Source-linked AI summary
From Signals to Evidence: Interpretable Underwater Acoustic Target Sensing with Large Language Models - Part II: Passive Detection
Zhiyuan Li, Yuge Zhang, Jie Chen, Xiruo Su, Zhengqiao Zhao, Xiling Yao, Lingji Xu, Jingdong Chen, Yixin Yang
TL;DR
Passive underwater detection systems often reduce array observations to a binary decision or scalar score, leaving the supporting physical evidence unclear, especially at low SNR. This paper adapts the UPD framework with physics-aware multichannel representations and LLM-based evidence reasoning. Experiments report reliable detection and accurate bearing estimates, while field-recording validation remains future work.
Problem
Existing detection outputs often collapse multichannel observations into binary decisions or scalar scores, while isolated spectral peaks and transient beamforming responses are ambiguous at low SNR.
Source: §1 · pp. 1–2Method
The approach combines physics-aware multichannel acoustic evidence tokens and environmental priors with an LLM that evaluates target-present and noise-only hypotheses and generates structured detection reports.
Source: §1 · p. 2, §4.2.1 · p. 6Results
The experiments report reliable detection through complementary physical cues and multimodal reasoning, with consistent associations among spectral structures, directional responses, temporal persistence, and propagation priors.
Source: §1 · p. 3Takeaways & Limitations
The results provide initial evidence that the proposed approach is feasible for passive acoustic detection using simulated targets under white and mixed backgrounds.
Source: §7 · p. 12Validation with field recordings and evaluation under changing propagation conditions, diverse sensor configurations, and continuous moving-target observations remain future work.
Source: §7 · p. 12
Abstract
from arXiv · showhide
Underwater acoustic passive detection is essential to maritime surveillance and intelligent ocean sensing. Existing detection systems typically collapse multichannel array observations into a binary decision or a scalar confidence score, offering limited insight into the physical evidence that supports or contradicts a target declaration. This limitation is acute at low signal-to-noise ratio (SNR), where isolated spectral peaks and transient beamforming responses are inherently ambiguous and physically verifiable targets are characterized by an ensemble of coupled reinforcing cues. Building on the unified from-signals-to-evidence framework introduced in the companion Part I, this paper instantiates the framework for passive target detection. A physics-aware multichannel acoustic encoder encodes the array observations into complementary acoustic evidence tokens through a direction-guided architecture that combines minimum variance distortionless response (MVDR) processing with time-frequency and fine-frequency representations. These tokens are combined with environmental priors and fed into a pretrained large language model (LLM) that evaluates target-present and noise-only hypotheses and generates a structured detection report detailing target presence, bearing, spectral-line locations, received SNR, range, and per-field confidence scores. A physics-parameterized data engine extended to realistic oceanic channels provides traceable supervision at both discrete-label and structured-description levels. Experiments under white-noise and mixed ocean-background conditions demonstrate reliable detection, accurate bearing estimation, and physically interpretable structured reports, validating the feasibility of evidence-grounded underwater acoustic detection.
1 INTRODUCTION
Passive underwater detection must recover target evidence from multichannel signals degraded by noise, reverberation, and interference, while conventional outputs often reduce observations to a binary decision or scalar score. This paper adapts UPD to detection with physics-aware evidence tokens and LLM reasoning, reporting reliable detection and physically interpretable descriptions.
Motivation
- At low SNR, isolated spectral peaks and transient beamforming peaks are ambiguous, whereas physically verifiable targets exhibit coupled cues such as persistent tonals and directionally consistent phase differences.Noise, reverberation, and competing interferers can mask or distort the spatial and spectral signatures used for detection.Source: §1 · pp. 1–2
Contributions
- The proposed system frames passive detection as physics-grounded multi-evidence hypothesis reasoning, with target declarations supported by explicit temporal, spatial, spectral, propagation, and background evidence.It extends the UPD framework from recognition to passive detection.Source: §1 · p. 3
- A physics-aware multichannel encoder preserves weak structures across time, frequency, and array dimensions and represents them as physically meaningful evidence tokens.Tokens include spectral persistence, temporal continuity, directional concentration, inter-channel coherence, array-manifold consistency, and background descriptors.Source: §1 · p. 3
- The LLM evaluates supporting and conflicting observations and generates structured descriptions of target presence, direction, spectral components, temporal behavior, physical consistency, environment, and uncertainty.The framework conditions evidence tokens on environmental priors and produces a report with explicit confidence bounds.Source: §1 · pp. 2–3
- Extensive experiments report reliable detection through complementary physical cues and multimodal reasoning, with consistent associations among spectral structure, directional response, temporal persistence, and propagation priors.The authors present these results as validating the feasibility of evidence-grounded underwater acoustic detection.Source: §1 · p. 3
2 RELATED WORK
Prior work spans array-based detection and localization, temporal tracking, and audio-language systems for spatial sound description. The paper positions passive underwater detection as a task requiring spatial, spectral, and statistical information to be represented together while retaining physical meaning and numerical accuracy.
