A methodology to define underwater acoustic radiated noise norms for small commercial vessel classes using neural networks

 

Amy Deeb and Mae L. Seto

 

Citation: Proc. Mtgs. Acoust. 47 , 070019 (2022); doi: 10.1121/2.0001633

 

View online: https://doi.org/10.1121/2.0001633

 

View Table of Contents: https://asa.scitation.org/toc/pma/47/1

 

Published by the Acoustical Society of America

 

ARTICLES YOU MAY BE INTERESTED IN

 

Decay behavior of a reverberant sound field in a deep fjord

 

Proceedings of Meetings on Acoustics 47, 070020 (2022); https://doi.org/10.1121/2.0001634

 

Measurements of shipping, fin whales, earthquakes and other soundscape components at the Lofoten-Vesterålen Observatory, Norway (2018-2019)

 

Proceedings of Meetings on Acoustics 47, 070017 (2022); https://doi.org/10.1121/2.0001619

 

Automatic recognition of underwater munitions from multi-view sonar surveys using semi supervised machine learning: a simulation study

 

Proceedings of Meetings on Acoustics 47, 070018 (2022); https://doi.org/10.1121/2.0001632

 

Estimation of uncertainties in underwater sound measurements of ships

 

Proceedings of Meetings on Acoustics 47, 070002 (2022); https://doi.org/10.1121/2.0001571

 

Sonar array beampattern bounds and an interval arithmetic toolbox

 

Proceedings of Meetings on Acoustics 47, 055002 (2022); https://doi.org/10.1121/2.0001613

 

Determining optimal time interval and frequency band of ship noise spectrograms for seabed classification

 

Proceedings of Meetings on Acoustics 47, 070016 (2022); https://doi.org/10.1121/2.0001625

 

 

 

A methodology to define underwater acoustic radiated noise norms for small commercial vessel classes using neural networks

 

Amy Deeb and Mae L. Seto

 

Department of Mechanical Engineering, Dalhousie University Faculty of Engineering, Halifax, Nova Scotia, B3H 4R2, CANADA; amy.deeb@dal.ca; mae.seto@dal.ca

 

To address the International Maritime Organization’s underwater radiated noise (URN) reduction guidelines (MEPC.1/Circ.833), vessels undertake rangings to evaluate their URN against classification society norms. However, norms for smaller vessels, like Canada’s inshore fishing fleet, are not established. The proposed methodology informs URN class norms for small commercial vessels, while minimizing individual vessel rangings, through three steps. Firstly, representative class members are acoustically ranged over the range of operating conditions. URN is logged and correlated with on-board measures of structure-borne noise, machinery states, hull fouling, and weather. Secondly, a neural network (NN) is trained to predict the URN from these logged measures. Third and finally, the network’s sensitivity to logged measures is analyzed using permutation importance and dropout. Sensitive features demarcate the class, or must be logged for each vessel, towards accurate predictions. This methodology is demonstrated on the Cape Islanders class. The trained NN predicted the decidecade URN spectrum (100Hz - 50kHz) to an accuracy of 6.6dB re: 1microPascal at 1m. URN prediction sensitivity to in-water conditions, engine speed, engine power, cavitation-induced hull vibrations, and hull fouling extent indicates important features to log. This shows that class-wide analysis of URN can inform small vessel class norms using the proposed methodology.

 

1. INTRODUCTION

 

A growing body of literature shows low-frequency underwater acoustic ambient noise levels have more than doubled in each decade since 1950 and that this can be attributed to the underwater radiated noise (URN) generated by commercial shipping.¹ Marine fauna rely on sound to communicate, find food, and understand and navigate their environment. Increases in ambient noise levels hinder these behaviours. Physical damage to fauna can be caused by high intensity acoustic pulses, while high chronic ambient noise intensity levels can alter marine fauna behaviours. Not all changes in behaviour are harmful, however humans have a responsibility to minimize their impact on the marine fauna.

