Search

Efforts to mentor undergraduates in research require time, energy, and resources but can yield significant benefits. A review of education research on undergraduate research is presented followed by a case study of how technical and educational principles from the literature have been implemented and practiced in a new underwater acoustics lab. This case study involves the students who have joined this lab over its first three years. The results highlight the importance of several key factors that mentors can implement to improve their students' research experience. In order to meet the steep learning curve that new students face, faculty mentors should establish a culture of research (i.e., well established norms, expectations, and practices associated with being a part of a research group) through clear expectations, open communication, and student-led peer mentorship supported by carefully selected or designed resources (i.e., scaffolding). This paper seeks to share these education research resources with the acoustics community while providing a case study for review as a specific application of these key principles into an acoustics laboratory setting.

Sounds to Astound is an acoustics demonstration show, produced for the community twice yearly by the Brigham Young University Student Chapter of the Acoustical Society of America. The free, interactive demonstration show explores the science of sound for a target audience of fifth- to eighth-grade students. Introductory acoustics concepts, such as longitudinal wave motion, wave properties, propagation effects, and standing waves, are taught with live demonstrations, animations, and videos. The goal of this paper is to inspire and encourage readers in their outreach efforts by describing the purposes of Sounds to Astound and technical details of several entertaining and educational demonstrations. Lessons learned from a decade of these student-produced shows serve as an aid for future efforts and highlight the benefits of outreach efforts, particularly for the students involved.

Merchant ship-radiated noise, recorded on a single receiver in the 360-1100 Hz frequency band over 20 min, is employed for seabed classification using an ensemble of deep learning (DL) algorithms. Five different convolutional neural network architectures and one residual neural network are trained on synthetic data generated using 34 seabed types, which span from soft-muddy to hard-sandy environments. The accuracy of all of the networks using fivefold cross-validation was above 97%. Furthermore, the impact of the sound speed and water depth mismatch on the predictions is evaluated using five simulated test cases, where the deeper and more complex architectures proved to be more robust against this variability. In addition, to assess the generalizability performance of the ensemble DL, the networks were tested on data measured on three vertical line arrays in the Seabed Characterization Experiment in 2017, where 94% of the predictions indicated that mud over sand environments inferred in previous geoacoustic inversions for the same area were the most likely sediments. This work presents evidence that the ensemble of DL algorithms has learned how the signature of the sediments is encoded in the ship-radiated noise, providing a unified classification result when tested on data collected at-sea.

An examination of the received spectrogram levels of about twenty merchant ship recordings on two vertical line arrays deployed on the New England continental shelf during the Seabed Characterization Experiment 2017 has identified an acoustic feature that can be attributed to the group velocities of modes 1 and 2 being equal at a frequency f = F. The observation of such a feature is a result of β n m ( 2 π F ) = ∞, where βnm is the waveguide invariant for modes n and m. For the New England Mudpatch, the average value of F is about 24.5 Hz. An effective seabed model is inferred from a feature inversion method that has a deep sediment layer which lies between 190 m and 290 m beneath the seafloor with sound speeds on the order of 1810 m/s. This effective sediment model appears to be consistent with a previous seismic survey on the New England shelf that identified a deep low speed layer about 250 m beneath the water sediment interface.

Broadband spectrograms from surface ships are employed in convolutional neural networks (CNNs) to predict the seabed type, ship speed, and closest point of approach (CPA) range. Three CNN architectures of differing size and depth are trained on different representations of the spectrograms. Multitask learning is employed; the seabed type prediction comes from classification, and the ship speed and CPA range are estimated via regression. Due to the lack of labeled field data, the CNNs are trained on synthetic data generated using measured sound speed profiles, four seabed types, and a random distribution of source parameters. Additional synthetic datasets are used to evaluate the ability of the trained CNNs to interpolate and extrapolate source parameters. The trained models are then applied to a measured data sample from the 2017 Seabed Characterization Experiment (SBCEX 2017). While the largest network provides slightly more accurate predictions on tests with synthetic data, the smallest network generalized better to the measured data sample. With regard to the input data type, complex pressure spectral values gave the most accurate and consistent results for the ship speed and CPA predictions with the smallest network, whereas using absolute values of the pressure provided more accurate results compared to the expected seabed types.

