Ocean ambient noise modeling method based on multi-stage modeling regression of wind speed

Abstract

The application belongs to the field of ocean acoustics, and discloses a marine environmental noise modeling method based on multi-stage modeling regression of wind speed, which firstly collects and matches marine noise and wind speed data, and extracts noise power spectrum; then fits the noise spectrum into a gamma distribution, and extracts shape and scale parameters as physical statistical characteristics; further, a two-stage prediction model is constructed: the first stage uses Gaussian process regression to establish a mapping from wind speed to distribution parameters; the second stage uses a hybrid kernel support vector regression model that fuses Wenz empirical physical kernel and RBF data kernel to predict the noise spectrum in the full frequency band. Through the framework of "physical statistical characteristic guidance", the application decomposes complex prediction problems, effectively fuses physical priori and data-driven learning, significantly improves prediction accuracy, generalization ability and model robustness under data scarcity conditions, and provides a high-precision noise prediction tool for marine wind field monitoring and other applications.

Description

Technical Field

[0001] This invention relates to the field of marine acoustics, and more particularly to a method for modeling marine environmental noise based on multi-stage modeling regression of wind speed. Background Technology

[0002] The spectral characteristics of marine environmental noise are closely related to sea surface wind speed. However, traditional monitoring methods typically rely on direct measurement using hydrophones. While this method offers high accuracy, it suffers from limitations such as limited applicability, high cost, and difficulty in data acquisition. Marine environmental noise modeling techniques can predict marine environmental noise based on readily available data (e.g., wind speed, precipitation). Existing noise modeling techniques are mainly divided into physical models and data-driven models. While the former can explain the basic relationship between wind speed and noise level, it ignores the non-stationary nature of marine environmental noise, leading to significant prediction errors under complex sea conditions. Pure data-driven models primarily incorporate various deep learning methods. Although they can capture non-linear relationships, they suffer from a lack of physical constraints, difficulty in data acquisition, and susceptibility to false predictions when training data is insufficient or noise pollution is severe. However, a gap exists between data-driven and model-driven approaches; simple combinations often fail to substantially improve prediction accuracy. Summary of the Invention

[0003] The purpose of this invention is to provide a modeling method for marine environmental noise based on multi-stage modeling regression of wind speed, so as to solve the above-mentioned technical problems.

[0004] To address the aforementioned technical challenges, this invention proposes a marine environmental noise modeling method based on multi-stage modeling and hybrid physical kernel support vector regression. The core of this method is the construction of a hybrid modeling framework guided by "physical statistical features." Its innovation lies first in the progressive mapping from "wind speed → statistical distribution parameters → noise spectrum," which decomposes the complex end-to-end prediction problem into two more interpretable and stable sub-problems, achieving deep synergy between the two stages. Specifically, the synergistic mechanism involves the gamma distribution parameters extracted in the first stage serving as "physical guiding features," injecting prior physical statistical constraints into the purely data-driven SVR model of the second stage. This effectively limits the model's hypothesis space, thereby improving fitting accuracy while enhancing the model's generalization ability and physical rationality under conditions of scarce data or noise interference.

[0005] The specific technical solution of the marine environmental noise modeling method based on wind speed multi-stage modeling regression of the present invention is as follows:

[0006] A multi-stage modeling regression method for marine environmental noise based on wind speed includes the following steps: Step 1: Data acquisition and alignment: Acquire measured marine environmental noise data and corresponding spatiotemporal wind speed data of the target sea area, and construct a sample library of noise and wind speed pairings; Step 2: Noise spectrum feature extraction: Perform power spectral density analysis on the noise data, extract power spectrum features within the preset wind-induced noise dominant frequency band, and form a noise power spectrum dataset; Step 3: Physical statistical parameter inversion: Fit each noise spectrum in the noise spectrum dataset to a two-parameter probability distribution model, thereby compressing the high-dimensional spectrum into two low-dimensional physical statistical parameters representing its energy distribution pattern, forming a corresponding training set of wind speed and physical statistical parameters; Step 4: One-stage statistical parameter prediction modeling: Based on the training set of wind speed and physical statistical parameters, establish a prediction model from wind speed to the two physical statistical parameters as the first-stage model; Step 5: Two-stage hybrid kernel regression modeling: Use the physical statistical parameters predicted by the first-stage model and the original wind speed as input features, and the noise spectrum as the output target to construct a support vector regression model.

