Wind noise modeling method and system based on machine learning and application

Abstract

The invention relates to a wind noise modeling method and system based on machine learning and application, and belongs to the field of marine acoustics and environmental noise modeling, and the method comprises the steps of data preprocessing and feature extraction, hybrid network model building, hybrid network model training and verification. According to the method, a hybrid network model of multiple linear regression and a multi-layer perceptron is constructed based on a known physical mechanism of wind noise and two main noise generation mechanisms of surface turbulence and bubble oscillation, a linear relation and a non-linear relation are modeled respectively, outputs of the two models are integrated through a weighted fusion strategy, and a multi-layer perceptron model is constructed. Smooth transition modeling from a low-wind-speed linear relation to a high-wind-speed nonlinear relation is achieved.

Description

Technical Field

[0001] This invention belongs to the field of marine acoustics and environmental noise modeling, specifically involving a machine learning-based method, system, and application for modeling wind noise. Background Technology

[0002] Aeolian environmental noise is an important component of marine environmental noise and has a significant impact on applications such as underwater acoustic detection, communication, and navigation. Accurately modeling the relationship between aeolian environmental noise and wind speed is of great significance for marine acoustic research and engineering applications.

[0003] Traditional modeling methods for aeolian environmental noise typically employ linear models, assuming a simple linear dependence between noise level and wind speed. However, actual observations reveal that the generation of aeolian environmental noise involves multiple physical mechanisms, primarily surface turbulence excitation and bubble oscillation excitation. Under low wind speed conditions, surface turbulence is the dominant noise generation mechanism, with a linear relationship between noise level and wind speed. However, under high wind speed conditions, the bubble oscillation mechanism gradually becomes dominant, and the relationship between noise level and wind speed becomes nonlinear. This complex nonlinearity makes it difficult for traditional linear models to accurately describe the complete characteristics of aeolian environmental noise.

[0004] While some existing studies have attempted to consider the nonlinear characteristics of aeolian environmental noise, most employ empirical formulas or simplified physical models, making it difficult to fully capture the transition characteristics from surface turbulence to bubble oscillations. Furthermore, existing methods often neglect the frequency and depth dependencies of aeolian environmental noise, leading to significant errors in practical applications.

[0005] On the other hand, the acquisition and processing of actual observation data also face many challenges. Environmental noise data collected by hydrophones usually has a high time resolution (e.g., 5-minute intervals), while wind speed observation data has a relatively low resolution (e.g., 1-hour intervals). This mismatch in data time scales further increases the difficulty of modeling.

[0006] Therefore, a new technical solution is needed that can fully consider the various physical mechanisms, frequency dependence, and depth dependence of aeolian environmental noise, and effectively handle the time scale mismatch of observation data, so as to achieve accurate modeling and prediction of aeolian environmental noise. Summary of the Invention

[0007] This invention aims to address the problems of insufficient accuracy, difficulty in handling nonlinear relationships and complex dependencies in existing wind-generated environmental noise modeling techniques. It proposes a wind-generated environmental noise modeling method based on a hybrid machine learning model. This method is based on the known physical mechanisms of wind-generated noise, using surface turbulence and bubble oscillation as the two main noise generation mechanisms as its theoretical foundation. It constructs a hybrid network model of multiple linear regression and multilayer perceptron to model linear and nonlinear relationships respectively. A weighted fusion strategy is then used to integrate the outputs of the two models, achieving a smooth transition from low-wind-speed linear relationships to high-wind-speed nonlinear relationships.

[0008] To achieve the above objectives, the present invention is implemented through the following technical solution: A method for modeling wind-induced environmental noise based on a hybrid machine learning model, the method including data preprocessing and feature extraction, building a hybrid network model, training the hybrid network model, and validation; The data preprocessing and feature extraction involved selecting noise data in the frequency range above 10kHz as modeling data. The hybrid network model is a hybrid network model of multiple linear regression and multilayer perceptron. The multiple linear regression part is responsible for modeling the linear relationship under low wind speed conditions and capturing the basic positive correlation between noise level and wind speed. The multilayer perceptron part is responsible for modeling the nonlinear relationship under high wind speed conditions and capturing the transition characteristics from surface turbulence to bubble oscillation. The training of the hybrid network model involves setting the loss function to the mean squared error between the weighted sum of the linear and nonlinear outputs and the observed values, and using the gradient descent optimization method to train the hybrid model. By adjusting the weight coefficients, the contributions of the linear and nonlinear components to the final prediction result are balanced. The verification process involves using observation data of different frequencies and depths to validate the hybrid network model.