While source localization and seabed classification are often approached separately, the convolutional neural networks (CNNs) in this paper simultaneously predict seabed type, source depth and speed, and the closest point of approach. Different CNN architectures are applied to mid-frequency tonal levels from a moving source recorded on a 16-channel vertical line array (VLA). After training each CNN on synthetic data, a statistical representation of predictions on test cases is presented. The performance of a single regression-based CNN is compared to a multitask CNN in which regression is used for the source parameters and classification for the seabed type. The impact of water sound speed profile and seabed variations on the predictions is evaluated using simulated test cases. Environmental mismatch between the training and testing data has a negative impact on source depth estimates, while the remaining labels are estimated tolerably well but with a bias towards shorter ranges. Similar results are found for data measured on two VLAs during Seabed Characterization Experiment 2017. This work shows the superiority of multitask learning and the potential for using a CNN to localize an acoustic source and detect the surficial seabed properties from mid-frequency sounds.

Noise from a tactical aircraft can impact operations due to concerns regarding military personnel noise exposure and community annoyance and disturbance. The efficacy of mission planning can increase when the distinct, complex acoustic source mechanisms creating the noise are better understood. For each type of noise, equivalent acoustic source distributions are obtained from a tied-down F-35B operating at various engine conditions using the hybrid method for acoustic source imaging of Padois, Gauthier, and Berry [J. Sound Vib. 333, 6858-6868 (2014)]. The source distributions for the distinct noise types are obtained using different sections of a 71 element, ground-based linear array. Using a subarray close to the nozzle exit plane, source distributions are obtained for fine-scale turbulent mixing noise and broadband shock-associated noise, although grating lobes complicate interpretations at higher frequencies. Results for a subarray spanning the maximum sound region show that the multiple frequency peaks in tactical aircraft noise appear to originate from overlapping source regions. The observation of overlapping spatial extent of competing noise sources is supported by the coherence properties of the source distributions for the different subarrays.

In ocean acoustics, many types of optimizations have been employed to locate acoustic sources and estimate the properties of the seabed. How these tasks can take advantage of recent advances in deep learning remains as open questions, especially due to the lack of labeled field data. In this work, a Convolutional Neural Network (CNN) is used to find seabed type and source range simultaneously from 1 s pressure time series from impulsive sounds. Simulated data are used to train the CNN before application to signals from a single hydrophone signal during the 2017 Seabed Characterization Experiment. The training data includes four seabeds representing deep mud, mud over sand, sandy silt, and sand, and a wide range of source parameters. When applied to measured data, the trained CNN predicts expected seabed types and obtains ranges within 0.5 km when the source-receiver range is greater than 5 km, showing the potential for such algorithms to address these problems.

The traditional method for intensity-based sound power estimates often used in engineering applications is limited in bandwidth by microphone phase mismatch at low frequencies and by microphone spacing at high frequencies. To overcome these limitations, the Phase and Amplitude Gradient Estimator (PAGE) method [Gee, Neilsen, Sommerfeldt, Akamine, and Okamoto, J. Acoust. Soc. Am. 141(4), EL357-EL362 (2017)] is applied to sound power for a reference sound source, a blender, and a vacuum cleaner. Sound power measurements taken according to ISO 3741:2010 (2010) are compared against traditional- and PAGE-processed intensity-based sound power estimates measured according to ANSI S12.12-1992 (R2017). While the traditional method underestimates the sound power at the spatial Nyquist frequency by 7-10 dB, the PAGE-based sound power is accurate up to the spatial Nyquist frequency, and above when phase unwrapping is successful.

Environmental inversions in ocean acoustics require the simultaneous estimation of many unknown parameters. Broad studies of multi-parameter models from diverse fields have shown that such inference problems are universally sloppy. Sloppiness is a phenomenon in which the predictions of a model are insensitive to all but a few key combinations of parameters. Sloppy model analysis is based on information geometry, an application of differential geometry to statistics. In this talk, we give an overview of information geometry for underwater acoustics models for environmental inversion, focusing on transmission loss (TL) in range-independent normal mode models. We demonstrate that these models are sloppy and how the geometry directly relates the information content of acoustic data to the relevance of environmental parameters. In particular, the model manifold quantifies what environmental information is encoded in ocean sounds and how unidentifiable parameters can be removed from the model to give a simplified, identifiable ocean acoustics model of comparable accuracy. We summarize physical insights revealed by this information geometry analysis for ocean inversion. [Work supported by Office of Naval Research]

During Seabed Characterization Experiment 2017 in the New England Mudpatch under near isospeed conditions of the water column, observations were made of well-defined intensity striations in time-frequency spectrograms in the 1.5–4.0 kHz band. These observations provided the basis for the hypothesis that the compressional sound speed profile for a homogenous fine-grained sediment possessed a gradient in the upper portions of the seabed. This hypothesis was tested with a statistical inference method that extracts probability density functions for parameter values representing an empirical-based nonlinear profile whose origins go back to Hamilton. Inferred was a surface sound speed that is less than the sound speed of the water at the seafloor and a surface sound speed gradient of about 6 1/s. In 2022 the experiment was repeated during non-isospeed conditions of the water column. The results for 2017 and 2022 are compared with the intent of understanding the response of the geoacoustic properties of the upper portions of a fine-grained sediment to changes in the bottom water temperature and salinity of the water column. [Work supported by Office of Naval Research.]