2.1 Underwater Passive Detection
- Passive-detection research includes classical beamforming and high-resolution spatial methods, neural DOA and localization models, matched-field processing, and methods for tracking temporal bearing and frequency evolution.The cited approaches address different estimation and sensing tasks.Source: §2.1 · p. 3
- Existing underwater sensing systems often handle detection, DOA estimation, ranging, environmental characterization, and tracking with separate models and task-specific representations.The paper notes that this fragmentation constrains broader scene understanding.Source: §2.1 · p. 3
2.2 Large Audio-Language Models for Acoustic Sensing
- Audio-language systems use feature projection or query-based alignment to connect specialized acoustic representations with language models for audio question answering and description generation.Examples include SALMONN and Q-Former-based approaches.Source: §2.2 · pp. 3–4
- Spatial audio-language studies combine acoustic encoders and language models to describe sound events, directions, distances, and other physical attributes.The present work addresses upstream target detection and parameter estimation, including target presence, location, tonal characteristics, and signal quality.Source: §2.2 · p. 4
3 FROM-SIGNALS-TO-EVIDENCE FRAMEWORK: A BRIEF RECALL
The recalled UPD framework organizes sensing as a feature–evidence–decision chain in which decisions are conditioned on explicit evidence rather than directly on features. Its passive-detection adaptation processes raw multichannel observations and integrates acoustic physics and environmental information more deeply than recognition.
3 FROM-SIGNALS-TO-EVIDENCE FRAMEWORK: A BRIEF RECALL
- The section recalls the UPD framework established in Part I and presents its framework architecture in Fig. 2.The figure caption identifies it as the architecture of the from-signals-to-evidence framework.Source: §3 · p. 4, Figure 2 · p. 5
3.1 Feature–Evidence–Decision Chain
- UPD organizes sensing as an ordered feature–evidence–decision chain, with the decision conditioned on stated evidence rather than directly on features.The framework emphasizes physical grounding and a shared semantic space for acoustic signals, environmental imagery, and textual priors.Source: §3.1 · p. 4
3.2 Multimodal Input Layer
- The input layer accepts acoustic observations, environmental information, and textual physical priors, with new sources incorporated through encoders or extended textual context.Environmental inputs can include sound-speed profiles, bathymetry, sea state, and seabed properties.Source: §3.2 · p. 4
3.3 LLM Reasoning Core and Evidence-Grounded Output
- The LLM processes a unified token sequence and generates an evidence-then-decision description linking each evidence item to a verifiable subset of acoustic features.The framework freezes pretrained parameters, adapts through low-rank modules, and supervises the answer portion of sequences.Source: §3.3 · p. 4
3.4 From Recognition to Detection: What Changes
- Compared with recognition, detection handles raw multichannel observations over extended durations, adds spatial outputs such as bearing, and integrates more physical processing and environmental modeling.Detection must account for noise fields, fading, and channel-induced distortion in addition to how propagation shapes observed features.Source: §3.4 · pp. 4–5
4 DETECTION INSTANTIATION: PHYSICS-GUIDED ACOUSTIC ENCODING
UPD frames passive detection as physics-grounded reasoning over complementary evidence, using direction-guided acoustic branches and task-specific supervision before aligning acoustic and environmental inputs with a language model for structured reporting.