[0009] Furthermore, the data preprocessing and feature extraction involve: performing spectral analysis on the noise data, calculating the noise level at each frequency point, and aligning it with the wind speed data within the corresponding time window to resolve the data time scale mismatch problem.

[0010] Furthermore, the multilayer perceptron portion of the constructed hybrid network model adopts a three-level network structure, including an input layer, hidden layers, and an output layer. The hidden layers use the ReLU activation function, and the output layer uses a linear activation function. The hidden layers consist of a five-layer deep feedforward network structure, with the number of neurons in each layer decreasing progressively to enhance the hierarchical representation of features at different scales. The hidden layers uniformly use the ReLU activation function to improve the model's ability to fit nonlinear patterns. ; The output layer maintains a linear mapping to ensure that the model can directly generate continuous real values ​​in the sense of physical quantities when predicting noise levels, thereby avoiding bias caused by activation function saturation or range limitation.

[0011] Furthermore, wind speeds less than 10 m / s are considered low wind speeds, and wind speeds greater than 10 m / s are considered high wind speeds.

[0012] Furthermore, the training hybrid network model introduces weight coefficients representing the relative contributions of the linear and nonlinear components, constructing them as a function of wind speed. function form The parameters are incorporated into the network parameter set as training parameters. During backpropagation, the weight coefficients are adaptively updated using the gradient descent algorithm, enabling the model to adaptively learn the mapping relationship between wind speed and weight allocation. This allows the model to automatically adjust the influence of both components on the final prediction result based on wind speed. During training, regularization techniques (L2 regularization and dropout) are used to prevent overfitting and ensure the model's generalization ability.

[0013] Furthermore, the final output of the hybrid network model For the linear part and nonlinear part Weighted sum: ; in This is a weighting coefficient that is dynamically adjusted based on wind speed. At low wind speeds... Approaching 1, at high wind speeds Approaching 0. To achieve this dynamic adjustment mechanism, [the following will be implemented]: The wind speed is constructed as a function and incorporated as a trainable submodule into the hybrid network. Specifically, an additional learning unit can be introduced into the network to calculate the wind speed. As input, it undergoes one or more linear transformations and then passes through the Sigmoid activation function. The mapping yields: ; Thus The range of values ​​is automatically limited to During backpropagation, the weight parameters Bias parameters Updated along with network parameters, this allows the network to adaptively learn the optimal weight allocation rule based on the wind speed-noise relationship in the training data: in the low wind speed range, where the linear relationship dominates, the learning units automatically learn a larger weight allocation rule. This strengthens the linear output; in the high wind speed range, the nonlinear relationship is enhanced. The nonlinear branch is automatically suppressed, giving it a higher weight. This enables dynamic modeling and smooth transition of noise excitation mechanisms under different wind speed conditions.

[0014] Furthermore, the verification process selected environmental noise level data collected from two nodes of the hydrophone section at multiple representative frequencies, and combined them with the corresponding sea surface wind speed records for verification analysis. The deployment depths of the two nodes differed to test the applicability of the model under different depth conditions. At the same time, cross-frequency band comparisons were used to verify the reliability of the model in describing frequency change trends.

[0015] The present invention also provides a system for constructing a model using the method described above, the system comprising a data input module, a data processing and feature extraction module, a hybrid network model computation module, and a result output module; The data processing and feature extraction module described above executes the data processing and feature extraction method in the wind-induced environmental noise modeling method based on a hybrid machine learning model. The hybrid network model computation module describes a hybrid network model computation method in a wind-induced environmental noise modeling method based on a hybrid machine learning model. This invention also provides applications of the system, which, based on input sea surface wind speed data and noise level calculation conditions, predicts the wind-induced environmental noise level at target frequencies and depths using a constructed hybrid network model, and further generates corresponding sound field prediction results. This system can automatically calculate the noise response at each frequency point according to different wind speed scenarios, and, combined with set depth conditions, provide an estimate of noise variation with water depth.