The variation of temperature in the ocean causes changes in sound propagation. To simulate the naturally occurring variability in an ocean environment, this project quantified the sound speed variation achievable in a laboratory water tank. The rectangular tank (1.2 m × 3.6 m with 0.5 m depth) has anechoic paneling that minimizes lateral reflections. In this experiment, temperature was measured from four different temperature sensors. The temperature was used to calculate the sound speed in the tank as a function of time while the water was cooled and heated. Sound speed values were compared from four different empirical equations. We found that while the temperature varies in time, the temperature is relatively uniform at different locations in the tank. In experiments involving three hours of heating and adding 100 pounds of ice, there was a 10-degree Celsius difference in temperature, corresponding to a 20 m/s difference in water sound speed. By adding an additional degree of variability to the tank measurements, a portion of the variability seen in the ocean can be replicated. This sound speed variability can then be used to test the robustness of machine learning algorithms.

Remote sensing using passive sonar in the ocean is a challenging problem due to variations in the geoacoustic structure of the seabed and unknown source location and strength. One way to improve remote sensing is to perform an optimization for geoacoustic and source parameters. We use a Bayesian maximum entropy (BME) approach with a viscous-grain shearing parameterization for two sediment layers. The statistical optimization provides probability distributions for porosity and thickness of the sediment layers as well as ship speed, closest point of approach, and the source strength for the Wales-Heitmeyer empirical source level spectrum. We use this approach on spectrograms of transiting ships collected on a vertical line array during the 2017 seabed characterization experiment. We compare the resulting parameter distributions from distinct ships as well as previous estimates of geoacoustic values and source properties. This research shows that the BME approach obtains estimates for porosity and source strength that have narrow posterior probability distributions. [Work supported by the Office of Naval Research.]

The connection between ocean environmental parameters and transmission loss (TL) can be modeled using sound propagation models. We can analyze and improve these models using information theoretic tools such as the Fisher information matrix (FIM) and methods for model reduction, where we identify parameters that can be removed from a model without sacrificing accuracy. Critical to these methods for model reduction is the ability to evaluate derivatives of the model predictions of TL with respect to model parameters. Calculating derivatives of TL models using finite difference methods with sufficient accuracy is challenging due to the complexity of sound fields in an ocean environment; instead, we train a surrogate machine learning model to which we can apply automatic differentiation (AD). We propose a general method for model reduction in which a ML model is used as a surrogate model for physical sound propagation ocean models, which is then used to find potential simplifications of the original model. We demonstrate this method using the Pekeris waveguide model of the ocean, which calculates underwater acoustic TL in an ocean environment due to seafloor characteristics. [Work supported by Office of Naval Research]

Dimensional reduction methods based on feature extraction are critical for statistical inference application in variable ocean environments. Presented are analyses for selected acoustic measurements made during the Seabed Characterization Experiment in 2021-2022. Broadband sources were deployed on the New England continental slope between the 200 and 400 m depth contours. The method of analysis includes first analyzing short range (higher grazing angles) acoustic measurements provided by both a broadband towed source and explosive sources. Features utilized for short-ranged data include those associated with seabed interferometry. Then, longer range data (lower grazing angles) are utilized to provide improved estimates of such quantities as the attenuations and sound speed gradients. Features utilized for long-ranged data include those associated with the dispersion of signals generated by explosive sources. As a general guide the mid-frequency band is utilized to infer characteristics about the upper portions of the sediment whereas the lower frequencies are utilized to infer the deeper portions of the seabed. [Work supported by Office of Naval Research.]

Deep learning has great potential to be useful for ocean acoustics problems. However, uncertainty in an ocean environment must be taken into account, because it affects the ability of trained machine learning models to generalize. Because ocean measurements are costly, we are using a laboratory water tank to help develop our deep learning neural networks. In addition, using a water tank in a lab increases some level of control over the amount of uncertainty in data used to train neural networks. A residual convolutional neural network architecture is trained to identify source-receiver range in an underwater tank environment at room temperature. The model is then tested on data taken with variable water temperature. For comparison, two different source signals are used (sine waves and linear chirps) to analyze the usefulness of broadband sounds. These tests explore how environmental uncertainty impacts source localization and will lead to insights on how to develop robust machine learning for ocean acoustics problems. [Work supported by the Office of Naval Research.]