[0007] Furthermore, in step 1, an underwater glider equipped with a high-precision hydrophone is used to continuously collect three-dimensional noise data in the target sea area. A spatiotemporally aligned dataset is constructed using a bilinear interpolation algorithm. Simultaneously, the wind speed true values ​​of the corresponding spatiotemporal points are obtained by accessing the marine meteorological service API. The raw data is stored in .dat format. After the depth, time, latitude and longitude and profile information are extracted by MATLAB, it is converted into an Excel spreadsheet for storage. The audio signal is saved in .wav format, forming an original dataset containing 138 samples.

[0008] Furthermore, in step 2, during the ascent / descent phase of each profile, sampling is performed every 50m within a depth range of 50m to 800m, with 0.25 latitude and longitude as the spatial resolution of the data, to construct a noise-wind speed dataset.

[0009] Furthermore, in step 3, the two-parameter probability distribution model is a gamma distribution, and the two low-dimensional physical statistical parameters are the shape parameter and scale parameter of the gamma distribution, respectively. The fitting process specifically includes: S31: using a Gaussian kernel function to estimate the kernel density of each noise spectrum to obtain a smooth spectral probability density curve; S32: using the maximum likelihood estimation method to fit the spectral probability density curve to the gamma distribution model and extract the optimal shape parameter and scale parameter; S33: using the Kolmogorov-Smirnov test to evaluate the goodness of fit of the gamma distribution.

[0010] The choice of the gamma distribution as the representation of the statistical characteristics of the noise spectrum is based on the following two considerations: First, the gamma distribution is a flexible two-parameter distribution whose shape and scale parameters can effectively describe different degrees of skewness and dispersion of the data, which is highly consistent with the characteristics of the energy distribution of marine environmental noise changing with wind speed. Second, compared with directly using high-dimensional raw power spectral density (PSD) data or simple statistical moments (such as mean and variance), the gamma distribution parameters condense the overall morphological characteristics of the noise spectrum with extremely low dimensionality (only two parameters). This compact representation not only reduces the input dimensionality and learning difficulty of the subsequent data-driven model, but it is also easier to establish a smooth and robust mapping relationship with wind speed, laying an interpretable physical and statistical foundation for the entire multi-stage model.

[0011] Furthermore, in step 4, the first stage model is a Gaussian process regression model; the Gaussian process regression model uses a constant mean function and a radial basis function kernel, and optimizes the model hyperparameters by maximizing the marginal likelihood.

[0012] Furthermore, in step 5, the support vector regression model employs a hybrid kernel function, which is a linear combination of a data-driven kernel and a physical kernel constructed based on physical empirical laws. The data-driven kernel is a radial basis function kernel. The physical kernel is constructed based on the Wenz marine environmental noise empirical curve. Its construction method is as follows: calculate the empirical noise spectrum level corresponding to the two input features according to the Wenz curve, and use a certain monotonically decreasing function value of the difference between the two noise spectrum levels as the value of the physical kernel.

[0013] Furthermore, the hybrid kernel function is a weighted sum of the data-driven kernel and the physical kernel, and its weight is controlled by a hybrid coefficient that is adjustable between 0 and 1. The hybrid coefficient, together with the penalty parameter, the insensitive loss parameter, and the parameters of each kernel function of the support vector regression model, are used as hyperparameters and are jointly optimized by grid search combined with K-fold cross-validation.

[0014] Furthermore, during the optimization process, the search range of the mixing coefficients is set to 0.3 to 0.7, and the optimization objective is to minimize the root mean square error on the validation set.

[0015] The above steps represent the collaborative and integrated stages of multi-stage modeling. The gamma distribution parameters (Shape and Scale) predicted by the GPR model in step 3 are used as key "physical guiding features," forming the input vector of the support vector regression model together with the original wind speed features. This design achieves deep coupling between the physical model and the data-driven model: the "wind speed-statistical distribution" relationship learned by the GPR model is injected into the SVR model in an explicit, low-dimensional feature form, providing strong physical prior constraints for the data-driven learning process. Based on this, a dual-core fusion SVR model is constructed:

[0016] Physical kernel: Embedded with the Wenz empirical curve kernel function to characterize the physical laws governing wind speed and noise level;

[0017] Data kernel: RBF kernel is used to capture nonlinear residuals.