4 DETECTION INSTANTIATION: PHYSICS-GUIDED ACOUSTIC ENCODING
- UPD estimates target presence, bearing, and spectral-line locations from synchronized array recordings, with directional features guiding detection and line prediction.Its three branches are bearing estimation (DOA), target detection (DET), and spectral-line detection (LINE).Source: §4 · p. 5
4.1 Physical Principles of Passive Detection
- Passive detection combines spatial, temporal, and spectral structure because a local energy peak at one sensor provides limited evidence.Propagation delays create inter-sensor phase relationships, while successive windows and fine-spectral statistics preserve temporal change, average strength, and variability.Source: §4.1 · pp. 5–6
4.2 Direction-Guided Acoustic Encoding
- The three branches share observation windows, while their analysis-window lengths set spectral grids and DOA features guide DET and LINE aggregation.DOA supplies window-level queries to DET and recording-level keys and values to LINE; the branches use 0.5 s and 2 s Fourier windows, respectively.Source: §4.2 · p. 6, Figure 3 · p. 6
4.2.1 Bearing estimation branch DOA
- DOA estimates window-level spatial covariance and MVDR directional spectra, then encodes angular profiles into directional features for DET and LINE.Window-level features guide DET, while recording-level features guide LINE.Source: §4.2.1 · p. 6
4.2.2 Target detection branch DET
- DET forms time-frequency features from each sensor’s short-time spectrum, models temporal and frequency axes, and uses directional queries to aggregate those features.A gated residual connection integrates the attention output into window-level directional features, which are averaged over bearings into a temporal sequence.Source: §4.2.2 · pp. 6–7
4.2.3 Spectral-line detection branch LINE
- LINE encodes fine-spectrum power, local prominence, and intersensor magnitude coherence, then fuses frequency features with recording-level directional features.Means and standard deviations across observation windows form six frequency channels for a multiscale frequency CNN.Source: §4.2.3 · p. 7
4.3 Task-Specific Acoustic Representation Learning
- Separate task summaries and prediction heads train detection, bearing, and frequency representations with binary, soft-bearing, and frequency-bin supervision.Frequency supervision balances positive and negative bins and adds weight to high-scoring negatives; target-absent observations use negative-bin terms.Source: §4.3 · p. 7
4.4 Multimodal Alignment and Parameter-Efficient Adaptation
- Acoustic sequences and task summaries are projected into language-model inputs alongside environmental tokens, then adapted in two stages for structured reporting.The language model generates target presence, bearing, range, and received SNR; a separate frequency head supplies line scores for the report.Source: §4.4 · pp. 7–8
5 PHYSICS-PARAMETERIZED DATA ENGINE FOR PASSIVE DETECTION
The data engine combines source synthesis, array propagation, controlled noise mixing, and traceable pairing of observations with source, geometry, environment, and target attributes. Its simulated sources combine tonal, continuous-spectrum, and modulation-spectrum components, while propagation and background construction model sensor observations under distinct noise conditions.
5 PHYSICS-PARAMETERIZED DATA ENGINE FOR PASSIVE DETECTION
- The data engine pairs each array observation with its source composition, geometry, propagation environment, and target attributes.White-noise and mixed-background datasets support learning acoustic representations and structured detection descriptions.Source: §5 · p. 8
5.1 Source Signal Synthesis
- The source model combines tonal, continuous-spectrum, and modulation-spectrum components, with periodic Gaussian pulses representing amplitude fluctuations.Pulse period and width set the fluctuation repetition rate and duration.Source: §5.1 · p. 8, §5.1.2 · p. 8, §5.1.3 · pp. 8–9
5.1.1 Tonal component
- Tonal sources include selected in-band harmonics of a fundamental frequency and independently sampled spectral lines, with harmonic amplitudes decreasing by order.Line frequency, amplitude, phase, and frequency-drift rate vary across the modeled lines.Source: §5.1.1 · p. 8
5.2 Propagation and Background Noise
- Each sensor’s propagated target signal is formed by convolving the source signal with its channel impulse response, using a common time reference across channels.The common reference preserves relative propagation delays across sensors.Source: §5.2 · p. 9
5.2.1 Environment and array geometry
- The simulated environment spans a sloping bathymetry, with the array-center depth set to 1.9 km and sources placed at 96 positions on four concentric circles.The circles have radii of 3, 5, 10, and 25 km; source and receiver depths are both 100 m.Source: §5.2.1 · p. 9
- BELLHOP computes propagation along a radial vertical slice from source to array center, and interpolated responses are RMS-normalized across the multichannel record.This common normalization preserves relative sensor amplitudes and phases.Source: §5.2.1 · p. 9
5.2.2 Background construction and SNR
- The backgrounds comprise independent Gaussian white noise or synchronized low-frequency ocean recordings combined with synthetic high-frequency noise.The measured-background construction uses the retained band through 250 Hz and synthetic noise from 250 to 1000 Hz.Source: §5.2.2 · p. 9
- Received mixtures use prescribed broadband SNR, and background-only observations reuse the corresponding target-present background draw before common RMS scaling.The resulting sampled and batched observations form the encoder input.Source: §5.2.2 · p. 9
5.3 Dual-Granularity Supervision
The study supplies both physics-aware acoustic representation labels and structured acoustic descriptions, then partitions observations by source realization. Its dataset configuration uses multiple target SNRs and background-specific observation schemes.
