A method and apparatus for constructing an acoustic prediction model for steel ball entering water.

Abstract

A method and apparatus for constructing an acoustic prediction model for steel ball entering water are disclosed. To address the low prediction accuracy problem in existing acoustic prediction models for steel ball entry into water, this invention proposes a method for constructing such a model. First, the time-domain signal of the steel ball entering the water is acquired using the near-field virtual source method and the reverberation method. After EMD denoising, the sound power level is calculated and the power spectrum is plotted. Then, filtering and fusion are performed to obtain the total radiated sound power of the steel ball entering the water. Simultaneously, the initial impact sound power level is extracted, and the bubble pulsation sound power is obtained through difference. Multiple operating condition experiments are repeated to plot the variation of sound power level with particle size and water drop height. A three-layer perceptron neural network model is then constructed and trained using a mean squared error loss function, Dropout strategy, and 5-fold cross-validation. Finally, a model capable of accurately predicting the acoustic characteristics of steel ball entering water is obtained. This method is applicable to underwater acoustics research and other technical fields.

Description

Technical Field

[0001] This invention belongs to the field of underwater acoustic signal processing and prediction modeling technology, specifically involving a method for predicting the acoustic performance of a steel ball entering water based on measuring the pulsating radiated acoustic power level of air bubbles entering the water and combining it with a neural network. Background Technology

[0002] With the development of underwater acoustic technology, the steel ball immersion experiment has been widely used as an important means of studying the acoustic properties of water bodies. By measuring and analyzing the impact sound and bubble pulsation sound during the steel ball immersion process, rich acoustic information can be obtained, providing reference for underwater structure inspection, environmental monitoring, and marine engineering. However, existing experimental analysis methods mainly rely on traditional signal processing techniques, which suffer from problems such as sensitivity to noise, low data processing efficiency, and difficulty in predicting analysis results, failing to meet the need for rapid evaluation of acoustic characteristics under multiple operating conditions and parameters.

[0003] In recent years, with the development of computing power and machine learning technology, methods such as neural networks have shown promising application prospects in acoustic signal modeling and prediction. Existing studies have attempted to combine experimental data with machine learning models to predict acoustic characteristics, but most of them focus on single signal processing methods or are limited to a small number of experimental conditions. The generalization ability and prediction accuracy of the models are difficult to guarantee, especially in processing the entire time-domain signal of water entry and distinguishing between initial impact sound and bubble pulsation sound.

[0004] Existing technologies also face challenges in experimental data acquisition and processing. The near-field virtual source method and the reverberation method each have their advantages and disadvantages in practical measurements. A single method cannot fully reflect the acoustic characteristics generated by the steel ball entering the water. Furthermore, noise interference significantly affects the power spectrum analysis results, limiting the accuracy of subsequent acoustic modeling and prediction. In addition, the variation patterns of acoustic characteristics caused by different steel ball sizes and water drop heights have not been systematically summarized, and reliable prediction models are lacking. This limits the use of experimental analysis results to retrospective studies, making it difficult to effectively predict unknown operating conditions.

[0005] Therefore, how to establish an acoustic model that can accurately predict the sound power level of a steel ball entering water, by combining signal optimization processing and neural network modeling techniques under existing experimental conditions, has become an urgent technical problem to be solved. Summary of the Invention

[0006] To address the issue of low prediction accuracy in existing underwater acoustic prediction models, this invention proposes the following solution: A method for constructing an acoustic prediction model for a steel ball entering water, the method comprising: S1. The time-domain signal 1 of the steel ball entering the water is acquired by the near-field virtual source method, and the time-domain signal 2 of the steel ball entering the water is acquired by the reverberation method. S2. Perform EMD preprocessing on both time-domain signals described in S1 to obtain denoised time-domain signal 1 and time-domain signal 2. S3. Calculate the sound power level of the denoised time-domain signal 1 and plot the first power spectrum; S4. Calculate the sound power level of the denoised time-domain signal 2 and plot the second power spectrum; S5. Filter and fuse the first power spectrum and the second power spectrum, and use integral calculation to obtain the radiated sound power level of the whole segment when the steel ball enters the water. S6. Extract the initial impact sound of the noise-reduced time-domain signal 1 and calculate the sound power level to obtain the radiated sound power level of the initial impact sound when the steel ball enters the water. S7. Subtract the radiated sound power level of the initial impact sound from the radiated sound power level of the entire segment to obtain the radiated sound power level of the bubble pulsation sound.