[0018] The model takes wind speed as input and outputs predicted noise spectrum values ​​across the entire frequency band. The kernel weight coefficients (λ=0.3-0.7) are optimized through grid search to improve prediction accuracy while maintaining physical consistency.

[0019] Furthermore, it also includes step 6: model training and prediction: using training data to train and optimize the first-stage model and the hybrid kernel support vector regression model to obtain a complete marine environmental noise prediction model, which is used to predict the noise spectrum based on the input wind speed.

[0020] Furthermore, step 6 is followed by a verification step: evaluating the performance of the marine environmental noise prediction model using an independent test set, with evaluation metrics including mean square error, root mean square error, mean absolute error, and mean relative error.

[0021] The marine environmental noise modeling method based on wind speed multi-stage modeling regression of the present invention has the following advantages:

[0022] (1) In the measured data of the South China Sea, the correlation coefficient between the predicted spectrum of the 100Hz~16kHz frequency band and the measured data spectrum of the underwater glider reached 0.93, which is 27% lower than the error of the classic marine environmental noise spectrum Wenz model. The prediction error is less than 0.6dB in the typical wind speed range of 8~12m / s. Its evaluation metrics are significantly better than traditional data models: the mean squared error (MSE) is 3.78, a 6.2% reduction compared to the traditional SVR's 4.03, and a 24.1% reduction compared to the theoretical model's 4.98; the root mean square error (RMSE) is 1.944, a 3.2% reduction compared to the traditional SVR's 2.008, and a 12.9% reduction compared to the theoretical model's 2.232; the mean absolute error (MAE) is 1.521, a 9.4% reduction compared to the traditional SVR's 1.679, and a 14.6% reduction compared to the theoretical model's 1.782; and the mean relative error (MRE) is 3.01%, a 7.7% reduction compared to the traditional SVR's 3.26%, and a 15.7% reduction compared to the theoretical model's 3.57%.

[0023] (2) Underwater noise data is less affected by meteorological interference and maintains stable performance under adverse conditions, improving the anti-interference ability of traditional methods by 40%.

[0024] (3) The mapping relationship between wind speed and noise spectrum is explicitly expressed by the gamma distribution parameters. The model output has a clear physical meaning, which is conducive to subsequent model fine-tuning according to regional characteristics.

[0025] (4) Furthermore, the multi-stage modeling framework adopted in this method fundamentally reduces the learning complexity of the data-driven stage and the dependence on the amount of training data by using gamma distribution parameters to provide a compact and low-dimensional representation of the complex features of the noise spectrum. This design enables the model to efficiently and robustly learn the mapping relationship between wind speed and noise spectrum even under limited experimental sample conditions (such as 138 sets of data in this embodiment). This is one of the intrinsic reasons why the model achieves excellent generalization ability and significantly reduces various error indicators.

[0026] This invention achieves high-precision prediction of noise spectrum across the entire frequency band from raw noise signals through a four-stage architecture of "data acquisition - feature analysis - physical parameter inversion - hybrid kernel modeling". It innovatively integrates the three-dimensional observation advantages of underwater gliders, the statistical physical constraints of gamma distribution, and the machine learning optimization capabilities of hybrid kernel SVR, constructing a new generation of marine environmental noise wind speed prediction method with high precision, strong real-time performance, and wide adaptability. This provides key technical support for fields such as marine wind field monitoring, climate model validation, and marine energy development. Attached Figure Description

[0027] Figure 1 This is a flowchart of a multi-stage marine environmental noise prediction model according to an embodiment of the present invention;

[0028] Figure 2 This is a comparison chart of the gamma coefficients between the nonlinear polynomial fitting and the GPR fitting in an embodiment of the present invention.

[0029] Figure 3 This is a comparison chart of the prediction results of the method of this invention and other models on the test set data. Detailed Implementation

[0030] To better understand the purpose, structure, and function of this invention, the invention will be described in further detail below with reference to the accompanying drawings.

[0031] The core of this invention's multi-stage modeling regression method for marine environmental noise based on wind speed lies in a hybrid modeling framework guided by physical statistical features. This framework does not simply string together existing models; instead, it decomposes the complex end-to-end prediction problem into two more interpretable and stable sub-stages through a progressive mapping of "wind speed → statistical distribution parameters → noise spectrum," achieving deep synergy between the two stages through the crucial hub of gamma distribution parameters. The first stage uses Gaussian process regression (GPR) to map wind speed into low-dimensional physical statistical features (gamma distribution parameters). The second stage inputs these features along with the original wind speed into a hybrid kernel support vector regression (SVR) model that incorporates physical experience, ultimately achieving high-precision full-band noise spectrum prediction. The specific implementation steps are described in detail below with reference to the accompanying drawings.

