PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep-learning source localization using multi-frequency magnitude-only data

Haiqiang Niu, Zaixiao Gong, Emma Ozanich, Peter Gerstoft, Haibin Wang, Zhenglin Li

arXiv:1903.12319v2physics.ao-pheess.SP

TL;DR

Machine-learning source localization is constrained by small datasets and environmental variation. This paper uses separate ResNet50 classifiers with a two-stage localization strategy, and reports range and depth accuracy on China Yellow Sea experimental data.

  • Problem

    Prior machine-learning localization studies used small datasets, while collecting recordings across source positions is impractical and environmental variation can degrade performance.

    Source: §I · p. 1

  • Method

    The method uses separate 50-layer ResNet classifiers for range and depth with coarse interval selection followed by within-interval estimation.

    Source: §I · p. 2, C. Localization algorithm · p. 4

  • Results

    88.8% of range predictions had absolute error below 1.5 km, and 83.8% of depth estimates had error within 10 m.

    Source: §VI · p. 12

  • Takeaways & Limitations

    Experimental data further verified localization performance in uncertain environments, including the China Yellow Sea validation reported by the paper.

    Source: §VI · pp. 11–12

    Time-varying water-column sound-speed profiles were not considered, and significantly different environments such as water depth require retraining.

    Source: C. Limitations and future work · p. 11

Abstract

from arXiv · show

A deep learning approach based on big data is proposed to locate broadband acoustic sources using a single hydrophone in ocean waveguides with uncertain bottom parameters. Several 50-layer residual neural networks, trained on a huge number of sound field replicas generated by an acoustic propagation model, are used to handle the bottom uncertainty in source localization. A two-step training strategy is presented to improve the training of the deep models. First, the range is discretized in a coarse (5 km) grid. Subsequently, the source range within the selected interval and source depth are discretized on a finer (0.1 km and 2 m) grid. The deep learning methods were demonstrated for simulated magnitude-only multi-frequency data in uncertain environments. Experimental data from the China Yellow Sea also validated the approach.

I. INTRODUCTION

The paper addresses sparse training data and environmental mismatch in single-sensor ocean source localization. It proposes ResNet50 classifiers using magnitude-only multi-frequency data, large replica-based training sets, and a two-step strategy for wide-region localization.

I. INTRODUCTION

  • The authors report prior machine-learning localization relied on small datasets (1,2,7–13), while exhaustive recordings across source positions are impractical and environmental variation can degrade performance.Historical observed datasets may be unavailable for some applications.Source: §I · p. 1
  • Acoustic-model replicas provide large training sets that reflect environmental uncertainty, and the resulting DNNs are reported to locate sources in various environments.The paper states that the training data comprise tens of millions of samples.Source: §I · pp. 1–2
  • The approach uses separate 50-layer ResNet classifiers for range and depth, with multi-frequency pressure magnitudes from one sensor as localization input.ResNet uses identity shortcuts and learns residual functions to support very deep structures.Source: §I · p. 2
  • A two-step training strategy is proposed to alleviate training difficulty for wide-region source localization.The paper frames the strategy as a response to the difficulty of training for wide-region localization.Source: §I · p. 2

II. LOCALIZATION BASED ON DEEP LEARNING

The localization pipeline applies ResNet50 to broadband data from one receiver. Its preprocessing forms a multi-frequency magnitude feature vector and normalizes it, including segment-wise normalization for slowly varying source magnitudes.

II. LOCALIZATION BASED ON DEEP LEARNING

  • ResNet50 solves source localization from broadband data recorded at one receiver.The section introduces data preprocessing, network structure, and the localization algorithm.Source: §II · p. 2

A. Input data preprocessing

  • The sensor pressure is transformed into a frequency-domain vector, and the input uses pressure magnitudes across F frequencies.The pressure model includes a source term, Green’s function, and noise.Source: A. Input data preprocessing · p. 2
  • Min–max normalization scales the frequency-magnitude vector so each element lies in [0, 1].The normalized vector is used to facilitate DNN training.Source: A. Input data preprocessing · p. 2
  • For slowly varying source magnitudes, each 20-frequency segment is normalized to its own maximum before forming the DNN feature vector.Values of segment length nf from 10 to 25 gave similar validation performance; this piecewise normalization is limited to the slowly varying-magnitude scenario considered.Source: A. Input data preprocessing · p. 2

B. Residual neural network

The section describes how convolutional and residual-network design supports deep feature learning, then specifies the ResNet50 adaptations used for localization.