[0016] The beneficial effects of this invention compared to the prior art are as follows: 1. Nonlinear relationship modeling capability: This invention effectively captures the transition characteristics of wind-induced environmental noise from surface turbulence to bubble oscillation through a hybrid machine learning model. It can achieve accurate modeling under both low and high wind speed conditions, overcoming the limitations of traditional linear models.

[0017] 2. Comprehensive Consideration of Multiple Physical Mechanisms: This invention is based on the physical characteristics of aeolian environmental noise and, according to the known laws governing the two main noise generation mechanisms—surface turbulence and bubble oscillation—distinguishes the dominant role in different wind speed ranges during model design. Under low wind speed conditions, surface turbulence is the primary noise source, with a linear relationship between noise level and wind speed. Under high wind speed conditions, the bubble oscillation mechanism gradually becomes dominant, and the noise level exhibits a nonlinear relationship with wind speed. To address this transitional characteristic, a hybrid modeling strategy is designed, using both linear and nonlinear models to capture the noise characteristics in different wind speed ranges. This fully considers the multiple physical generation mechanisms of aeolian environmental noise and achieves a smooth transition between different mechanisms through reasonable model design, better reflecting actual physical processes.

[0018] 3. Frequency and depth dependence processing: The proposed hybrid model was validated on observation data at different frequencies and depths. The results show that the model can accurately reflect the variable characteristics of wind-induced environmental noise under frequency and depth conditions.

[0019] 4. Effective solution to the problem of data time scale mismatch: This invention effectively addresses the time scale mismatch between hydrophone noise data and wind speed data through reasonable data preprocessing and time alignment strategies, thereby improving the accuracy of modeling.

[0020] 5. Good generalization ability and practicality: Through actual experimental verification, the present invention exhibits good predictive performance under different marine environmental conditions, providing reliable technical support for further research. Attached Figure Description

[0021] Figure 1 is an overall flowchart of the wind-induced environmental noise modeling method based on a hybrid machine learning model proposed in this invention; Figure 2 is a graph showing the relationship between noise level and wind speed collected by hydrophone 1 in an embodiment of the present invention, illustrating the linear relationship at low wind speeds and the nonlinear relationship at high wind speeds. Figure 3 is a comparison of the wind noise network prediction results and actual data of hydrophone node 1 at frequencies of 10kHz and 15kHz in the embodiment of the present invention; (a): comparison results at a frequency of 10kHz; (b): comparison results at a frequency of 15kHz. Figure 4 shows a comparison between the wind-induced noise network prediction results and actual data of hydrophone node 2 at frequencies of 10kHz and 15kHz in this embodiment of the invention. (a): Comparison results at 10kHz, (b): Comparison results at 15kHz. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0023] The flowchart of the wind-generated environmental noise modeling method based on a hybrid machine learning model proposed in this patent is shown below. First, data preprocessing and feature extraction are performed to solve the problem of data time scale mismatch; then, the physical mechanism of wind-generated environmental noise is analyzed and a hybrid modeling strategy is designed; next, a hybrid network model of multiple linear regression and multilayer perceptron is constructed; through model training and optimization, a smooth transition from linear to nonlinear relationships is achieved; finally, frequency dependence and depth dependence are considered to establish a complete modeling system.

[0024] The technical solution of the present invention will be described below with reference to a specific embodiment.

[0025] 1. For the environmental noise data and corresponding sea surface wind speed data recorded by the hydrophone, the data first undergoes quality control and screening. Considering the potential interference from non-wind speed factors in the low-frequency range (e.g., below 10kHz), environmental noise data in the frequency range above 10kHz is selected as the modeling data. The main influencing factor of environmental noise in this frequency range is sea surface wind speed. Spectral analysis is performed on the noise data to calculate the noise level at each frequency point, and time alignment is performed with the wind speed data within the corresponding time window to resolve the data time scale mismatch issue.

