Predicting ocean pressure field with a physics-informed neural network.

Ocean sound pressure field prediction, based on partially measured pressure magnitudes at different range-depths, is presented. Our proposed machine learning strategy employs a trained neural network with range-depth as input and outputs complex acoustic pressure at the location. We utilize a physics-informed neural network (PINN), fitting sampled data while considering the additional information provided by the partial differential equation (PDE) governing the ocean sound pressure field. In vast ocean environments with kilometer-scale ranges, pressure fields exhibit rapidly fluctuating phases, even at frequencies below 100 Hz, posing a challenge for neural networks to converge to accurate solutions. To address this, we utilize the envelope function from the parabolic-equation technique, fundamental in ocean sound propagation modeling. The envelope function shows slower variations across ranges, enabling PINNs to predict sound pressure in an ocean waveguide more effectively. Additional PDE information allows PINNs to capture PDE solutions even with a limited amount of training data, distinguishing them from purely data-driven machine learning approaches that require extensive datasets. Our approach is validated through simulations and using data from the SWellEx-96 experiment.

Bayesian optimization with Gaussian process surrogate model for source localization.

Source localization with a geoacoustic model requires optimizing the model over a parameter space of range and depth with the objective of matching a predicted sound field to a field measured on an array. We propose a sample-efficient sequential Bayesian optimization strategy that models the objective function as a Gaussian process (GP) surrogate model conditioned on observed data. Using the mean and covariance functions of the GP, a heuristic acquisition function proposes a candidate in parameter space to sample, balancing exploitation (sampling around the best observed objective function value) and exploration (sampling in regions of high variance in the GP). The candidate sample is evaluated, and the GP conditioned on the updated data. Optimization proceeds sequentially until a fixed budget of evaluations is expended. We demonstrate source localization for a shallow-water waveguide using Monte Carlo simulations and experimental data from an acoustic source tow. Compared to grid search and quasi-random sampling strategies, simulations and experimental results indicate the Bayesian optimization strategy converges on optimal solutions rapidly.

Uncertainty quantification for direction-of-arrival estimation with conformal prediction.

Uncertainty quantification (UQ) of deep learning (DL)-based acoustic estimation methods is useful for establishing confidence in the predictions. This is crucial to enable the real-world applicability of DL-based systems for acoustic tasks. Specifically, it is proposed to use conformal prediction (CP) for UQ in direction-of-arrival (DOA) estimation. CP is a statistically rigorous method to provide confidence intervals for an estimated quantity without making distributional assumptions. With CP, confidence intervals are computed via quantiles of user-defined scores. This easy-to-use method can be applied to any trained classification/regression model if an appropriate score function is chosen. The proposed approach shows the potential to enhance the real-time applicability of DL methods for DOA estimation. The advantages of CP are illustrated for different DL methods for DOA estimation in the presence of commonly occurring environmental uncertainty. Codes are available online (https://github.com/NoiseLabUCSD/ConformalPrediction).

Compressive frequency-difference direction-of-arrival estimation.

Direction-of-arrival estimation is difficult for signals spatially undersampled by more than half the wavelength. Frequency-difference beamforming [Abadi, Song, and Dowling (2012). J. Acoust. Soc. Am. 132, 3018-3029] offers an alternative approach to avoid such spatial aliasing by using multifrequency signals and processing them at a lower frequency, the difference-frequency. As with the conventional beamforming method, lowering the processing frequency sacrifices spatial resolution due to a beam broadening. Thus, unconventional beamforming is detrimental to the ability to distinguish between closely spaced targets. To overcome spatial resolution deterioration, we propose a simple yet effective method by formulating the frequency-difference beamforming as a sparse signal reconstruction problem. Similar to compressive beamforming, the improvement (compressive frequency-difference beamforming) promotes sparse nonzero elements to obtain a sharp estimate of the spatial direction-of-arrival spectrum. Analysis of the resolution limit demonstrates that the proposed method outperforms the conventional frequency-difference beamforming in terms of separation if the signal-to-noise ratio exceeds 4 dB. Ocean data from the FAF06 experiment support the validity.

Deep transfer learning-based variable Doppler underwater acoustic communications.