[0007] S8. Repeat steps S1-S7 to obtain the radiated sound power levels of the whole section, initial impact sound and bubble pulsation sound corresponding to different working conditions. S9. Plot the curves of the radiated sound power level of the whole section, the initial impact sound, and the bubble pulsation sound under different working conditions as a function of the steel ball particle size and the release height to obtain the variation law. S10. Based on the variation law, a neural network prediction model is constructed with steel ball diameter and release height as independent variables and the radiated sound power level of initial impact sound and bubble pulsation sound as dependent variables. S11. The neural network prediction model is trained using the mean squared error loss function. During the training process, the Dropout strategy is introduced, and the neural network prediction model is trained using 5-fold cross-validation. S12. Obtain the acoustic prediction model for the steel ball entering the water.

[0008] Furthermore, the neural network prediction model described in S10 is a three-layer perceptron model, including an input layer, a hidden layer, and an output layer; the input layer contains two input features, namely the steel ball diameter and the height of its fall into the water; the hidden layer includes a two-level structure, with the first hidden layer having 32 neurons and the second hidden layer having 16 neurons, both employing... Activation function; the output layer includes two output variables, namely the initial impact sound power level and the bubble pulsation sound power level.

[0009] Furthermore, the neural network prediction model can be expressed as: , in, For input variables, For output variables, This represents the input variable of the i-th multilayer perceptron. Its function.

[0010] Furthermore, the aforementioned The activation function is a rectified linear unit function, defined as: .

[0011] Furthermore, the method for training the neural network prediction model described in S11 is as follows: First, the mean squared error loss function is used to calculate the loss value for the model: , in, For loss function, The radiated sound power level predicted by the model. For binary variables, This represents the number of data sample groups; Secondly, various nonlinear fits to the model are compared, and the fitting effect is evaluated using the following three metrics: , , , in, It is the coefficient of determination. It is the root mean square error. It is the mean absolute error. It is the sum of squared residuals. It is the total sum of squares. It is the number of observations. This is the nth actual observation value. It is the first Each model predicts a value.

[0012] Furthermore, the expression for the Dropout strategy described in S12 is: , in, Represents element-wise multiplication. and It is a random 0-1 vector.

[0013] Furthermore, the 5-fold cross-validation method described in S12 is as follows: The dataset was divided into 5 subsets. One subset was selected as the test set and the rest were used as the training set. This process was repeated 5 times. The average of the 5 test results was then used as the performance evaluation result of the neural network prediction model.

[0014] Furthermore, the time-domain signal 1 acquired by the near-field virtual source method in S1 is obtained by placing two hydrophones underwater at a distance of 0.5 meters from the point where the steel ball enters the water; the time-domain signal 2 acquired by the reverberation method is obtained by deploying a hydrophone array at multiple points and using spatial averaging and far-field correction.

[0015] Furthermore, the EMD preprocessing method described in S2 is as follows: each time-domain signal is decomposed into an intrinsic mode function, and the intrinsic mode function is reconstructed to obtain a noise-removed near-field virtual source time-domain signal and a reverberation time-domain signal, namely, noise-removed time-domain signal 1 and time-domain signal 2.

[0016] Furthermore, the number of extreme points and the number of zero-crossing points of the intrinsic mode function are equal or differ by 1 throughout the entire data interval, and the average value of the local maximum envelope and the local minimum envelope at any point is 0.

[0017] Furthermore, the method for calculating the sound power level of the denoised time-domain signal 1 described in S3 is as follows: First, the noise-removed near-field virtual source time-domain signal is substituted into the sound pressure of the dipole near-field virtual source. The expression for particle velocity: , , , in, It is the radial particle velocity. It is the vertical particle vibration velocity. For dipole moments, k For wave number, Angular frequency, It is the imaginary unit, satisfying ; Secondly, the far-field radial sound intensity is obtained based on the sound pressure and particle velocity. and vertical radial sound intensity : , , in, For the length of time, The speed of sound in water. and These are the angular frequency components in the far-field radial and vertical radial directions, respectively; Then obtain the total sound intensity: , Further obtain the radiated acoustic power level of the denoised time-domain signal 1. and radiated sound power level : , , in, It is the overall sound strength. It means to integral, It is the radiated sound power level of the sound source. It is the reference sound power level. ; Finally, based on the radiated sound power level Plot the first power spectrum.