[0026] The data in this embodiment comes from actual ocean observations. Environmental noise data and corresponding wind speed data recorded by hydrophone node 1 and hydrophone node 2 during winter were selected for a modeling experiment. The two hydrophones were deployed at different depths. The hydrophone sampling frequency was 50kHz, and the noise level calculation time interval was 5 minutes; the time resolution of the wind speed data was 1 hour. The experimental frequency range was selected from 10kHz to 20kHz to avoid interference from low-frequency non-wind speed factors.

[0027] First, quality control was performed on the collected hydrophone environmental noise data to remove outliers and interference signals. Then, the noise time series was analyzed. Perform a Fast Fourier Transform (FFT) to obtain the frequency domain representation. : ; in, For time sampling point index, ; Frequency index; Indicates the number of FFT sampling points. This represents the complex exponential basis function.

[0028] Based on this, calculate each frequency point. sound pressure power spectral density : ; in The number of sampling points. Frequency resolution. Frequency. Noise level at Convert using the following formula: ; in The reference sound pressure level is used. Considering the significant non-wind speed interference in the frequency range below 10kHz, frequency data above 10kHz are selected for modeling.

[0029] To address the time scale mismatch between hydrophone noise data (5-minute intervals) and wind speed data (1-hour intervals), this study employs a time scale alignment strategy based on fixed time windows. The time axis is divided into continuous, non-overlapping windows in hours. For the first... Each window records the wind speed as follows: This can be considered as a representative value of wind speed within that time window. , recorded as Within the corresponding time window, the hydrophone provides 12 noise level samples: in Indicates the first Within an hour, the The noise level corresponding to each time interval Each time interval is 5 minutes.

[0030] To achieve scale consistency, noise samples within the same window are paired with uniform hourly wind speed values ​​to form aligned sample pairs: , Indicates the first Within the first time window The matched samples include noise levels and corresponding wind speeds. This method unifies noise and wind speed data in the time dimension, thereby avoiding modeling biases caused by differences in temporal resolution and improving the stability of hybrid network training.

[0031] Step 2: Physical Mechanism Analysis and Modeling Strategy Design Based on the physical characteristics of wind-induced environmental noise, the generation mechanisms of two main noise sources—surface turbulence and bubble oscillation—are analyzed. For example... As shown, at low wind speeds (typically less than 10 m / s), there is a linear relationship between noise level and wind speed, mainly dominated by surface turbulence mechanism; at high wind speeds (typically greater than 10 m / s), there is a higher power-law relationship between noise level and wind speed, mainly dominated by bubble oscillation mechanism.

[0032] To address the transitional characteristics of this mechanism, a hybrid modeling strategy is designed: a multivariate linear regression model is used to capture the linear relationship under low wind speed conditions, a multilayer perceptron model is used to capture the nonlinear relationship under high wind speed conditions, and a smooth transition is achieved through dynamic weighted fusion.

[0033] Step 3: Building a Hybrid Machine Learning Model A hybrid network model combining multiple linear regression and a multilayer perceptron was constructed. The multiple linear regression part adopted the standard linear regression form, with wind speed data as input and linearly predicted noise level as output. The multilayer perceptron part employed a three-layer network structure, including an input layer, hidden layers, and an output layer. The hidden layers consisted of a five-layer deep feedforward network structure, with the number of neurons in each layer decreasing progressively to enhance the hierarchical representation of features at different scales. The hidden layers uniformly used the ReLU activation function to improve the model's ability to fit nonlinear patterns. ; The output layer maintains a linear mapping to ensure that the model can directly generate continuous real values ​​in the sense of physical quantities when predicting noise levels, thereby avoiding bias caused by activation function saturation or range limitation.

[0034] Specifically, the final output of the hybrid network model For the linear part and nonlinear part Weighted sum: ; in This is a weighting coefficient that is dynamically adjusted based on wind speed. At low wind speeds... Approaching 1, at high wind speeds Approaching 0. To achieve this dynamic adjustment mechanism, [the following will be implemented]: It is constructed as a function of wind speed and incorporated into the hybrid network as a trainable submodule.