Because ocean measurements are costly, a laboratory water tank provides opportunities to take scaled measurements and develop approaches to machine learning that are transferable to ocean acoustic problems. These applications are better achieved if reflections off the side walls are significantly reduced. To reduce these reflections, the side walls of the tank are lined with Aptile SF5048 echo reducing tiles made by Precision Acoustics. This work studies the extent to which two ocean sound propagation models (namely ORCA and BELLHOP) can predict relative transmission loss as a function of source-receiver range and frequency. The modeling parameters of sound speed and density that determine the reflections at the bottom of the tank were refined through a Bayesian optimization algorithm. We found that the optimized parameters improve the model-data agreement and that the models fit the data at certain frequencies corresponding mainly to peaks in the echo reduction for the Aptile panels. This agreement implies that at certain frequencies, sound propagation in our tank acts as it would in a large body of water. Using our tank as a testbed for developing machine learning algorithms is a reasonable approach.

Since the ocean is not a convenient place to conduct research, data are often collected from offshore buoys. However, the data collected from the buoys usually do not have labels (e.g., source of the sound) attached to them. Manually labeling the data is time-consuming, but labeled data are necessary to use powerful supervised machine learning methods. We propose that unsupervised learning can be used to assign preliminary labels to data samples. To show the utility of this approach, we took an echosounder to Utah Lake and collected sonar images, some of which showed evidence of fish. We used Python to crop the images, similar to how regions of interest are identified in image processing. These cropped images are used in an unsupervised k-means clustering algorithm. The images are clustered based on visual features. An analysis of the resulting clusters and the feasibility of using unsupervised learning to provide preliminary labels to data is discussed.

Deep learning can assist in characterizing seabeds using sources of opportunity such as shipping noise. While previous work focused on seabed classification, this study uses a residual convolutional neural network to find individual seabed properties. The training data were labeled with sound speed, density, attenuation, and thickness of the layer values of the top sediment layer. A comparison was made between predictive capabilities of ResNet-18 networks when trained to learn a single parameter and those trained to simultaneously learn multiple parameters. For stiff parameters—those with high information content in the data—learning an individual parameter performed better. These single parameter predictions are fundamentally different from a geoacoustic inversion for one parameter. In geoacoustic inversion, all other parameters are held at a fixed value. In deep learning, variability in all other parameters is contained in the training data, but the network focuses on features in the data related to a single property. The trained networks are applied to ship noise measured during the 2017 Seabed Characterization Experiment. [Work supported by the Office of Naval Research and the National Science Foundation’s REU program.]

A risk factor shared by all musicians is a potential for noise-induced hearing loss. A study was conducted to explore what factors determine noise dosage during musical events experienced by college student musicians. First, noise exposure during specific activities was explored, including during ensemble rehearsal and personal practice. Next, full-day noise exposure was investigated by measuring levels experienced by student musicians during a typical daily routine. Factors such as instrument played, type of activity, location within ensemble, and room environment were related to noise dosage. Disparities in results using different standards to calculate noise dosage were also explored. Risk of noise overexposure was found to be greater in some instrument classes, such as wind instruments, than in others, such as string instruments, and can yield vastly different results depending on the metric used for noise dosage calculation.

Calculation of acoustic intensity using the phase and amplitude gradient estimator (PAGE) method has been shown to increase the effective upper frequency limit beyond the traditional p-p method when the source of interest is broadband in frequency [Torrie, Whiting, Gee, Neilsen, and Sommerfeldt, Proc. Mtgs. Acoust. 23, 030005 (2015)]. PAGE processing calculates intensity for narrowband sources without bias error up to the spatial Nyquist frequency [Succo, Sommerfeldt, Gee, and Neilsen, Proc. Mtgs. Acoust. 30, 030015 (2018)]. The present work demonstrates that for narrowband sources with frequency content above the spatial Nyquist frequency, additive low-level broadband noise can improve intensity calculations. To be effective, the angular separation between the source and additive noise source should be less than 30 °, while using phase unwrapping with a smaller angular separation will increase the usable bandwidth. The upper frequency limit for the bandwidth extension depends on angular separation, sound speed, and probe microphone spacing. Assuming the signal-to-additive-noise ratio ( SNR a) is larger than 10 dB, the maximum level and angular bias errors incurred by the additive broadband noise beneath the frequency limit—or up until probe scattering effects must be taken into account—are less than 0.5 dB and 2.5 °, respectively. Smaller angular separation yields smaller bias errors.