[0018] Furthermore, the method for calculating the sound power level of the denoised time-domain signal 2 described in S4 is as follows: First, using normal mode theory, we assume the distance between the measuring point and the sound source... The radius vector in the water tank under the action of a directional sound source is obtained as Mean square sound pressure at spatial points : , in, Represents the function of direct sound. Represents reverberation sound pressure level. The intensity of a directional sound source, The reverberation radius of a point source equivalent to a directional sound source; Secondly, regarding the mean square sound pressure and radiated sound power level Request for sound level: , , in, It is the pool constant. The average sound absorption coefficient of the pool wall. The area of ​​the pool wall. It is the density of water. It is the speed of sound in water, in the formula The spatial average sound pressure level of the reverberant time-domain signal. The radiated acoustic power level of the reverberant time-domain signal; Then, the spatially averaged sound pressure level of the reverberant time-domain signal. and free field sound source level There is a correction amount. : , , Further obtain the radiated acoustic power level of the denoised time-domain signal 2. : , Ultimately based on the radiated sound power level Plot the second power spectrum.

[0019] Furthermore, the method by which S5 obtains the radiated sound power level of the entire section when the steel ball enters the water is as follows: S51. Use a low-pass filter to extract the first power spectrum below the cutoff frequency. The power spectrum portion is labeled power spectrum 1; S52. Use a high-pass filter to extract the first power spectrum above the cutoff frequency. The power spectrum portion is labeled power spectrum 2; S53. Use a high-pass filter to extract the second power spectrum above the cutoff frequency. The power spectrum portion is labeled power spectrum 3; S54. Average power spectra 2 and 3 to obtain power spectrum 4; S55. Combine power spectrum 1 and power spectrum 4 to obtain the power spectrum of the entire segment when the steel ball enters the water. S56. Calculate the power spectrum of the entire section when the steel ball enters the water by integration, and obtain the radiated sound power level of the entire section when the steel ball enters the water.

[0020] Based on the same inventive concept, this invention also proposes an apparatus for constructing an acoustic prediction model of a steel ball entering water, the apparatus comprising: Acquisition module: Acquires time-domain signal 1 of steel ball entering water using near-field virtual source method, and acquires time-domain signal 2 of steel ball entering water using reverberation method; EMD module: Used to perform EMD preprocessing on both types of time-domain signals in the acquisition module to obtain noise-reduced time-domain signal 1 and time-domain signal 2; Power Spectrum Module 1: Used to calculate the sound power level of the denoised time-domain signal 1 and plot the first power spectrum; Power Spectrum Module 2: Used to calculate the sound power level of the denoised time-domain signal 2 and plot the second power spectrum; Fusion module: used to filter and fuse the first power spectrum and the second power spectrum, and use integral calculation to obtain the radiated sound power level of the whole segment when the steel ball enters the water; The initial impact sound module is used to extract the initial impact sound of the noise-reduced time-domain signal 1 and calculate the sound power level to obtain the radiated sound power level of the initial impact sound when the steel ball enters the water. The bubble pulsation module is used to obtain the radiated sound power level of the bubble pulsation sound by subtracting the radiated sound power level of the initial impact sound from the radiated sound power level of the entire segment. Multi-condition module: used to repeatedly collect data from the module to the bubble pulsation module to obtain the radiated sound power levels of the whole section, initial impact sound and bubble pulsation sound corresponding to different conditions; Dynamic Curve Module: Used to plot the radiated sound power level of the whole section, initial impact sound and bubble pulsation sound under different working conditions as a function of steel ball particle size and release height, and to obtain the variation law; Model building module: used to build a neural network prediction model based on the change pattern, with steel ball diameter and release height as independent variables and the radiated sound power level of initial impact sound and bubble pulsation sound as dependent variables; Model training module: used to train the neural network prediction model using the mean squared error loss function. Dropout strategy is introduced during the training process, and 5-fold cross-validation neural network prediction model is used. Output module: Used to obtain the acoustic prediction model of the steel ball entering the water.

[0021] Based on the same inventive concept, the present invention also proposes a computer program product, which, when read, implements the method described in the present invention.

[0022] Based on the same inventive concept, the present invention also proposes a computer storage medium for storing a computer program, wherein when the computer program is read by a computer, the computer executes the method described in the present invention.

[0023] Compared with the prior art, the present invention has the following beneficial effects: 1. The method for constructing an acoustic prediction model of a steel ball entering water, as described in this invention, combines the near-field virtual source method with the reverberation method, thus achieving a balance between the accuracy of near-field signals and the stability of far-field signals. The near-field virtual source method collects signals near the point of entry of the steel ball into the water, effectively capturing the intrinsic characteristics of the initial impact sound and bubble pulsation sound. The reverberation method, through multi-point deployment and spatial averaging, reduces the uncertainty and environmental impact of single-point measurements. Compared with existing methods that rely solely on sound pressure characterization, this scheme can more comprehensively and stably reflect the true radiation characteristics of the sound source, thereby obtaining reliable radiated sound power level results.