 

The International Maritime Organization’s (IMO) URN reduction guidelines² provide guidance on set- ting targets to reduce sources of noise from shipping including i) machinery noise transmitted through the hull, ii) cavitation noise from propellers, and iii) hull hydrodynamic noise. Classification society norms (e.g., Lloyd’s Register’s ShipRight procedures on URN⁶) support stakeholders to make informed plans to reduce their levels relative to other vessels in their class (i.e., vessels with similar hull forms, length over all, propulsion systems, auxiliary systems, payload equipment, function, and operating profiles). Similar norms are not yet established for smaller vessels like Canada’s inshore fishing fleet. While smaller vessels radiate lower URN intensity levels, they operate in potentially sensitive regions for extended duration. Mitson and Knudsen showed herring avoided a small vessel⁷ whose URN was not controlled. For fishing vessels, fish avoiding traditional fishing grounds is a critical problem.

 

A proposed transfer function (TF) approach predicts a vessel’s URN through on-board measurements. This facilitates the Captain’s operational decisions to maintain their URN levels below a class threshold. TF for large vessels can be unique due to the complex machinery states and numerous operating setpoints. A TF for small vessels may be easier to develop due to their mechanical simplicity and greater class uniformity.

 

The contribution of this work is a methodology to establish a TF to predict the URN for a small vessel class. Further, the similarities among members of a small vessel class are exploited to create an URN TF to apply to any class member. This class-wide TF is a measure of class norms. Understanding the vessel characteristics, operating conditions, and on-board measurements helps to predict the URN. This provides insight to refine the class definition and design an on-board sensor suite for near real-time URN prediction.

 

This paper also shows the class URN TF can be found from applying machine learning methods. Finally, the proposed methodology is demonstrated on the Cape Islander class as a proof-of-concept. This paper briefly reviews the background, then describes the proposed methodology. Then, the methodology is applied to the Cape Islander class and the results are presented and discussed. Finally, conclusions and recommendations for future work are presented.

 

2. BACKGROUND

 

A. PRACTICES IN SHIP RANGING

 

Ship acoustic ranging measures the vessel URN under a wide range of vessel operation and environmental conditions. These conditions are vessel speed, draft, machinery states (e.g., on-board and in-water propulsion system, auxiliary systems, wet sensors, etc.), crew activity, sea state, and aspect into the sea state, among many others. Together, these contribute to a vessel’s radiated acoustic signature. Typically, an acoustic ranging consists of a vessel transiting towards submerged hydrophones, which measure URN on approach and retreat. ISO 17208³ and others^4,5 define standards to acoustically range mostly large vessels.

 

Small vessel ranging has had less attention. However, new IMO URN reduction guidance 2 will apply to all vessels so small vessel ranging becomes relevant. Small vessel acoustic ranging can differ due to fewer machinery operating states, more possible hydrophone configurations, different responses in sea states, more potential use of shallower water, and other effects of the physics of underwater sound propagation. Further, small vessels lend themselves more readily to class-wide analyses as they may have more similar URN sources across the class, than larger vessels.

 

Such class-wide analyses for small vessels can be revealing towards establishing norms for small craft URN. In particular, if a class-wide TF is meaningful, the impact of a small number of rangings is amplified to inform all vessels of that class. To achieve this, complex and non-linear conditions that relate in situ on-board measurements to farfield URN need to be captured in the TF. The problem appears amenable to a data-driven treatment since a machine learning (ML) approach (i) captures the many conditions that affect a vessel URN, and their sometimes opaque inter-relationships; (ii) takes advantage of the large volumes of URN measurements made in each instant of a ranging, and (iii) the burden of diverse data sources is reduced due to the hypothesized smaller diversity within a small vessel class.