This paper proposes a deep transfer learning (DTL)-based variable Doppler frequency-hopping binary frequency-shift keying underwater acoustic communication system. The system uses a convolutional neural network (CNN) as the demodulation module of the receiver. This approach directly demodulates the received signal without estimating the Doppler. The DTL first uses the simulated communication signal data to complete the CNN training. It then copies a part of the convolution layers from the pre-trained CNN to the target CNN. After randomly initializing the remaining layers for the target CNN, it is trained by the data samples from the specific communication scenarios. During the training process, the CNN learns the corresponding frequency from each symbol in the selected frequency-hopping group through the Mel-spectrograms. Simulation and experimental data processing results show that the performance of the proposed system is better than conventional systems, especially when the transmitter and receiver of the communication system are in variable speed motion in shallow water acoustic channels.

Recovery of spatially varying acoustical properties via automated partial differential equation identification.

Observable dynamics, such as waves propagating on a surface, are generally governed by partial differential equations (PDEs), which are determined by the physical properties of the propagation media. The spatial variations of these properties lead to spatially dependent PDEs. It is useful in many fields to recover the variations from the observations of dynamical behaviors on the material. A method is proposed to form a map of the physical properties' spatial variations for a material via data-driven spatially dependent PDE identification and applied to recover acoustical properties (viscosity, attenuation, and phase speeds) for propagating waves. The proposed data-driven PDE identification scheme is based on ℓ1-norm minimization. It does not require any PDE term that is assumed active from the prior knowledge and is the first approach that is capable of identifying spatially dependent PDEs from measurements of phenomena. In addition, the method is efficient as a result of its non-iterative nature and can be robust against noise if used with an integration transformation technique. It is demonstrated in multiple experimental settings, including real laser measurements of a vibrating aluminum plate. Codes and data are available online at https://tinyurl.com/4wza8vxs.

Generative models for sound field reconstruction.

This work examines the use of generative adversarial networks for reconstructing sound fields from experimental data. It is investigated whether generative models, which learn the underlying statistics of a given signal or process, can improve the spatio-temporal reconstruction of a sound field by extending its bandwidth. The problem is significant as acoustic array processing is naturally band limited by the spatial sampling of the sound field (due to the difficulty to satisfy the Nyquist criterion in space domain at high frequencies). In this study, the reconstruction of spatial room impulse responses in a conventional room is tested based on three different generative adversarial models. The results indicate that the models can improve the reconstruction, mostly by recovering some of the sound field energy that would otherwise be lost at high frequencies. There is an encouraging outlook in the use of statistical learning models to overcome the bandwidth limitations of acoustic sensor arrays. The approach can be of interest in other areas, such as computational acoustics, to alleviate the classical computational burden at high frequencies.

Inversion in an uncertain ocean using Gaussian processes.

Gaussian processes (GPs) can capture correlation of the acoustic field at different depths in the ocean. This feature is exploited in this work for pre-processing acoustic data before these are employed for source localization and environmental inversion using matched field inversion (MFI) in an underwater waveguide. Via the application of GPs, the data are denoised and interpolated, generating densely populated acoustic fields at virtual arrays, which are then used as data in MFI. Replicas are also computed at the virtual receivers at which field predictions are made. The correlations among field measurements at distinct spatial points are manifested through the selection of kernel functions. These rely on hyperparameters, that are estimated through a maximum likelihood process for optimal denoising and interpolation. The approach, employing Gaussian and Matérn kernels, is tested on synthetic and real data with both an exhaustive search and genetic algorithms and is found to be superior to conventional beamformer MFI. It is also shown that the Matérn kernel, providing more degrees of freedom because of an increased number of hyperparameters, is preferable over the frequently used Gaussian kernel.

Graph-based sequential beamforming.

This paper presents a Bayesian estimation method for sequential direction finding. The proposed method estimates the number of directions of arrivals (DOAs) and their DOAs performing operations on the factor graph. The graph represents a statistical model for sequential beamforming. At each time step, belief propagation predicts the number of DOAs and their DOAs using posterior probability density functions (pdfs) from the previous time and a different Bernoulli-von Mises state transition model. Variational Bayesian inference then updates the number of DOAs and their DOAs. The method promotes sparse solutions through a Bernoulli-Gaussian amplitude model, is gridless, and provides marginal posterior pdfs from which DOA estimates and their uncertainties can be extracted. Compared to nonsequential approaches, the method can reduce DOA estimation errors in scenarios involving multiple time steps and time-varying DOAs. Simulation results demonstrate performance improvements compared to state-of-the-art methods. The proposed method is evaluated using ocean acoustic experimental data.

Blind source separation by long-term monitoring: A variational autoencoder to validate the clustering analysis.