[0032] Step 1: Data Acquisition and Spatiotemporal Alignment

[0033] A high-precision hydrophone was used in an underwater glider to continuously collect three-dimensional noise data (time, latitude, longitude, and depth) in the target sea area, with a sampling frequency of 16kHz. A spatiotemporally aligned dataset was constructed using a bilinear interpolation algorithm, and the ground truth wind speed values ​​for the corresponding spatiotemporal points were obtained simultaneously from the European Centre for Medium-Range Weather Forecasts (ECMWF) meteorological service API. A "noise-wind speed" paired sample library was formed, covering the diurnal and tidal cycles to ensure sample representativeness. The raw data was stored in .dat format. After extracting depth, time, latitude, longitude, and profile information using MATLAB, the data was converted to an Excel spreadsheet and stored. The audio signals were saved in .wav format, forming an original dataset containing 138 samples.

[0034] Step 2: Power spectral density analysis and feature extraction

[0035] Power density spectrum (PSD) analysis was performed on the noise signal (.wav format audio file). Based on the physical mechanism of wind speed affecting marine environmental noise, mid-to-high frequency bands highly correlated with wind speed were selected, retaining the dominant frequency bands for wind-generated noise. Preferably, PSD features within the 100Hz~16000Hz frequency band were extracted, discarding frequency bands with negligible wind speed influence to avoid overfitting due to variations in other conditions. During the ascent / descent phases of each profile, sampling was performed every 50m within a depth range of 50m to 800m, using 0.25 latitude and longitude as the spatial resolution of the data to construct a noise-wind speed dataset. Partial data is shown in Table 1.

[0036] Table 1: Excerpt of the dataset from the example

[0037]

[0038] Step 3: Gamma distribution parameter inversion

[0039] This step aims to compress the high-dimensional, non-stationary noise power spectrum into a low-dimensional feature vector with clear physical and statistical meaning. First, for each sample's noise power spectrum (PSD) data, a Gaussian kernel function is used to perform kernel density estimation (KDE) to obtain its smooth probability density distribution curve. The Gaussian kernel provides good smoothness and continuity. Then, based on the noise density spectrum, the maximum likelihood estimation (MLE) method is used to fit a gamma distribution. The gamma distribution is chosen as an ideal model for characterizing the distribution of noise energy due to its two-parameter flexibility and its good ability to characterize positively skewed and long-tailed data.

[0040] To obtain the best fit, the bandwidth parameter of the KDE needs to be adjusted to balance curve smoothness and detail preservation. After fitting, the goodness of fit is evaluated using the Kolmogorov-Smirnov (KS) test. Experimental data show that the average p-value of the KS test for the fitting results is greater than 0.05, indicating that the gamma distribution is statistically acceptable. Therefore, we uniquely represent the complex noise spectrum under each wind speed condition with two parameters (Shape, Scale), where Shape represents the shape parameter and Scale represents the scale parameter. These two parameters encapsulate the core morphological information of the spectrum, such as energy concentration and skewness characteristics, and constitute the training set for subsequent regression analysis. This process realizes the transformation from the original signal to physical statistical characteristics, such as... Figure 1 As shown in the flowchart, the fitting effect of random samples is as follows: Figure 2 As shown.

[0041] Step 4: Prediction of distribution parameters based on Gaussian process regression (one-stage modeling)

[0042] After obtaining the training set, a prediction model needs to be established from wind speed to gamma distribution parameters (Shape, Scale). This is the first stage of the multi-stage framework. Gaussian process regression (GPR) is chosen here, mainly based on its two major advantages: First, as a non-parametric Bayesian model, GPR can flexibly capture the potential nonlinear and complex functional relationship between wind speed and parameters without pre-specifying the functional form; second, GPR can provide variance estimates for each prediction point, i.e., quantify the uncertainty of the prediction, which provides important information for evaluating the reliability of the first-stage output.