B. Residual neural network

  • Convolutional filters learn spatial features, while weight sharing reduces unknown parameters compared with fully connected networks.The filters are smaller than the inputs, allowing them to capture small-scale features.Source: B. Residual neural network · p. 2
  • Identity shortcuts let ResNet model the residual F(x) in H(x) = F(x)+x, addressing vanishing-gradient training saturation.The paper describes shortcut connections as adding identity mappings to convolutional outputs.Source: B. Residual neural network · p. 3
  • The bottleneck block stacks 1×1, 3×3, and 1×1 convolutions with 64, 64, and 256 filters, reducing and then enlarging output dimensions.The three-layer structure is stacked to form common ResNet models.Source: B. Residual neural network · p. 3, Figure 2 · p. 3
  • For localization, the 50-layer ResNet replaces 2-D with 1-D convolutions and removes maxpooling because the inputs are small.The paper uses 50-layer ResNets given the size of the localization data sets.Source: B. Residual neural network · p. 3

C. Localization algorithm

The localization algorithm first assigns a coarse range interval, then estimates range and depth within that interval using separate classifiers.

C. Localization algorithm

  • A first ResNet50 classifies each input into one of four 5-km range intervals spanning 1–20 km.The intervals are [1,5), [5,10), [10,15), and [15,20].Source: C. Localization algorithm · pp. 3–4
  • For each selected interval, separate ResNet50 classifiers estimate fine-grid range and depth, using eight second-step models overall.One range model and one depth model are trained for each of the four intervals.Source: C. Localization algorithm · p. 4
  • The classification labels are binary vectors, with range labels designed differently for the coarse and fine tasks..Source: C. Localization algorithm · p. 4

D. Performance metric

The section defines separate metrics for the two localization stages and describes the simulated environment and data generation used for evaluation.

D. Performance metric

  • First-step performance is measured by classification accuracy: correctly classified range intervals divided by all samples.The metric applies to the coarse range-interval classification task.Source: D. Performance metric · p. 4
  • Second-step prediction performance is measured by mean absolute error between predicted and ground-truth range or depth.The metric applies to either source parameter.Source: D. Performance metric · p. 4

A. Environmental model and data sets generation

  • The simulation uses a shallow-water waveguide with a measured winter sound-speed profile, sediment, a fluid halfspace basement, and a receiver 0.2 m above the bottom.Source frequencies span 100–200 Hz in 1-Hz increments.Source: A. Environmental model and data sets… · p. 4, Figure 4 · p. 5
  • KRAKEN generates training, validation, and test sets; environmental parameters are sampled uniformly, yielding 139392 possible environments at 0.1-km range and 2-m depth resolution.Basement P-wave attenuation is fixed at 0.05 dB/λ and basement density is fixed to reduce problem complexity.Source: A. Environmental model and data sets… · pp. 4–5

B. Sensitivity analysis

Sensitivity analysis identifies which waveguide parameters most shape the input features and informs the training-set design. The localization pipeline uses ResNet50 models in two steps, with overlapping range intervals and finer range and depth grids in the second step.

B. Sensitivity analysis

  • The input feature is the modal interference pattern, and parameter sensitivity is evaluated by mean squared error against a baseline feature.A larger MSE indicates greater sensitivity; the baseline holds the other waveguide parameters fixed while one parameter varies.Source: B. Sensitivity analysis · p. 5
  • Source range and depth, water depth, sediment thickness, and sediment sound speed are more sensitive than other parameters; basement density and attenuation vary little.The training set therefore includes more possibilities for the more sensitive parameters.Source: B. Sensitivity analysis · p. 5, Figure 5 · p. 6

C. ResNet50 model parameters

  • ResNet50 first selects a range interval, then four model pairs estimate range and depth on finer grids.The second-step grid uses 0.1 km range and 2 m depth increments, with overlapping ranges to improve localization robustness.Source: C. ResNet50 model parameters · pp. 5–6

D. Results

The two-step models estimate source range and depth with low errors across four test environments outside the training parameter set, under noiseless and constant-source-magnitude conditions. Range-interval classification is also accurate, and minimum-validation-loss models are selected for final prediction.