Noise exposure influences the comfort and well-being of people in several contexts, such as work or learning environments. For instance, in offices, different kind of noises can increase or drop the employees' productivity. Thus, the ability of separating sound sources in real contexts plays a key role in assessing sound environments. Long-term monitoring provide large amounts of data that can be analyzed through machine and deep learning algorithms. Based on previous works, an entire working day was recorded through a sound level meter. Both sound pressure levels and the digital audio recording were collected. Then, a dual clustering analysis was carried out to separate the two main sound sources experienced by workers: traffic and speech noises. The first method exploited the occurrences of sound pressure levels via Gaussian mixture model and K-means clustering. The second analysis performed a semi-supervised deep clustering analyzing the latent space of a variational autoencoder. Results show that both approaches were able to separate the sound sources. Spectral matching and the latent space of the variational autoencoder validated the assumptions underlying the proposed clustering methods.

Sound source localization using multiple ad hoc distributed microphone arrays.

Sound source localization is crucial for communication and sound scene analysis. This study uses direction-of-arrival estimates of multiple ad hoc distributed microphone arrays to localize sound sources in a room. An affine mapping between the independent array estimates and the source coordinates is derived from a set of calibration points. Experiments show that the affine model is sufficient to locate a source and can be calibrated to physical dimensions. A projection of the local array estimates increases localization accuracy, particularly further away from the calibrated region. Localization tests in three dimensions compare the affine approach to a nonlinear neural network.

Reducing echo state network size with controllability matrices.

Echo state networks are a fast training variant of recurrent neural networks excelling at approximating nonlinear dynamical systems and time series prediction. These machine learning models act as nonlinear fading memory filters. While these models benefit from quick training and low complexity, computation demands from a large reservoir matrix are a bottleneck. Using control theory, a reduced size replacement reservoir matrix is found. Starting from a large, task-effective reservoir matrix, we form a controllability matrix whose rank indicates the active sub-manifold and candidate replacement reservoir size. Resulting time speed-ups and reduced memory usage come with minimal error increase to chaotic climate reconstruction or short term prediction. Experiments are performed on simple time series signals and the Lorenz-1963 and Mackey-Glass complex chaotic signals. Observing low error models shows variation of active rank and memory along a sequence of predictions.

Gridless sparse covariance-based beamforming via alternating projections including co-prime arrays.

This paper presents gridless sparse processing for direction-of-arrival (DOA) estimation. The method solves a gridless version of sparse covariance-based estimation using alternating projections. Gridless sparse DOA estimation is represented by the reconstruction of Toeplitz-structured low-rank matrices, which our method recovers by alternatively projecting a solution matrix. Compared to the existing gridless sparse methods, our method improves speed and accuracy and considers non-uniformly configured linear arrays. High-resolution and reliable DOA estimation are achieved even with single-snapshot data, coherent sources, and non-uniform arrays. Simulation results demonstrate performance improvements compared to the existing DOA estimators, including gridless sparse methods. The method is illustrated using experimental data from a real ocean experiment.

Sound source localization based on multi-task learning and image translation network.

Supervised learning-based sound source localization (SSL) methods have been shown to achieve a promising localization accuracy in the past. In this paper, MTIT, SSL for indoors using Multi-Task learning and Image Translation network, an image translation-based deep neural networks (DNNs) framework for SSL is presented to predict the locations of sound sources with random positions in a continuous space. We extract and represent the spatial features of the sound signals as beam response at each direction which can indicate the chance of the source in each point of the room. We utilize the multi-task learning (MTL) based training framework. There are one encoder and two decoders in our DNN. The encoder aims to obtain a compressed representation of the input beamspectrum surfaces while the two decoders focus on two tasks in parallel. One decoder focuses on resolving the multipath caused by reverberation and the other decoder predicts the source location. Since these two decoders share the same encoder, by training these two decoders in parallel, the shared representations are refined. We comprehensively evaluate the localization performance of our method in the simulated data, measured impulse response and real recordings datasets and compare it with multiple signal classification, steered response power with phase transform, and a competing convolutional neural network approach. It turns out that MTIT can outperform all of the baseline methods in a dynamic environment and also can achieve a good generalization performance.

Automated partial differential equation identification.

Inspired by recent developments in data-driven methods for partial differential equation (PDE) estimation, we use sparse modeling techniques to automatically estimate PDEs from data. A dictionary consisting of hypothetical PDE terms is constructed using numerical differentiation. Given data, PDE terms are selected assuming a parsimonious representation, which is enforced using a sparsity constraint. Unlike previous PDE identification schemes, we make no assumptions about which PDE terms are responsible for a given field. The approach is demonstrated on synthetic and real video data, with physical phenomena governed by wave, Burgers, and Helmholtz equations. Codes are available at https://github.com/NoiseLab-RLiu/Automate-PDE-identification.