Because the seabed impacts sound propagation in the ocean, machine learning is being used for both seabed classification and to obtain estimates of individual seabed properties. This paper proposes a method to simultaneously estimate these properties and an associated uncertainty label. A residual neural network is trained and validated using synthetic ship noise spectrograms generated with a range-independent normal mode sound propagation model and a ship noise source spectrum. The data set includes 140 seabeds: In each, the top sediment layer has a random thickness and properties randomly chosen from five sets of bounds, which roughly correspond to clay, mud, sand, silt, and gravel. For each of the 140 sediments, a random selection of 405 combinations of ship speed, closest-point-of-approach range, and source depth are used resulting in 22k data samples. Each data sample is labeled with the true values of the sediment parameters as well as a label for the uncertainty level of each. The uncertainty levels are obtained from the Fisher information and qualify of the information content in the ship spectrogram about each parameter. Examples of how the residual neural network learns to perform regression for the parameter value and the uncertainty level will be presented.

Graduate studies in acoustics at Brigham Young University prepare students for industry, research, and academia by complementing in-depth coursework with publishable research. Coursework provides a solid foundation in core acoustical principles and practices and measurement skills, including a strong foundation in experimental techniques and technical writing. Labs across the curriculum cover calibration, directivity, scattering, absorption, laser Doppler vibrometry, experimental methods for dynamic structures, lumped-element mechanical systems, equivalent circuit modeling, arrays, filters, room acoustics, active noise control, and near-field acoustical holography. Recent thesis and dissertation topics include active noise control, directivity, room acoustics, energy-based acoustics, time reversal, nondestructive evaluation, vibration and acoustics of aerospace vehicles, biomedical applications, flow-based acoustics, voice production, aeroacoustics, sound propagation modeling, nonlinear propagation, high-amplitude noise analyses, machine and deep learning applied to ambient noise level prediction, crowd noise interpretation, and underwater acoustic source localization, and ocean environment classification. Graduate students are expected to present research at professional meetings and publish in peer-reviewed acoustics journals. Additionally, graduate students often serve as peer mentors to undergraduate students on related projects and may participate in field experiments to gain additional experience.@BYUAcoustics

In this study the characteristics of the modal dispersion of acoustic waves in littoral ocean environments are utilized to infer both source properties (range, depth, speed, and source level) and geoacoustic properties of the seabed. An analysis is made of broadband acoustic data (10–4500 Hz) taken during 2021–2022 on the New England continental slope over water depths from 200 to 500 m. The acoustic data are generated by both moving broadband sources and explosive sources over range scales of 1–30 km. Such range scales can reveal the nature of the modal dispersion. Conductivity, Temperature, and Depth (CTD) and Expendable Bathythermographs (XBTs) measurements provide a sparse sampling of the temporal and spatial variability of the ocean during the acoustic measurements. Along with bathymetry measurements the sound speed profiles derived from the CTDs and XBTs provide prior information for an inverse problem from which insight can be obtained on the effects of the seabed on the dispersion of broadband propagation.

The waveguide invariant βmn (f), m,n = 1,2,…,N where N is the number of observed modes is inferred from sound emitted by merchant ships and from a towed mid-frequency source emitting CHIRP signals. The acoustic data were recorded on VLAs deployed in the New England Mudpatch during the Seabed Characterization Experiment 2017 in about 75 m of water. The received spectrograms show clear striation patterns out to 4 kHz. The seabed consists of about 12 m of mud over multiple sand layers extending down to about 300 m. Starting at β12(F) = ∞ at F about 25 Hz, βmn(f) is extended to 4 kHz. Below 500 Hz βmn exhibits a non-Pekeris behavior which results from the depth dependence of the multilayer seabed. Above 500 Hz the behavior of βmn begins to exhibit more of a Pekeris behavior, although there exist features that result from the finite thickness of the mud layer. A conclusion is that βmn (f) is sensitive to the geoacoustic structure over a large bandwidth, and thus has the potential to be utilized for remote sensing of the seabed with wideband sources.