 

B. MACHINE LEARNING IN UNDERWATER ACOUSTICS

 

Machine learning (ML) analyses have been applied to underwater acoustics problems to address challenges⁸ like data corruption, missing or sparse measurements, indirect measurement of source phenomena and large data volumes. Unlike traditional analyses, ML reduces the human effort to identify acoustic features in time or frequency-domain measurements.

 

In general, ML develops a model that relates input features to one or more output variables. For an URN TF, input features can include vessel characteristics (e.g., dimensions, draft, machinery states, etc.), vessel and environment interactions (e.g., water temperature, extent of hull fouling, maneuvers, sea state, etc.) and operational conditions (e.g. speed, aspect into sea state, etc.). The output could be decidecade URN frequency-domain spectra that relate to the on-board measurements. This becomes a regression problem since the output is a vector of real-valued scalars. Since labelled examples of expected outputs for input feature vectors are available to train the model, supervised learning can be applied. Given sufficient training examples, the model correctly predicts the output for an input feature vector not used in training.

 

Supervised regression is applied in several ways. Linear regression is limited to model linear relationships, while support vector machines, neural networks, and random forest regressors capture non-linear relationships. In the early phase of this project, neural networks showed the most promise of the three. This is attributed to random forests being better suited to large datasets, and an inappropriate support vector machine kernel being used due to incomplete understanding of the patterns within the complete dataset at that time. Revisiting these or other regressors with the complete dataset may improve performance, however this was not within the scope of this study.

 

Neural networks are applied to problems with large numbers of input features with complex relationships to the output vector that are unknown. A neural network, also a multi-layer perceptron (MLP), is a fully connected network where values from one layer are passed through weighted links to subsequent layers. Training an MLP is the process of identifying the weights for each link in the network.⁹ A disadvantage of MLPs is their lack of transparency in the predicted output and thus an inability to express this in an analytic form. Feature analysis techniques can address this limitation. Two examples of MLP feature analyses are dropout¹⁰ and permutation importance¹¹ and will be discussed later.

 

3. METHODS

 

A. SMALL VESSEL CLASS-WIDE URN TF

 

To support small vessels in establishing their URN baseline levels, a class-wide methodology is proposed. This methodology is formulated based on two hypotheses: (1) for a particular vessel, its URN can be predicted if its operational conditions are known, and, (2) a class can be defined such that the URN of a class member is similar to others in its class under the same operating conditions.

 

For vessel classes where these two hypotheses hold, the vessel URN can be predicted based on a vessel class TF. The advantage is that class operators can understand, and take actions to reduce, their URN levels, without ranging the URN of each individual vessel under every operating condition. The challenge is to ensure the predicted URN class’ accuracy is sufficient to meet regulatory limits, and provides enough information for vessel operators to make informed decisions on their vessel’s URN levels.

 

 

Figure 1: High-level overview of the three steps for the proposed Small Vessel Class-wide URN TF Methodology.

 

The proposed methodology to establish small vessel TF (Fig. 1) has three steps. Firstly, representative class members are acoustically ranged over a range of vessel operation and environmental conditions. Each vessel’s recorded URN is correlated with on-board measurements of its structure-borne noise, machinery states, hull fouling, and weather. This assumes sufficient class insight to identify representative members. Otherwise, a larger set of class members are selected for ranging. Then, unique class features are identified. For small vessels, these features are proposed to be: (i) length overall (LOA), draft, and tonnage; (ii) propulsion system; (iii) auxiliary equipment; (iv) hull form and hydrodynamic features; (v) cavitation inception speed, the extent that cavitation dominates their URN, and (vi) operating speeds.

 

In the second step, a TF is computed to predict the vessel’s URN based on on-board measurements. A neural network is proposed as such networks can learn complex, non-linear relationships in multi-variate systems. In this case, the on-board measurements are the input features while the URN is the output. A supervised learning approach is proposed using measurements collected from the first step. While training the neural network, correlations in input features are identified as correlated features do not provide additional information to the network and can obscure the learned relationships. After training, the learned URN TF predicts the vessel’s near-real-time URN through a sensor suite which measures the conditions. To design such a sensor suite the features of the neural network are analyzed in the third step.