[0043] The specific implementation is as follows: Using the wind speed-parameter observation set obtained in step 3 as training data, independent GPR models are established for the shape parameter and scale parameter, respectively. The model uses a constant mean function and a radial basis function (RBF) kernel. The training process includes two optimization steps: first, optimizing the maximum likelihood estimate of the gamma distribution parameter at each wind speed; then, optimizing the hyperparameters of the GPR itself (such as length scale and signal variance) to maximize the marginal likelihood of the model. As shown in Tables 2 and 3, the prediction determination coefficients R² of the GPR model for the two parameters reach 0.8768 and 0.8498, respectively, significantly outperforming comparative methods such as linear regression and Lasso regression. This verifies the superiority of GPR in capturing the complex and smooth mapping relationship between "wind speed and statistical parameters," laying a solid foundation for the accuracy of the entire multi-stage model. See the prediction curves and comparisons below. Figure 2 .

[0044] Table 2: Fitting Effects of Different Regression Methods on Shape

[0045]

[0046] Table 3: Fitting Effects of Different Regression Methods on Scale

[0047]

[0048] Step 5: Hybrid kernel support vector regression incorporating physical experience (two-stage modeling)

[0049] This stage is the integration and final output of the multi-stage model. Its core innovation lies in constructing a hybrid kernel function SVR model that integrates prior physical knowledge and data-driven learning. This model uses the gamma distribution parameters predicted by the previous stage's GPR (as key "physical guidance features") and the original wind speed as input features, and the full-band noise spectrum as output.

[0050] In this invention, a physical kernel function combining the RBF and a Wenz empirical noise curve is used. The hybrid kernel function Kmix is ​​defined as a convex combination of the data-driven kernel KRBF and the physical kernel Kphysical.

[0051]

[0052] in The input feature vector, For traditional RBF kernel functions, This is a physical kernel function constructed based on the Wenz curve, reflecting the physical influence of characteristics and frequencies on the noise spectrum. This is a mixing coefficient used to adjust the weights between the data-driven and physical-constraint components.

[0053] To embed known physical empirical laws, we construct a physical kernel based on the classic Wenz empirical curve for marine environmental noise. The Wenz curve is a classic empirical model in the field of marine environmental noise, and the Wenz curve gives the wind speed... and frequency With noise spectrum level The empirical relationship is expressed as follows:

[0054]

[0055] in For frequency, For wind speed, These are empirical coefficients; if the two samples... The closer the values ​​are, the more similar we consider them under physical laws. Therefore, we use the square of the difference in their noise levels as a measure of "distance": based on this, a physical kernel function can be constructed:

[0056]

[0057] Combining the above RBF kernel function, the final expression for the hybrid physics kernel function is:

[0058]

[0059] Will Substituting the values, we obtain the SVR optimization objective function based on the Wenz curve and the hybrid physics kernel, thus yielding the corresponding optimization objective function. By training the model on the dataset, we can ultimately obtain a prediction model that integrates physical priors and data-driven characteristics.

[0060] In this hybrid kernel function, the physical kernel ensures that the prediction conforms to the basic physical trend. It ensures that the model tends to give predictions that conform to the Wenz empirical law (i.e. the basic physical trend) during the learning process, preventing the model from generating outliers in areas with sparse data or high noise, and enhancing the robustness of extrapolation. The data-driven kernel is responsible for learning the residuals and nonlinear patterns in the data to improve accuracy, thereby improving the absolute accuracy of the model in specific sea areas and complex conditions.

[0061] Mixing coefficient This determines the relative degree of reliance of the model on data-driven laws and physical prior laws. As the value approaches 1, the model becomes close to being purely data-driven. When the model approaches zero, it is strongly constrained to the physical experience curve.

[0062] Finally, substituting the aforementioned hybrid kernel function into the optimization objective function of Support Vector Regression (SVR) yields the final prediction model. Model training includes the following key steps: a) Hyperparameter optimization: Model hyperparameters include RBF kernel parameters. Physical nuclear parameters SVR penalty parameters Insensitive loss parameters and mixing coefficient Optimization is achieved using a strategy combining grid search and cross-validation. Specifically, the training dataset is further divided into training and validation subsets. The optimal hyperparameter combination is determined by traversing a predefined parameter grid and minimizing the root mean square error (RMSE) on the validation set. b) Weight coefficients Determination: Through the above grid search process, As a key hyperparameter, it was optimized together. The search range was set to 0.3 to 0.7 with a step size of 0.1. Experimental results show that the optimal λ value stabilizes in the range of 0.4-0.5, indicating that the contributions of the physical kernel and the data kernel are nearly balanced, verifying the effectiveness of the hybrid design. The final model performance was evaluated on an independent test set.