Bias errors for two-dimensional active acoustic intensity using multi-microphone probes have been previously calculated for both the traditional cross-spectral and the Phase and Amplitude Gradient Estimator (PAGE) methods [Whiting, Lawrence, Gee, Neilsen, and Sommerfeldt, J. Acoust. Soc. Am. 142, 2208–2218 (2017)]. Here, these calculations are expanded to include errors due to contaminating noise, as well as probe orientation. The noise can either be uncorrelated at each microphone location or self-correlated; the self-correlated noise is modeled as a plane-wave with a varying angle of incidence. The intensity errors in both magnitude and direction are dependent on the signal-to-noise ratio (SNR), frequency, source properties, incidence angles, probe configuration, and processing method. The PAGE method is generally found to give more accurate results, especially in direction; however, uncorrelated noise with a low SNR (below 10–15 dB) and low frequency (wavelengths more than 1/4 the microphone spacing) can yield larger errors in magnitude than the traditional method—though a correction for this is possible. Additionally, contaminating noise does not necessarily impact the possibility of using the PAGE method for broadband signals beyond a probe's spatial Nyquist frequency.Bias errors for two-dimensional active acoustic intensity using multi-microphone probes have been previously calculated for both the traditional cross-spectral and the Phase and Amplitude Gradient Estimator (PAGE) methods [Whiting, Lawrence, Gee, Neilsen, and Sommerfeldt, J. Acoust. Soc. Am. 142, 2208–2218 (2017)]. Here, these calculations are expanded to include errors due to contaminating noise, as well as probe orientation. The noise can either be uncorrelated at each microphone location or self-correlated; the self-correlated noise is modeled as a plane-wave with a varying angle of incidence. The intensity errors in both magnitude and direction are dependent on the signal-to-noise ratio (SNR), frequency, source properties, incidence angles, probe configuration, and processing method. The PAGE method is generally found to give more accurate results, especially in direction; however, uncorrelated noise with a low SNR (below 10–15 dB) and low frequency (wavelengths more than 1/4 the microphone spacing) c...

The impact of individual seabed properties on sound propagation in the ocean depends on many factors including source-receiver range and frequency band of interest. In this talk, estimates of the sound speed ratio across the water-sediment interface are obtained using a maximum entropy approach and ResNet18, a supervised machine learning model. The input data are spectrograms of surface ship noise from shipping lanes. Synthetic spectrograms are modeled using a ship noise source spectrum and a range-independent normal mode model, ORCA, with a wide range of environments and ship parameters. Experimental data from the New England Mud Patch are used with both inverse methods. The maximum entropy approach uses data-model mismatch to obtain a posterior probability distribution for the parameters of interest. The ResNet18 is trained on the synthetic spectrograms, augmented with additive noise, and then applied to the experimental data. A comparison of the results from these two methods for a variety of ships using different frequency bands will be presented, along with a discussion of the advantages and limitations of each method.

Supersonic jet noise reduction efforts benefit from targeted source feature extraction and high-resolution acoustic imaging. Another useful tool for feature extraction is partial field decomposition of sources into independent contributors. Since such decomposition processes are nonunique, care must be taken in the physical interpretation of decomposed partially coherent aeroacoustic fields. The optimized-location virtual reference method (OLVR) is a partial field decomposition designed to extract physically meaningful source and field information through the strategic placement of virtual references within a reconstructed field. The OLVR method is applied here to obtain spatially distinct and ordered partial sources at multiple frequencies of a full-scale, high-performance supersonic jet engine operating at 100% engine power. Partial sources are shown to mimic behaviors of the total source distributions including monotonic growth and decay. Because of finite spatial coherence, multiple partial sources are used to reproduce far-field radiation away from the main lobe, and the number of required sources increases with increasing frequency. An analytical multiwavepacket model is fitted to the partial sources to demonstrate how OLVR partial fields can be leveraged to produce reduced-order models.

Results of the first formal perceptual study of jet crackle are presented. Prior studies examined noise waveform properties believed to be linked to the jet crackle percept from a physics perspective or using signal processing and informal subjective evaluation. This investigation involves 31 listeners that rated 15 jet noise waveforms with a category subdivision scaling test. Results reveal a strong log-linear correlation between the pressure waveform time derivative's skewness and crackle rating. A regression analysis establishes practical derivative skewness bounds for a five-point categorical crackle scale and results in the suggested definition of the crepit as the unit of crackliness.

In acoustic intensity estimation, adding a microphone at the probe center removes errors associated with pressure averaging. Analytical bias errors are presented for a one-dimensional, three-microphone probe for active intensity, reactive intensity, and specific acoustic impedance in a monopole field. Traditional estimation is compared with the Phase and Amplitude Gradient Estimator (PAGE) method; the PAGE method shows an increased bandwidth for all three quantities. The two- and three-microphone methods are compared experimentally, showing reduced bias errors with three-microphone PAGE for active and reactive intensity, whereas using two microphones is preferred for specific acoustic impedance.