 

The final step analyzes the sensitivity of the predicted URN to the on-board measurements. Features that the predicted URN is sensitive to provides insight to: a) refine the class definition to recognize these sensitive features as common to a class and b) define the necessary on-board measurements for individual vessels to achieve prediction accuracy. Dropout and permutation importance are proposed for feature analysis. As a proof-of-concept, this methodology was applied to the Cape Islander class.¹² This specific implementation is described in the following sections.

 

B. URN DATASET COLLECTION

 

Between fall 2020 and fall 2021, five members of the Cape Islander class were ranged three times in Shad Bay, Nova Scotia, Canada. This created a dataset spanning 15 trials where each trial consisted of four to six runs. A run consists of two hydrophone passes on reciprocal headings which were averaged to compensate for directional effects like current, wind, or vessel aspect into the seas. Each run was performed at different engine speeds. In all trials, the boat path’s midpoint is 100 m from the hydrophone and starting and ending points are 200 m apart. All trials were conducted during daylight hours. The bay is located on Nova Scotia’s South Shore approximately 24 km from the Port of Halifax. The sea bottom is primarily sandy with plant growth particularly in the summer months. The channel where the runs occurred was 3.5 km long by 0.5 km wide and 20 m deep. While there was limited ship/boat traffic in the area, and testing halted when other vessels transited through, there were uncontrollable ambient underwater noise sources like Port of Halifax shipping activities and above-water activities like railways (at ∼ 20 km) and lawn mowers (at ∼ 500 m) which insonified the water. The underwater ambient before and after each trial was recorded.

 

Between trials, the weather and vessel conditions changed. Most notably, the hulls were cleaned and a novel hull and propeller coating, the XGIT-Fuel and XGIT-Prop, respectively, were applied.¹³ The impact of the hull and propeller coatings are reported separately. The hull fouling extent is noted firstly, with whether the hull was cleaned before the trial and secondly, fouling was evaluated through the Naval Ships’ Technical Manual (NSTM) guidance. These are both input features.

 

An accelerometer was placed on the hull above the propeller to capture cavitation (which insonified the hull) and three others were placed on the engine mount to measure its source levels of hull-transferred noise. A spectrum for each accelerometer was generated by averaging the two passes in a run. Similarly, the average spectrum captured by a sound meter in the engine compartment measured the air-borne noise.

 

For URN, the sound pressure is collected in 1.5 s duration records at a sampling frequency of 204.8 kHz and corrected as suggested by the hydrophone manufacturer. After accounting for propagation loss and transforming to frequency domain, the resulting radiated noise level (RNL) spectra are averaged to produce one URN spectrum per run. The hydrophone specifications indicate good response between 100 Hz and 50 kHz. Since each run corresponds to a different engine speed for that trial, it is not possible to assess the accuracy of the run’s spectrum by comparing multiple measurements for the same condition. Past experience showed a measurement uncertainty of 5 dB for the system.

 

C. URN TRANSFER FUNCTION COMPUTATION

 

An MLP predicted the URN level (RNL) for each bin of the vessel’s decidecade URN spectrum based on on-board measurements. The measured URNs were divided into k subsets (folds) with ( k − 1 ) folds used for training and the final fold for testing (verification). K-folds cross-validation assessed the prediction accuracy’s consistency. This was achieved by comparing the measured URN to the predicted one and calculating the average difference across frequency bins for the verification examples.Training and verification are repeated k-times for each combination. For the Cape Islanders, the 72 runs were divided into five folds. Each fold had at least one run that represented each: i) vessel, ii) trial, and iii) engine speed to ensure the greatest measurement diversity and that each verification set was representative of the complete dataset.