[0063] Substituting this hybrid kernel function into the support vector regression model yields the optimized objective function that incorporates physical priors. By training this model on the dataset, a wind speed prediction model that combines physical mechanisms with data learning capabilities is finally obtained.

[0064] After the model was built, its accuracy was verified using a test set, and the results are as follows: Figure 3As shown, the multi-stage fusion physical kernel support vector regression performed best on all evaluation metrics, followed by traditional support vector regression. The error evaluation results of the multi-stage fusion physical kernel support vector regression on the test set are as follows: mean squared error (MSE) of 3.78, root mean squared error (RMSE) of 1.944, mean absolute error (MAO) of 1.521, and mean relative error (MRE) of 3.01%. This indicates that the model has a good overall fit, small prediction error, and a low relative error, demonstrating strong generalization ability. Compared with the traditional support vector regression model, the multi-stage fusion physical kernel method improves on all error metrics, with a 6.21% decrease in MSE, a 3.19% decrease in RMS, a 9.39% decrease in MAO, and a 7.67% decrease in MMR, indicating that introducing a physical kernel function helps improve prediction accuracy and model robustness.

[0065] Ablation experiments and analysis:

[0066] To rigorously verify the necessity and synergistic contribution of the two core innovations in this invention—the "multi-stage modeling framework" and the "hybrid physical kernel"—we designed a systematic ablation experiment and compared the following models on the same dataset:

[0067] Model A (single-stage, data-driven model): directly maps wind speed to the noise spectrum using an RBF kernel. This model represents the traditional data-driven approach.

[0068] Model B (Two-stage, pure data kernel): This model employs the two-stage framework described in this invention, but the second-stage SVR uses only the RBF kernel (i.e., the mixing coefficient λ=0). This model is used to verify the value of the physical statistical characteristics themselves.

[0069] Model C (Two-stage, Pure Physics Kernel): Employs a two-stage framework, but the second-stage SVR uses only the aforementioned Wenz physics kernel (i.e., λ=1). This model is used to verify the value of pure physics and empirical formulas themselves.

[0070] The model of this invention (two-stage, hybrid kernel): The complete method of this invention uses a hybrid kernel and λ is determined through multiple iterations of optimization (0.4-0.5).

[0071] The performance comparison of each model on the same test set is shown in Table 4 below:

[0072] Table 4 Comparison of Experimental Results for Different SVR Models

[0073]

[0074] Experimental Analysis:

[0075] Effectiveness of physical statistical features: Model B (MSE: 3.89) shows a significant performance improvement over the baseline model A (MSE: 4.03), with an error reduction of approximately 3.5%. This demonstrates that mapping wind speed to the low-dimensional, smooth physical statistical feature of gamma distribution parameters first can effectively constrain the model learning in the second stage, improving prediction accuracy and stability.

[0076] Limitations of purely physical constraints: Model C (MSE: 4.20) performs even slightly worse than the baseline, indicating that relying solely on classical Wenz physical laws cannot fully capture the complex environment and nonlinear residuals of a specific sea area, and data-driven corrections must be introduced.

[0077] The necessity of multi-stage hybrid kernel collaboration: The complete model of this invention (MSE: 3.78) achieved the best performance, significantly outperforming Model B and Model C. This proves that the hybrid design of physical kernel and data kernel is not a simple superposition, but rather produces a synergistic effect: the physical kernel ensures the physical rationality of the prediction and the robustness of the extrapolation, while the data kernel learns the fine-grained biases in specific scenarios; neither can be dispensed with.

[0078] In summary, the ablation experiments quantitatively demonstrate that the "multi-stage framework" and "hybrid physical core" proposed in this invention are both necessary innovations that bring performance gains, and their synergistic design is the key reason for achieving significantly better results than existing technologies.

[0079] This method collects ocean noise data by underwater gliders and combines it with meteorological wind speed data to construct a dataset; it extracts multi-band noise features and uses gamma distribution fitting to obtain physical parameters; finally, it establishes a hybrid support vector regression model that integrates the Wenz physical kernel and the RBF kernel to achieve high-precision prediction of sea surface wind speed, which is superior to traditional data-driven models in terms of both error and generalization ability.