An 18-layer residual neural network (ResNet-18) and a convolutional neural network (CNN) are used for source localization and seabed classification using broadband spectrograms obtained from ship-radiated noise. The propagation of the broadband noise from those ships of opportunity (SOO) contains significant information about the waveguide and seafloor. In this work, a ResNet-18 and a CNN are trained using multi-task learning to estimate the closest point of approach, ship speed, and seabed type. Training data are simulated using a range-independent normal modes model and several sediment profiles obtained by inversions in different regions. The input data samples correspond to 20 min long broadband spectrograms in the 300–1500 Hz band. During the training stage, data augmentation techniques are also applied to evaluate their effects on the predictions. The metrics used to evaluate performance are accuracy for seabed classification and root mean square error for source localization. Results demonstrate the ability of deep learning models to estimate source localization and seabed classification along with the impact of data augmentation techniques using SOO spectrograms. [Work supported by ONR, Contract No. N00014-19-C-2001.]

Research on supervised deep learning methods have shown potential for ocean acoustic applications. Supervised methods require labeled data samples for training. However, ocean acoustics data often do not contain labels. In experimental situations it is possible to estimate source range and speed labels as inferred by GPS data, but the environmental labels may not be easily defined. Applications of machine learning in ocean acoustics will be greatly enhanced by developing ways to utilize semi-supervised learning. In other applications, semi-supervised learning has been accomplished via contrastive learning or self-attention methods. Our work applies semi-supervised learning by developing a contrastive learning framework for acoustic data. First the model does self-training using an augmentation policy on unlabeled data to initialize representations of key features in the data. Then a small set of cleanly labelled samples are given to a model for supervised learning to teach the desired prediction task. Our work in contrastive learning will be applied to transiting surface ship spectrograms to demonstrate the effectiveness for predicting source labels and seabed type. Preliminary work will be presented to illustrate the advantages of semi-supervised learning. [Work supported by ONR contract N00014-19-C-20001.]

The phase and amplitude gradient estimator (PAGE) method for active acoustic intensity uses pairwise microphone transfer functions to obtain the phase gradient, which improves the calculation bandwidth over the traditional weighted quadspectral method. Additionally, for broadband sources, the PAGE theory indicates that the transfer function phase can be unwrapped to further extend the usable frequency range to beyond the spatial Nyquist frequency. Here, two experiments demonstrate intensity bandwidth extension by more than an order of magnitude using phase unwrapping. First, plane-wave tube results show accurate one-dimensional intensity calculations with the microphones separated by five wavelengths, 30 times the traditional limit. Second, two-dimensional measurements of a laboratory-scale jet with a four-microphone probe yield physically reasonable calculations at frequencies 15 times the traditional limit.

Lack of labeled field data in ocean acoustics is a limitation for applying supervised machine learning techniques. Unsupervised techniques have shown potential for classifying signals, but questions remain as to how much additional information unsupervised techniques can learn from ocean acoustic data. This paper evaluates the ability of an unsupervised k-mean clustering algorithm to provide information about the seabed and location of a moving source. The training data are synthetic spectrograms from surface ships based on a wide variety of ship speeds, closest distance to the hydrophone, and seabed types. The resulting clusters are evaluated to discover trends into which types of spectrograms are in each cluster, and the characteristics of the cluster centroids are investigated. These studies highlight the speed-distance ambiguity seen in ship spectrograms as well as the ambiguity between distance and bottom loss of the seabed. The trained algorithm is tested on both synthetic spectrograms and ones from merchant and cargo ships recorded during the Seabed Characterization Experiment in 2017. The future of unsupervised machine learning in ocean acoustics depends on correctly interpreting the results and testing the ability of trained networks to generalize.

Understanding sound propagation in a shallow ocean environment is complicated but important for naval technologies. Some researchers are currently applying a variety of supervised deep learning methods, such as convolution neural networks, for either source localization or seabed classification. One of the costs of supervised learning is the requirement for large amounts of labeled data. In the present study, unlabeled data are used to train a Transformer model. Transformer-based models have demonstrated ability for predicting sequential data; examples are Google’s BERT and OpenAI’s GPT-3. In our technique, we train a transformer-based model in a twopart process. First, self-supervised learning is implemented using synthetic ship spectrograms for various shallow ocean environments. The model is trained as an encoder/decoder to perform sequence-to-sequence prediction. Second, the transformer model is fine-tuned to predict the source location and seabed class from a small set of labeled synthetic samples. Data samples measured during the Seabed Characterization Experiment 2017 are utilized as a testing dataset. The advantages of this approach are the ability to train a model with a larger variety of data, including the use of unlabeled data and data of variable input length.