1. Training and validation errors

  • Final range and depth models were selected at minimum validation loss; the range model required more training steps because it determined more classes.The representative second-step models used 10 epochs for range and 2 for depth.Source: §1 · p. 6, Figure 6 · p. 7

2. Test data

  • The four test environments were excluded from training parameter combinations, and evaluation considered environmental variation with noiseless data and constant source magnitude.Water depth and sediment parameters remained within the training bounds.Source: §2 · p. 7
  • Range-interval classification accuracy was 97.9%, 97.9%, 94.8%, and 95.8% across Cases 1–4.Misclassifications occurred at interval junctions; overlapping intervals in the second step reduce their effect on localization robustness.Source: §2 · p. 7, Figure 7 · p. 7
  • Range MAE was 0.08, 0.14, 0.15, and 0.11 km, while depth MAE was 0.4, 0.2, 0.2, and 0.5 m across Cases 1–4.The results support low-error range and depth estimation across different bottom parameters.Source: §2 · pp. 7–8, Figure 8 · p. 8

3. Source magnitude

Piecewise normalization reduces the effect of slowly varying source magnitudes on localization. In noisy tests, prediction errors decrease as SNR rises, with the authors reporting convincing predictions at SNRs of at least 15 dB.

3. Source magnitude

  • Piecewise normalization transformed the frequency spectrogram to be independent of source magnitude and reduced training–test magnitude differences, resulting in more accurate localization.The demonstration used slowly varying source magnitude and the 100–200 Hz band.Source: §3 · p. 8, Figure 9 · p. 9

4. SNR

  • At 5 dB SNR, range MAE was 1.09 km and depth MAE was 20.0 m; at 20 dB, they were 0.02 km and 0.13 m.Range-interval accuracy reached 95.0% at 15 dB, and the authors state that SNR ≥15 dB is necessary for convincing prediction; training data were noiseless while test data were noisy.Source: §4 · p. 8, Figure 10 · p. 9

IV. EXPERIMENTAL DATA

The China Yellow Sea experiment used one hydrophone from a seafloor-moored array to record towed-airgun signals. Localization used 100–200 Hz data, with a range-independent assumption for simulated sound fields.

A. Experiment description

  • One hydrophone from the 32-element horizontal array recorded the airgun signals; the array was moored alongside a 16-element vertical array on the seafloor.The experiment took place in the China Yellow Sea on December 18–20, 2011.Source: A. Experiment description · p. 9, Figure 11 · p. 10
  • The towed airgun emitted low-frequency signals at ranges of 3.0–16.0 km, with source positions obtained from ship GPS and source depth measured by a sensor.The ship traveled at about 2 m/s; water depth along the path varied from 68 to 73 m.Source: A. Experiment description · p. 9
  • Localization used the 100–200 Hz band at 1 Hz increments, and simulated sound fields were generated under a range-independent assumption.The 80 pulses were recorded over 95 minutes, each with a 2-second duration.Source: A. Experiment description · p. 9, Figure 11 · p. 10

B. Results

On experimental data, deep learning estimates range more accurately than SAGA and has comparable depth performance. Pre-trained models also predict much faster than SAGA.

B. Results

  • Deep-learning range predictions had a 0.70 km MAE versus SAGA’s 1.54 km; depth MAEs were 7.7 m and 8.1 m, respectively.Most depth predictions fell within 2–16 m, consistent with the measured 5–8 m towing depth; depth performance degraded at far ranges due to low SNR.Source: Figure 12 · p. 10, B. Results · pp. 10–11
  • 88.8% of range predictions had error below 1.5 km, versus 78.8% for SAGA; depth predictions below 10 m were 83.8% and 87.5%, respectively.The paper characterizes range estimation as better and depth prediction as comparable to SAGA.Source: Figure 13 · p. 11, B. Results · p. 11

A. Computation time

  • Pre-trained models predicted one sample in about 125 ms on one CPU core, compared with about 600 s for SAGA.Training took 6 days for ResNet50-1, 3 days per ResNet50-2-x-D model, and 15 days per ResNet50-2-x-R model.Source: A. Computation time · p. 11

B. Extension to low-SNR and multi-source cases

Experimental results support localization in uncertain environments, while the proposed low-SNR and multi-source extension relies on array processing. The paper also identifies boundaries in the simulated environments and water-depth range.

B. Extension to low-SNR and multi-source cases

  • For low-SNR or multiple sources, the proposed route uses a horizontal array for DOA estimation and spatial filtering before localization.The paper describes this extension using conventional array signal processing.Source: B. Extension to low-SNR and… · p. 11

C. Limitations and future work

VI. SUMMARY

  • The method uses single-sensor absolute pressure and synthetic acoustic-model data, first selecting a range interval and then estimating range and depth with interval-specific models.The two-step strategy is proposed to ease training the deep models with big data.Source: §VI · p. 11
  • 88.8% of range predictions had absolute error below 1.5 km, and 83.8% of depth estimates were within 10 m in experimental validation.The reported simulated tests also performed well across various environments, slowly varying source magnitudes, and high SNRs.Source: §VI · pp. 11–12
Loading 1903.12319v2…