[0080] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. A method for modeling marine environmental noise based on multi-stage modeling regression of wind speed, characterized in that, Includes the following steps: Step 1: Data Acquisition and Alignment: Obtain measured marine environmental noise data and corresponding spatiotemporal point sea surface wind speed data for the target sea area, and construct a sample library of noise and wind speed pairings; Step 2: Noise spectrum feature extraction: Perform power spectral density analysis on the noise data to extract the power spectral features within the preset wind-generated noise dominant frequency band, forming a noise power spectrum dataset; Step 3: Physical statistical parameter inversion: Fit each noise spectrum in the noise power spectrum dataset to a two-parameter probability distribution model, thereby compressing the high-dimensional spectrum into two low-dimensional physical statistical parameters that represent its energy distribution pattern, forming a corresponding training set of wind speed and physical statistical parameters; Step 4: First-stage statistical parameter prediction modeling: Based on the training set of the wind speed and physical statistical parameters, establish a prediction model from wind speed to the two low-dimensional physical statistical parameters, as the first-stage model; Step 5: Two-stage hybrid kernel regression modeling: The physical statistical parameters predicted by the first-stage model and the original wind speed are used as input features, and the noise spectrum is used as the output target to construct a support vector regression model.

2. The method according to claim 1, characterized in that, In step 1, an underwater glider equipped with a high-precision hydrophone is used to continuously collect three-dimensional noise data in the target sea area. A spatiotemporally aligned dataset is constructed using a bilinear interpolation algorithm. Simultaneously, the wind speed true values ​​of the corresponding spatiotemporal points are obtained by accessing the marine meteorological service API. The raw data is stored in .dat format. After the depth, time, latitude and longitude and profile information are extracted by MATLAB, it is converted into an Excel spreadsheet for storage. The audio signal is saved in .wav format, forming an original dataset containing 138 samples.

3. The method according to claim 1, characterized in that, In step 2, during the ascent / descent phase of each profile, sampling is performed every 50m within a depth range of 50m to 800m, and a noise-wind speed dataset is constructed using 0.25 latitude and longitude as the spatial resolution of the data.

4. The method according to claim 1, characterized in that, In step 3, the two-parameter probability distribution model is a gamma distribution, and the two low-dimensional physical statistical parameters are the shape parameter and scale parameter of the gamma distribution, respectively. The fitting process specifically includes: S31: Use a Gaussian kernel function to estimate the kernel density of each noise spectrum to obtain a smooth spectral probability density curve; S32: Using the maximum likelihood estimation method, the spectral probability density curve is fitted to the gamma distribution model to extract the optimal shape and scale parameters; S33: The Kolmogorov-Smirnov test was used to evaluate the goodness of fit of the gamma distribution.

5. The method according to claim 1, characterized in that, In step 4, the first stage model is a Gaussian process regression model; the Gaussian process regression model uses a constant mean function and a radial basis function kernel, and optimizes the model hyperparameters by maximizing the marginal likelihood.

6. The method according to claim 1, characterized in that, In step 5, the support vector regression model uses a hybrid kernel function, which is a linear combination of a data-driven kernel and a physical kernel constructed based on physical empirical laws. The data-driven kernel is a radial basis function kernel. The physical kernel is constructed based on the Wenz marine environmental noise empirical curve. Its construction method is as follows: calculate the empirical noise spectrum level corresponding to the two input features according to the Wenz curve, and use a certain monotonically decreasing function value of the difference between the two noise spectrum levels as the value of the physical kernel.

7. The method according to claim 6, characterized in that, The hybrid kernel function is a weighted sum of the data-driven kernel and the physical kernel, and its weight is controlled by a hybrid coefficient that is adjustable between 0 and 1. The hybrid coefficient, together with the penalty parameter, the insensitive loss parameter, and the parameters of each kernel function of the support vector regression model, are used as hyperparameters and are jointly optimized by grid search and K-fold cross-validation.

8. The method according to claim 7, characterized in that, During the optimization process, the search range of the mixing coefficients is set to 0.3 to 0.7, and the optimization objective is to minimize the root mean square error on the validation set.

9. The method according to claim 1, characterized in that, It also includes step 6: model training and prediction: using training data to train and optimize the first-stage model and the support vector regression model to obtain a complete marine environmental noise prediction model, which is used to predict the noise spectrum based on the input wind speed.

10. The method according to claim 9, characterized in that, Step 6 is followed by a verification step: evaluating the performance of the marine environmental noise prediction model using an independent test set, with evaluation metrics including mean square error, root mean square error, mean absolute error, and mean relative error.

Images