An ensemble of deep learning algorithms is employed for seabed classification using merchant ship-radiated noise recorded on a single receiver. Five different convolutional neural network (CNN) architectures and one residual neural network (ResNet) are trained on synthetic data generated using 34 seabed types found in the literature from different areas of the world that span from soft-muddy to hard-sandy sediments. Validation results show an accuracy above 97.15% for all the networks using the five-fold cross-validation method. To assess the generalizability performance of the ensemble deep learning, the networks were tested on 69 data samples measured on three VLAs in the Seabed Characterization Experiment in 2017 (SBCEX 2017) from merchant ships passing by close to the receivers. Results show that the most likely sediments were the mud over sand environments inferred in previous geoacoustic inversions for the SBCEX 2017 area with 94% of the predictions from all the trained networks. This work presents evidence that the ensemble of deep learning algorithms has learned how the signature of the sediments is encoded in the ship-radiated noise providing a unified classification result when tested on real-world data. [Work supported by ONR, Contract No. N00014-19-C-2001.]

In ocean acoustics, finding acoustic signals within long recordings is usually time consuming. The need to optimize this process we propose an experiment to discover the optimal acoustic signal classification model using the PyTorch deep learning package. This machine learning algorithm was designed to recognize and classify various sources from an input of spectrograms compiled from raw sound data. In the case of continuous audio files, this model’s purpose is to identify areas of interest for human operators so they don’t need to listen to each hour of the audio files marking down noises as they are heard. Four Convolutional Neural Networks (CNN) with differing numbers of layers took in one minute spectrograms. The most successful of the CNNs acted on time-averaged spectrograms, and resulted in a model which achieved a high degree of accuracy. This machine learning algorithm can help identify underwater sound signal sources, and can more efficiently identify when different signals are present in long audio files. The results of these tests imply that time averaging spectrograms may improve the identification of long term signal sources by a CNN. [ Research supported by the NSF REU program.]

One of the main problems in ocean acoustics is the lack of training data that assist networks to generalize better understand undersea environments. In deep learning, data augmentation is commonly used to add variability to the data by applying small transformations to the samples with the purpose of improving the performance and generalization capabilities of the networks. Here, this definition is extended to broadband spectrograms from merchant ships collected during the Seabed Characterization Experiment in 2017. In this study, residual neural networks were used as the machine learning algorithm for performing source localization and environment classification. Additionally, nine different data augmentation techniques were considered for studying their performance predictions. Each of the transformations was applied to the data set during the training stage, and the results were compared to establish the effects in the performance for both regression and classification tasks. The metrics used were the root mean squared error for regression and accuracy in the case of seabed classification. The results show a favorable potential of residual-based deep learning models to differentiate the seabed types and the source position based on these types of data, in addition to the improvements in performance after applying complex transformations during the training. [Work supported by the Office of Naval Research, Contract No. N00014-19-C-2001.]

An examination of the received spectrogram levels of merchant ship recordings has identified an important acoustic feature at the very-low frequencies (VLF). The data were recorded on vertical line arrays deployed in the New England Mudpatch during the Seabed Characterization Experiment 2017 in about 75 m of water. The VLF feature occurs when the group velocities of modes 1 and 2 are equal at a frequency f = F. The average value of F from about twenty ship signatures observed in the Mudpatch is about 24.5 Hz. Generally, the observation of this VLF feature is a result of βmn( F) = infinity where βmn is the waveguide invariant for modes n and m. The sign of the waveguide invariant is determined by the sign of the difference between the phase speeds at F. Also, we predict what F would be in two areas south of the Mudpatch, including a second smaller Mudpatch in about 125 m of water and an area near the slope in about 250 m of water where the sediment is reported to be sand over multiple layers of softer sediments. Finally, F is predicted on the Atlantis II seamount with a limestone seabed in about 1650 m of water.

In this paper, quantitative understanding of a frequency-domain nonlinearity indicator is developed. The indicator is derived from an ensemble-averaged, frequency-domain version of the generalized Burgers equation, which can be rearranged in order to directly compare the effects of nonlinearity, absorption, and geometric spreading on the pressure spectrum level with frequency and distance. The nonlinear effect is calculated using pressure-squared-pressure quadspectrum. Further theoretical development has given an expression for the role of the normalized quadspectrum, referred to as Q/S by Morfey and Howell [AIAA J. 19, 986-992 (1981)], in the spatial rate of change of the pressure spectrum level. To explore this finding, an investigation of the change in level for initial sinusoids propagating as plane waves through inviscid and thermoviscous media has been conducted. The decibel change with distance, calculated through Q/S, captures the growth and decay of the harmonics and indicates that the most significant changes in level occur prior to sawtooth formation. At large distances, the inviscid case results in a spatial rate of change that is uniform across all harmonics. For thermoviscous media, large positive nonlinear gains are observed but offset by absorption, which leads to a greater overall negative spatial rate of change for higher harmonics.