5.3 Dual-Granularity Supervision
- Physics-aware representation labels supervise target presence, bearing, and spectral-line locations, while structured descriptions target presence, bearing, range, SNR, and frequencies.Range and SNR tasks also receive bathymetry, receiver location, array geometry, and sound-speed parameters as input.Source: §5.3 · p. 9
5.4 Dataset Configuration and Partitioning
- Each 120 s observation is sampled at 2 kHz, and eight source realizations at each of 96 geometries yield 768 source instances across five target-present SNR conditions.The SNRs are 0, −10, −20, −25, and −30 dB.Source: §5.4 · p. 9
- Source realizations and clean waveforms remain disjoint across TRAIN, DEV, CAL, and TEST, while the geometry grid and environment are represented in each partition.Measured-background recordings are divided into nonoverlapping temporal blocks separated by unused intervals.Source: §5.4 · pp. 9–10
- TRAIN uses one background draw per source, TEST uses three draws across all SNRs, and CAL contains 288 background-only observations per background.DEV pairs a −30 dB observation with its background; Table 3 reports the resulting observation counts.Source: §5.4 · p. 10, Table 3 · p. 10
6 EXPERIMENTS
The experiments evaluate target presence, bearing, spectral lines, received SNR, and horizontal range under white-noise and mixed-background conditions. The shared-window Fourier analysis settings are summarized in Table 4.
6 EXPERIMENTS
- The experiments evaluate target presence, bearing, spectral lines, received SNR, and horizontal range under white noise and a mixed background.Source: §6 · p. 10
- Table 4 summarizes Fourier analysis settings within the shared observation windows.Source: Table 4 · p. 10
6.1 Experimental Setup
The implementation uses shared observation windows with separate Fourier settings for directional and fine-frequency processing. After acoustic pretraining, the encoder is frozen and language adaptation proceeds in task-specific stages.
6.1.1 Implementation Details
- DOA observations contain 39 spectral snapshots, and the bearing grid spans 0°–359° in 1° steps; the LINE grid has 2,001 frequency bins over 0–1,000 Hz.Source: §6.1.1 · p. 10
- The acoustic encoder forms 149 local tokens and three task summaries, occupying 152 acoustic-content positions.The detection, bearing, and frequency summaries have dimensions 128, 128, and 256, respectively.Source: §6.1.1 · p. 10
- The language backbone is Qwen3-4B-Instruct-2507, with a 201 × 201 bathymetric grid mapped to 16 tokens and receiver metadata supplied as text.The multimodal prefix contains 169 nontext positions plus instruction and metadata tokens.Source: §6.1.1 · pp. 10–11
6.1.2 Training Protocol
- After acoustic pretraining, the encoder is frozen; language adaptation then trains detection and bearing, frequency, and finally SNR, range, and explanation tasks in sequence.Environmental text and bathymetry are added in the final adaptation step.Source: §6.1.2 · p. 11
6.2 Evaluation Protocol
The evaluation defines measures for target detection, bearing, frequency-line outputs, SNR, range, and report validity. It also specifies the test partitions and decision rules used for scoring.