[0035] Specifically, an additional learning unit can be introduced into the network to measure wind speed. As input, it undergoes one or more linear transformations and then passes through the Sigmoid activation function. The mapping yields: ; Thus The range of values ​​is automatically limited to During backpropagation, the weight parameters Bias parameters Updated along with network parameters, this allows the network to adaptively learn the optimal weight allocation rule based on the wind speed-noise relationship in the training data: in the low wind speed range, where the linear relationship dominates, the learning units automatically learn a larger weight allocation rule. This strengthens the linear output; in the high wind speed range, the nonlinear relationship is enhanced. The nonlinear branch is automatically suppressed, giving it a higher weight. This enables dynamic modeling and smooth transition of noise excitation mechanisms under different wind speed conditions.

[0036] Step 4: Train the hybrid model The mixture model was trained using gradient descent optimization, with the loss function set to the mean squared error between predicted and observed values. L2 regularization was employed to prevent overfitting during training, with a learning rate of 0.001 and 1000 training epochs. Optimal weight coefficients were determined through cross-validation. The functional relationship between wind speed and the model ensures a smooth transition from linear to nonlinear relationships. Weighting coefficients representing the relative contributions of the linear and nonlinear components are introduced, constructing a function that varies with wind speed. function form The parameters are incorporated into the network parameter set as training parameters. During backpropagation, the weight coefficients are adaptively updated using the gradient descent algorithm, enabling the model to adaptively learn the mapping relationship between wind speed and weight allocation. This allows the model to automatically adjust the influence of both components on the final prediction result based on wind speed. During training, regularization techniques (L2 regularization and dropout) are used to prevent overfitting and ensure the model's generalization ability.

[0037] Step 5: Verification To analyze the characteristics of wind-generated environmental noise at different frequencies, the performance of the hybrid model was examined using data from the same hydrophone node (node ​​1). For example... As shown, the results indicate that the proposed hybrid network has higher prediction accuracy at both 10kHz and 15kHz frequencies compared to the traditional linear regression method. Further comparison reveals that, under the same wind speed conditions, the noise level at 10kHz is higher than that at 15kHz, indicating that wind-generated noise level gradually decreases with increasing frequency.

[0038] To further evaluate the impact of observation depth on model adaptability, hydrophone node 2 (with a depth difference of approximately 30 meters from hydrophone 1) was selected for comparative verification. The results are as follows: Figure 4 As shown in the figure. The results indicate that the hybrid model can effectively model and predict wind-generated environmental noise at different observation depths, and the model output is generally consistent with the measured data. Further comparison of the wind-generated noise levels at two nodes under the same frequency and wind speed revealed a difference of 0.5-2 dB, indicating that the wind-generated environmental noise level is dependent on depth.

[0039] The predictive performance of the model was quantitatively evaluated using observation data from the same hydrophone nodes. Experimental results show that the proposed hybrid model can achieve accurate predictions under both low and high wind speed conditions, with a prediction error lower than that of the traditional linear model. Compared with traditional methods, the hybrid model can accurately capture the nonlinear relationship between noise level and wind speed, improving prediction accuracy by approximately 40%.

[0040] Table 1 shows a comparison of the prediction errors of the hybrid model and the traditional linear model at different frequencies. The results show that the prediction error of the hybrid model is smaller than that of the linear model at all frequencies, demonstrating better prediction performance.

[0041] Table 1. Comparison of Prediction Errors between Hybrid Models and Traditional Linear Models ; The above embodiments demonstrate that the wind-generated environmental noise modeling method based on a hybrid machine learning model proposed in this invention can effectively overcome the limitations of traditional linear models and achieve accurate modeling and prediction of wind-generated environmental noise.

