loud tercile without being told. If added noise helps uniformly, the ambient-awareness premise is wrong.

The second check is the correlation test. At ρ ≥ 0.6, spread thresholds should beat identical thresholds in bit-derived Fisher information and in classification accuracy, contradicting the collapse to identical thresholds that the independent-noise SSR theory predicts. The simulated phase diagram should show where that collapse point moves as correlation and tail index change.

The third check is the estimation test. Bit-derived correlation estimates should track the true synthetic ρ, and the fitted tail index should track the impulsive fraction. If they do not, a label-free controller is not viable.

The fourth check is the matched-budget accuracy test on DeepShip at −5 and −10 dB against sigma-delta, multi-bit, and the oracle noise-boosted quantizer. It is the headline number, reported with confidence intervals over five seeds.