 

Implementation of the MLP was achieved with the scikit-learn MLP Regressor.¹⁴ Multiple MLP structures were evaluated. The best performance was achieved for an MLP with a single hidden layer of 200 nodes. The input layer had 184 nodes (the number of input features) and the output layer had 27 nodes (the number of bins in the URN decidecade spectrum between 100 Hz and 50 kHz). Hyper-parameters were evaluated and the default values were most appropriate for all except two: the identity activation function per- formed better than the rectified linear function and the Limited-memory Broyden–Fletcher–Goldfarb–Shanno (LBFGS) solver out-performed ‘adam’. While these hyper-parameters were best for this dataset, that may not be so for all URN TFs.

 

D. MACHINE LEARNING FEATURE ANALYSIS

 

After computing the vessel class URN TF, the next step was to explain the predicted URN for the given on-board measurements. Three forms of feature analysis helped develop insights into the URN prediction: correlation, dropout and permutation importance. Correlation highlights relationships across input features. Correlated features do not improve prediction accuracy and can obscure underlying relationships between input and output vectors. As the underlying relationships are a priority, only one feature of a correlated set was retained. Dropout evaluates the impact of removing a feature from the input vector. If the prediction accuracy drops, the model is sensitive to that feature, it should be measured and it could be a class identifier. For example, models sensitive to vessel dimensions may not perform well for vessels with dimensions that vary from those represented in the training set. With permutation importance, a feature’s measurement is replaced with a random number based on the feature’s measurement distribution in the training set. If such a random number is as good as the measured value then: i) that feature may be consistent across all members in the training set; ii) the measurement uncertainty may be large, or iii) the feature may be irrelevant to predicting the URN. On the other hand, features with high permutation importance should be measured in-situ to ensure the URN is predicted accurately. Taken together, these feature analyses identify a small set of features to refine the class definition and to develop an appropriate sensor suite to measure on-board conditions used by the class-wide URN TF.

 

4. RESULTS AND ANALYSIS

 

A. CORRELATION ANALYSIS

 

Correlations between on-board measurements in the Cape Islander’s rangings are presented in the correlation matrix of Fig. 2. Squares with high intensity colours (either red or blue) indicate strong correlations, while pale colours indicate little correlation. The highest correlations are noted with an asterisk (*). For clarity, this correlation analysis excludes the on-board measurements of the four accelerometers’ vibrations and the sound metre’s air-borne sound. Correlations between portions of those spectra are expected and segmenting the spectra used in training was not considered in this study.

 

 

Figure 2: Correlation matrix for on-board measurements for each pass. Squares with intense colours indicate strong correlations. The strongest correlations are marked with an asterisk (*). For example, Beam and HP are strongly positively correlated, NSTM and whether the hull was clean are strongly negatively correlated.

 

Fig. 2 shows strong correlations between five groups of variables. In each group, the correlated features were evaluated to remove those that were not expected to provide insight. The strongly correlated groups with features that were retained are bold-faced below:

 

• NSTM level and clean boolean

• planned speed, speed over ground, reported RPM, RPM, and fuel flow rate

• vessel dimensions (beam, draft, length over all, dry weight)

• vessel dimensions , fuel level, and engine HP

• water temperature and air temperature.

 

Since some highly correlated variables were retained, this must be considered when evaluating the feature dropout results. This is due to lower impact from dropping out one of a pair of highly correlated variables as the remaining variable provides similar information to the model.

 

B. SPECTRUM PREDICTION

 

After removing the variables suggested by the correlation analysis and processing the measurements for each pass in a run, the MLP training was performed using 5-fold cross-validation. The prediction errors are summarized in Table 1. It is also meaningful to evaluate the prediction error in each decidecade frequency bin as shown in Fig. 3 with colours representing each fold and the error bars showing one standard deviation from the mean prediction error for that fold’s verification set.