Introduction to the special issue on machine learning in acoustics.

The use of machine learning (ML) in acoustics has received much attention in the last decade. ML is unique in that it can be applied to all areas of acoustics. ML has transformative potentials as it can extract statistically based new information about events observed in acoustic data. Acoustic data provide scientific and engineering insight ranging from biology and communications to ocean and Earth science. This special issue included 61 papers, illustrating the very diverse applications of ML in acoustics.

Reinforcement learning applied to metamaterial design.

This paper presents a semi-analytical method of suppressing acoustic scattering using reinforcement learning (RL) algorithms. We give a RL agent control over design parameters of a planar configuration of cylindrical scatterers in water. These design parameters control the position and radius of the scatterers. As these cylinders encounter an incident acoustic wave, the scattering pattern is described by a function called total scattering cross section (TSCS). Through evaluating the gradients of TSCS and other information about the state of the configuration, the RL agent perturbatively adjusts design parameters, considering multiple scattering between the scatterers. As each adjustment is made, the RL agent receives a reward negatively proportional to the root mean square of the TSCS across a range of wavenumbers. Through maximizing its reward per episode, the agent discovers designs with low scattering. Specifically, the double deep Q-learning network and the deep deterministic policy gradient algorithms are employed in our models. Designs discovered by the RL algorithms performed well when compared to a state-of-the-art optimization algorithm using fmincon.

Mode separation with one hydrophone in shallow water: A sparse Bayesian learning approach based on phase speed.

An approach of broadband mode separation in shallow water is proposed using phase speed extracted from one hydrophone and solved with sparse Bayesian learning (SBL). The approximate modal dispersion relation, connecting the horizontal wavenumbers (phase velocities) for multiple frequencies, is used to build the dictionary matrix for SBL. Given a multi-frequency pressure vector on one hydrophone, SBL estimates a set of sparse coefficients for a large number of atoms in the dictionary. With the estimated coefficients and corresponding atoms, the separated normal modes are retrieved. The presented method can be used for impulsive or known-form signals in a shallow-water environment while no bottom information is required. The simulation results demonstrate that the proposed approach is adapted to the environment where both the reflected and refracted modes coexist, whereas the performance of the time warping transformation degrades significantly in this scenario.

Deep embedded clustering of coral reef bioacoustics.

Deep clustering was applied to unlabeled, automatically detected signals in a coral reef soundscape to distinguish fish pulse calls from segments of whale song. Deep embedded clustering (DEC) learned latent features and formed classification clusters using fixed-length power spectrograms of the signals. Handpicked spectral and temporal features were also extracted and clustered with Gaussian mixture models (GMM) and conventional clustering. DEC, GMM, and conventional clustering were tested on simulated datasets of fish pulse calls (fish) and whale song units (whale) with randomized bandwidth, duration, and SNR. Both GMM and DEC achieved high accuracy and identified clusters with fish, whale, and overlapping fish and whale signals. Conventional clustering methods had low accuracy in scenarios with unequal-sized clusters or overlapping signals. Fish and whale signals recorded near Hawaii in February-March 2020 were clustered with DEC, GMM, and conventional clustering. DEC features demonstrated the highest accuracy of 77.5% on a small, manually labeled dataset for classifying signals into fish and whale clusters.

Sequential sparse Bayesian learning for time-varying direction of arrival.

This paper presents methods for the estimation of the time-varying directions of arrival (DOAs) of signals emitted by moving sources. Following the sparse Bayesian learning (SBL) framework, prior information of unknown source amplitudes is modeled as a multi-variate Gaussian distribution with zero-mean and time-varying variance parameters. For sequential estimation of the unknown variance, we present two sequential SBL-based methods that propagate statistical information across time to improve DOA estimation performance. The first method heuristically calculates the parameters of an inverse-gamma hyperprior based on the source signal estimate from the previous time step. In addition, a second sequential SBL method is proposed, which performs a prediction step to calculate the prior distribution of the current variance parameter from the variance parameter estimated at the previous time step. The SBL-based sequential processing provides high-resolution DOA tracking capabilities. Performance improvements are demonstrated by using simulated data as well as real data from the SWellEx-96 experiment.