Claims

1. A method for modeling wind-induced environmental noise based on a hybrid machine learning model, characterized in that, The method includes data preprocessing and feature extraction, building a hybrid network model, training the hybrid network model, and validation. The data preprocessing and feature extraction involved selecting noise data in the frequency range above 10kHz as modeling data. The hybrid network model is a hybrid network model of multiple linear regression and multilayer perceptron. The multiple linear regression part is responsible for modeling the linear relationship under low wind speed conditions and capturing the basic positive correlation between noise level and wind speed. The multilayer perceptron part is responsible for modeling the nonlinear relationship under high wind speed conditions and capturing the transition characteristics from surface turbulence to bubble oscillation. The training of the hybrid network model involves setting the loss function to the mean squared error between the weighted sum of the linear and nonlinear outputs and the observed values, and using the gradient descent optimization method to train the hybrid model. By adjusting the weight coefficients, the contributions of the linear and nonlinear components to the final prediction result are balanced. The verification process involves using observation data of different frequencies and depths to validate the hybrid network model.

2. The wind-induced environmental noise modeling method based on a hybrid machine learning model according to claim 1, characterized in that, The data preprocessing and feature extraction involve performing spectral analysis on the noise data, calculating the noise level at each frequency point, and aligning it with the wind speed data within the corresponding time window.

3. The wind-induced environmental noise modeling method based on a hybrid machine learning model according to claim 1, characterized in that, The multilayer perceptron in the constructed hybrid network model adopts a three-level network structure, including an input layer, hidden layers, and an output layer. The hidden layers use the ReLU activation function, and the output layer uses a linear activation function. The hidden layers consist of a five-layer deep feedforward network structure, with the number of neurons in each layer decreasing progressively to enhance the hierarchical representation of features at different scales. The hidden layers uniformly use the ReLU activation function to improve the model's ability to fit nonlinear patterns. ； The output layer maintains a linear mapping to ensure that the model can directly generate continuous real values ​​in the sense of physical quantities when predicting noise levels, thereby avoiding bias caused by activation function saturation or range limitation.

4. The wind-induced environmental noise modeling method based on a hybrid machine learning model according to claim 1, characterized in that, Wind speeds less than 10 m / s are considered low wind speeds, and wind speeds greater than 10 m / s are considered high wind speeds.

5. The wind-induced environmental noise modeling method based on a hybrid machine learning model according to claim 1, characterized in that, The training hybrid network model introduces weight coefficients representing the relative contributions of the linear and nonlinear components, constructing them as a function of wind speed. function form The parameters are incorporated into the network parameter set as training parameters. During backpropagation, the weight coefficients are adaptively updated using the gradient descent algorithm, enabling the network to adaptively learn the mapping relationship between wind speed and weight allocation. Regularization techniques are used to prevent overfitting.

6. The wind-induced environmental noise modeling method based on a hybrid machine learning model according to claim 5, characterized in that, The final output of the hybrid network model For the linear part and nonlinear part Weighted sum: ； in This is a weighting coefficient that is dynamically adjusted based on wind speed. At low wind speeds... Approaching 1, at high wind speeds Approaching 0, its for ； The range of values ​​is automatically limited to During backpropagation, the weight parameters Bias parameters Updated along with network parameters, where v represents wind speed.

7. The wind-induced environmental noise modeling method based on a hybrid machine learning model according to claim 1, characterized in that, The verification process involved selecting environmental noise level data collected from two nodes of the hydrophone section at multiple representative frequencies, and combining this data with corresponding sea surface wind speed records for verification analysis. The deployment depths of the two nodes differed to test the applicability of the model under different depth conditions. Simultaneously, cross-frequency band comparisons were used to verify the reliability of the model in describing frequency change trends.

8. A system for operating the method according to any one of claims 1-7, characterized in that, The system includes a data input module, a data processing and feature extraction module, a hybrid network model operation module, and a result output module. The data processing and feature extraction module described herein operates the data processing and feature extraction method in the wind-induced environmental noise modeling method based on a hybrid machine learning model as described in any one of claims 1-7. The hybrid network model computation module performs the hybrid network model computation method in the wind-induced environmental noise modeling method based on a hybrid machine learning model as described in any one of claims 1-7; 9. The application of the system according to claim 8, characterized in that, Based on the input sea surface wind speed data and noise level calculation conditions, a hybrid network model is used to predict the wind-induced environmental noise level at the target frequency and target depth, and further generate the corresponding sound field prediction results.

Images