 

Table 1: Prediction error of Cape Islander URN TF showing consistency among the folds. Good prediction accuracy and precision was achieved, given 5 dB measurement uncertainty.

 

 

The results across all five folds are consistent meaning no particular run was poorer in accuracy than the others, nor that the URN TF was less effective for one of the vessels or conditions. On average, the URN prediction error was 6.63 ± 4.57 dB re 1 µ Pa at 1 m. This was generally higher at lower frequencies (below 400 Hz), which may be due to the higher low frequency ambient noise. The prediction error is surprisingly low given the measurement uncertainty was estimated at 5 dB, and minimal control of ambient noise levels were applied. This shows confidence in the efficacy of the proposed TF methodology.

 

C. MACHINE LEARNING FEATURE ANALYSIS RESULTS

 

The dropout analysis considered only named features (Figure 2) and excluded the on-board vibration and in-air sound spectra. The five most impactful features based on the dropout analysis were: (1) Day of Year, (2) Engine HP, (3) Power, (4) Fuel Flow Rate, and (5) Wind Speed.

 

 

Figure 3: Prediction error in each decidecade frequency bin for each of the five folds showing consistent error across frequency bins and between folds. Thus, the TF does not prefer a particular data subset or frequency range.

 

Recall that retaining correlated features (i.e. both vessel and engine speed, all vessel dimensions, and water and air temperatures) makes it unclear whether these features are not on the list because the model is insensitive to them or when one of a correlated pair was removed the other provided the same information to the model. Permutation importance does not suffer from this limitation. The ten most important features based on that analysis were: (1) Day of Year, (2) RPM, (3) Engine HP, (4) PropVib 10 Hz, (5) PropVib 12.59 Hz, (6) PropVib 15.85 Hz, (7) Power, (8) NSTM, (9) PropVib 19.95 Hz, and (10) PropVib 7.94 Hz. In this list, PropVib refers to a frequency bin for the accelerometer on the hull above the propeller.

 

Dropout and permutation importance analyses show expected trends where the model was sensitive to cavitation, vessel dimensions, the drive train, and vessel cleanliness. A notable Cape Islander class feature is that their URN is cavitation dominated. This is in the high permutation importance scores for several low frequency bins of the accelerometer located above the propeller. In the correlation analysis the engine power was highly correlated with the LoA for the Cape Islanders (bigger engines on bigger vessels). As well, the engine power’s place in both analyses show the URN prediction generally depends on the vessel dimensions. Similarly, the impact of speed on URN prediction is shown by the presence of power, engine speed and fuel flow rate in these analyses which are correlated for diesel engines. The presence of NSTM in the permutation importance analysis reveals that hull fouling impacts the URN.

 

The sensitivity to the ‘day of year’, seen in both feature analyses, was unexpected. On reflection, ‘day of year’ represents environmental variables that impact underwater noise propagation like sound velocity profile, water temperature, salinity, underwater ambient, and sea bottom growth. Further, Cape Islander fishing operations are seasonal with annually varying installed equipment and maintenance (cleaning) activities. These seasonal operations were not captured in any features and may have increased the model’s sensitivity to the date. Future study is needed to capture the factors that result in this ‘day of year’ sensitivity.

 

5. CONCLUSION AND RECOMMENDATIONS

 

The proposed methodology to establish transfer functions (TF) that predict underwater radiated noise (URN) from on-board measurements for small vessel classes has merit for the Cape Islander class.

 

The Cape Islander class transfer function predicted the decidecade underwater radiated noise spectrum between 100 Hz and 50 kHz with an average error of 6.63 dB re: 1 µ Pa at 1 m. Feature analyses correctly revealed the Cape Islander class URN is dominated by cavitation and the propulsion diesel engine. Further, for accurate URN predictions, the vessel / engine speed and hull cleanliness must be measured or noted.