6.2 Evaluation Protocol
- The test uses disjoint source realizations, fixed model parameters, and 288 target-present observations at each of five SNRs, plus 288 shared target-absent observations.Encoder validation instead uses 96 target-present and 96 target-absent observations per background at −30 dB.Source: §6.2 · p. 11
6.2.1 Target Detection
- Target detection is scored from the generated target_present field, with the encoder declaring a target when presence probability exceeds 0.5.Detection and false-alarm rates use correctly detected targets and false alarms relative to the target-present and target-absent observation counts.Source: §6.2.1 · p. 11
6.2.2 Bearing Estimation
- Bearing error is measured circularly over all target-present observations, with 24 tested reference bearings spaced at 15° intervals.The reported bearing metric is mean absolute circular error.Source: §6.2.2 · p. 11
6.2.3 Frequency Outputs
- Frequency candidates are sigmoid-scored local maxima thresholded at 0.5, separated by more than 2 Hz, and matched one-to-one to reference lines within ±2 Hz.Precision, recall, and F1 are micro-averaged from pooled matched, unmatched candidate, and unmatched reference-line counts.Source: §6.2.3 · p. 11
6.2.4 SNR, Range, and Output Validity
- SNR and horizontal-range MAE use target-present observations with valid numerical predictions; numerical coverage and schema validity are also reported.Schema validity checks scalar field types and target-absent null rules.Source: §6.2.4 · p. 11
6.3 Acoustic Representation Performance
At −30 dB, separately trained encoders show strong detection and bearing results on white-noise and mixed-background validation sets, with different frequency F1 scores.
6.3 Acoustic Representation Performance
- At −30 dB, the white-noise encoder detects 96/96 targets with zero false alarms and 1.26° bearing MAE; the mixed-background encoder detects 95/96 with zero false alarms and 0.22° MAE.Frequency F1 is 96.72% for white noise and 81.79% for mixed background; the separately trained encoders use different background datasets.Source: §6.3 · p. 11
6.4 Detection and Structured Reports
The system produces structured physical fields and frequency outputs, with strong detection and bearing results under the mixed background. Frequency, range, and report-validity results qualify that performance, while examples expose both interpretable outputs and a range error.
6.4.1 Physical Output Performance
- Frequency precision, recall, and F1 are 67.11%, 43.95%, and 53.11%; 10,927 of 24,862 reference lines are matched.Reference-line counts fall from 6,086 at 0 dB to 3,424 at −30 dB because annotation depends on the received spectrum, so lower-SNR F1 involves different reference sets.Source: §6.4.1 · p. 11
- Overall SNR MAE is 0.632 dB and horizontal-range MAE is 8.803 km, while a 7.5 km training-median predictor has lower range MAE of 6.753 km on the same 1,439 observations.SNR and range have numerical outputs for 1,439/1,440 observations overall.Source: §6.4.1 · pp. 11–12
- Across 1,440 targets, the system detects every target with zero false alarms among 288 background observations; overall bearing MAE is 0.208°.Bearing MAE rises to 0.521° at −30 dB; 1,420 of 1,440 bearing outputs equal the reference and 20 differ by 15° across 24 tested bearings.Source: §6.4.1 · p. 11, Table 6 · p. 12
- Absolute received level is not a unique range cue, so useful range information must remain in relative spatial or propagation structure.This constrains how the range field can be physically inferred from the received signal.Source: §6.4.1 · p. 12
6.4.2 Structured-Report Examples
- The four scalar fields satisfy joint type and null-value rules in 1,727 of 1,728 reports; one −30 dB target-present report has null SNR and range.Detection, bearing, and frequency outputs remain available in that report.Source: §6.4.2 · p. 12
- In the positive example, six of the ten highest-scoring frequency candidates lie within ±2 Hz of multiples of a possible 12.2 Hz fundamental, but predicted range is 10 km versus a 25 km reference.The predicted 60° bearing agrees with the reference, and the report presents the proposed spectral relation for inspection.Source: §6.4.2 · p. 12
7 CONCLUSION
The paper presents a physics-guided audio-language approach for passive target detection, using multichannel acoustic representations, environmental information, and traceably supervised simulations. The authors report initial feasibility evidence and identify future validation settings.
7 CONCLUSION
- The proposed encoder combines spatial, temporal, and spectral information from multichannel observations, using directional features to guide extraction of weak acoustic structures.The representations are aligned with a pretrained language model and supplemented with environmental information to produce detection descriptions with target presence and physical attributes.Source: §7 · p. 12
- A physics-parameterized simulation pipeline provides multichannel recordings with traceable supervision for acoustic representation learning and language adaptation.Experiments with simulated targets under white and mixed backgrounds provide initial evidence for passive-detection feasibility.Source: §7 · p. 12
- Future work may examine changing propagation conditions, diverse sensor configurations, continuous observations of moving targets, and field recordings.The authors also identify examining how generated descriptions relate to the acoustic information underlying detection decisions as future work.Source: §7 · p. 12