[0024] 2. The method for constructing an acoustic prediction model for steel ball entering water, as described in this invention, introduces EMD (Empirical Mode Decomposition) preprocessing to decompose complex time-domain signals into intrinsic mode functions, and then removes noise through reconstruction to obtain a cleaner signal input. Compared to traditional filtering methods that easily lose key information or leave residual noise, EMD can adaptively decompose non-stationary signals, preserving the nonlinear characteristics in bubble pulsation sound, making subsequent sound power level calculations more accurate and stable. This not only ensures signal quality but also improves the ability to distinguish between impact sound and bubble pulsation sound, providing a solid foundation for subsequent fusion analysis.

[0025] 3. The method for constructing an acoustic prediction model for a steel ball entering water, as described in this invention, establishes theoretical models for both the near-field virtual source method and the reverberation method in the sound power level calculation stage. The near-field virtual source method derives sound intensity and power through the near-field expression of a dipole, enabling the recovery of the radiated sound power level of the sound source from the local signal; the reverberation method combines normal mode theory with spatial averaging correction to obtain a sound power level closer to reality in complex environments. Through this complementary dual-model approach, this invention avoids the biases that are prone to occur with a single method, achieving dual verification of the acoustic characteristics of the water entry process, and is more robust and reliable in quantitative analysis compared to existing technologies.

[0026] 4. The method for constructing an acoustic prediction model of a steel ball entering water, as described in this invention, employs a frequency-separation strategy in the fusion stage. Low-pass and high-pass filters are used to separate the power spectra of different frequency bands, and then weighted averaging is used to achieve optimized fusion. The advantages of the near-field virtual source method are retained in the low-frequency portion, while reverberation correction is introduced in the high-frequency portion, thereby obtaining an overall optimized power spectrum. This method effectively solves the problem of inconsistent results from different measurement methods in existing technologies, achieving complementarity and balance across signal frequency bands, greatly improving the reliability of power calculation results, and providing more accurate data support for further analysis of bubble pulsation sound.

[0027] 5. The method for constructing an acoustic prediction model for a steel ball entering water, as described in this invention, calculates the radiated sound power level of the entire segment separately from the radiated sound power level of the initial impact sound, ultimately obtaining the radiated sound power level of the bubble pulsation sound. This processing method not only clearly distinguishes the different sound source components during the water entry process but also solves the problem of the difficulty in independently quantifying bubble pulsation sound in existing technologies. Through this method, researchers can more intuitively grasp the energy characteristics of bubble pulsation sound, providing more valuable reference data for the study of water entry acoustic mechanisms, impact point localization, and target identification.

[0028] 6. The method for constructing an acoustic prediction model for steel ball entry into water, as described in this invention, systematically summarizes the laws governing the changes in acoustic characteristics with different operating conditions by plotting experimental data under different working conditions as curves showing the radiated sound power level as a function of steel ball diameter and water drop height. Compared with existing technologies, this method overcomes the lack of multi-parameter system analysis in previous studies, enabling researchers to intuitively understand the influence of various parameters on the acoustic characteristics of the steel ball entering the water. This provides a quantifiable basis for establishing accurate prediction models and enhances the scientific rigor and application value of experimental research.

[0029] 7. The method for constructing an acoustic prediction model for steel ball entering water, as described in this invention, enables rapid prediction of the radiated sound power level of a steel ball entering water by building a three-layer perceptron neural network prediction model based on variation patterns. Compared with existing technologies, this method not only overcomes the limitations of traditional empirical formulas and single-experiment analysis, but also directly obtains the corresponding sound power level after inputting the steel ball diameter and drop height, realizing the prediction of acoustic characteristics under unknown conditions. This technique significantly improves prediction efficiency and applicability, providing a powerful tool for experimental design and engineering applications.

[0030] 8. The method for constructing an acoustic prediction model for steel ball entering water, as described in this invention, effectively prevents overfitting and improves the stability and generalization ability of the prediction model by introducing a mean squared error loss function, Dropout strategy, and 5-fold cross-validation during neural network training. Compared with existing technologies, this method maintains high-precision prediction under multi-condition data, ensuring the reliability and repeatability of the model in practical applications, achieving an efficient closed loop from experimental measurement to predictive analysis, and enhancing the practicality of acoustic research.