 

This proof-of-concept confirms the hypothesis that small vessel classes are quite similar in terms of the form, material, and stiffness of their hulls, and the high levels of cavitation from their propellers. This results in their URN levels being sufficiently similar for class-wide analyses to be meaningful. Further, a machine learning approach yields a class-wide URN transfer function. However, this analysis was limited to lower sea states and the small number of operating states common to Cape Islanders in Canada’s inshore fishing fleet.

 

Future trials should determine which environmental features explain the underwater radiated noise transfer function’s sensitivity to the day of the year. Rangings of counter examples would demonstrate the limitations of the Cape Islander underwater radiated noise transfer function on non-Cape Islanders. Additionally, studies of other vessel classes would evaluate the proposed methodology for broader applications.

 

ACKNOWLEDGMENTS

 

This work was completed in the first author’s capacity as an industrial post-doctoral fellow supported in part by Mitacs through the Mitacs Accelerate program. The data used was collected under the “Underwater Radiated Noise (URN) and Green House Gas Reduction Program for Canada’s Inshore Fishing Craft” Project funded by the Transport Canada Innovation Centre. The authors are grateful to the Applied Technologies Group of Lloyd’s Register, EMCC Electronics Inc. and Graphite Innovation and Technologies (GIT) for their efforts in collecting, validating, and making available the data used in this work.

 

REFERENCES

 

1. N. Jones, “The Quest for Quieter Seas”, Nature 568, 158-161 (2019).
2. International Maritime Organization Marine Environment Protection Committee, “Guidelines for the Re- duction of Underwater Noise from Commercial Shipping to Address Adverse Impacts on Marine Life”. MEPC.1/Circ 833 Annex (2014).
3. Underwater acoustics — Quantities and procedures for description and measurement of underwater sound from ships, ISO 17208-1:2016 and ISO 17208-2:2019.
4. Quantities And Procedures For Description And Measurement Of Underwater Sound From Ships - Part 1: General Requirements, ANSI/ASA S12.64-2009/Part1 (2014).
5. Standards for Use When Measuring and Reporting Radiated Noise Characteristics of Surface Ships, Submarines, Helicopters, etc. in Relation to Sonar Detection and Torpedo Acquisition Risk, NATO - STANAG 1136 (1995).
6. Lloyd’s Register Group Limited, “ShipRight Additional Design and Construction Procedure for the De- termination of a Vessel’s Underwater Radiated Noise”, (2018). Available: https://www.lr.org/en/shipright-procedures/. Retrieved: April 25, 2022.
7. R. B. Mitson and H.P. Knudsen, “Causes and effects of underwater noise on fish abundance estimation”, Aquatic Living Resources 16, 255-263 (2003).
8. M.J. Bianco et al., “Machine learning in acoustics: Theory and applications”, J. Acoust. Soc. Am. 146, 3590-3628 (2019).
9. K. P. Murphy, “Feedforward neural networks (multilayer perceptrons)” in Machine Learning: A Probabilistic Perspective, Cambridge, Massachussetts: The MIT Press, 563-579 (2012).
10. A. Akalin, “Variable Importance” in Computational Genomics with R, (2020). [E-book] Available: http://compgenomr.github.io/book/.
11. C. Molnar, “Permutation Feature Importance” in Interpretable Machine Learning (2022). Available: https://christophm.github.io/interpretable-ml-book/.
12. W. M. O’Leary, “The coming of the cape boats” in Tancook Schooners, 178-181 (1994).
13. M. Gaier et al., “Using Paint to Reduce Underwater Noise for the Shipping Industry”, Paint and Coatings Industry magazine, February 2022 [Online serial] (2022).
14. Scikit-Learn Developers, “MLPRegressor”. Available: https://scikit-learn.org/stable/modules/generated/sklearn.neural_network.MLPRegressor.html. Retrieved: April 25, 2022.
15. F. Pedregosa et al., “Scikit-learn: Machine Learning in Python”, Journal of Machine Learning Research 12 2825-2830 (2011).