[0031] This invention has the ability to accurately measure, optimize signal processing, and predictively model the acoustic characteristics of a steel ball entering water. It can quickly and accurately predict the radiated sound power level of the initial impact sound and bubble pulsation sound, and is applicable to underwater acoustic research, marine engineering experiments, target signal identification, and related military and civilian underwater detection fields. Attached Figure Description

[0032] Figure 1 This is a flowchart of the method for constructing an acoustic prediction model for steel ball entering water according to the present invention; Figure 2 The result is the fast Fourier transform (FFT) of the sound pressure signal obtained from near-field measurements in the implementation method. Figure 3 This is a graph showing the variation of radiated sound power level with particle size for the entire segment in the implementation method. Figure 4 This is a graph showing the radiated sound power level of the initial impact sound as a function of particle size in the implementation method. Figure 5 This is a graph showing the variation of the radiated sound power level of bubble pulsation sound with particle size in the implementation method. Figure 6 This is a schematic diagram of a neural network node in the implementation method; Figure 7 This is a schematic diagram of the Dropout strategy implemented in this way; Figure 8 The image shows a 3D plot of the radiated sound power level of the bubble pulsating sound in the implementation method. Figure 9The figure shows a 3D plot of the radiated sound power level of the initial impact sound in the implementation method. Figure 10 The figure shows a 3D plot of the radiated sound power level for the entire segment in the implementation method. Detailed Implementation

[0033] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0034] Implementation Method 1 like Figure 1 As shown, a method for constructing an acoustic prediction model for a steel ball entering water is provided, the method comprising: S1. The time-domain signal 1 of the steel ball entering the water is acquired by the near-field virtual source method, and the time-domain signal 2 of the steel ball entering the water is acquired by the reverberation method. S2. Perform EMD preprocessing on both time-domain signals described in S1 to obtain denoised time-domain signal 1 and time-domain signal 2. S3. Calculate the sound power level of the denoised time-domain signal 1 and plot the first power spectrum; S4. Calculate the sound power level of the denoised time-domain signal 2 and plot the second power spectrum; S5. Filter and fuse the first power spectrum and the second power spectrum, and use integral calculation to obtain the radiated sound power level of the whole segment when the steel ball enters the water. S6. Extract the initial impact sound of the noise-reduced time-domain signal 1 and calculate the sound power level to obtain the radiated sound power level of the initial impact sound when the steel ball enters the water. S7. Subtract the radiated sound power level of the initial impact sound from the radiated sound power level of the entire segment to obtain the radiated sound power level of the bubble pulsation sound.

[0035] S8. Repeat steps S1-S7 to obtain the radiated sound power levels of the whole section, initial impact sound and bubble pulsation sound corresponding to different working conditions. S9. Plot the curves of the radiated sound power level of the whole section, the initial impact sound, and the bubble pulsation sound under different working conditions as a function of the steel ball particle size and the release height to obtain the variation law. S10. Based on the variation law, a neural network prediction model is constructed with steel ball diameter and release height as independent variables and the radiated sound power level of initial impact sound and bubble pulsation sound as dependent variables. S11. The neural network prediction model is trained using the mean squared error loss function. During the training process, the Dropout strategy is introduced, and the neural network prediction model is trained using 5-fold cross-validation. S12. Obtain the acoustic prediction model for the steel ball entering the water.

[0036] This implementation method acquires near-field and far-field acoustic signals of steel balls entering water, performs EMD noise reduction, power spectrum calculation, filtering and fusion, and separates the overall, initial impact sound and bubble pulsation sound. It can comprehensively and accurately obtain different acoustic characteristics of steel balls entering water and establish the sound power level variation law under different working conditions, providing a real and reliable data foundation for subsequent prediction model construction.

[0037] Furthermore, the neural network prediction model described in S10 is a three-layer perceptron model, including an input layer, a hidden layer, and an output layer; the input layer contains two input features, namely the steel ball diameter and the height of its fall into the water; the hidden layer includes a two-level structure, with the first hidden layer having 32 neurons and the second hidden layer having 16 neurons, both employing... Activation function; the output layer includes two output variables, namely the initial impact sound power level and the bubble pulsation sound power level.

[0038] This implementation uses a three-layer perceptron model, with two features set in the input layer: steel ball diameter and falling height. In the hidden layer, 32 and 16 neurons are configured, and the ReLU activation function is used. This effectively extracts the complex nonlinear relationship between the input features and the output acoustic parameters, thereby improving the fitting and generalization capabilities of the prediction model.

[0039] Furthermore, the neural network prediction model can be expressed as: , in, For input variables, For output variables, This represents the input variable of the i-th multilayer perceptron. The function, , among which, the The weight matrix of the layer is The bias vector is .

[0040] This implementation method, by clearly defining the mathematical expression of the neural network prediction model, can intuitively represent the process by which input variables are mapped step by step through a multilayer perceptron and output results. This is beneficial to improving the interpretability and reproducibility of the model, while ensuring the rationality of the model design.

[0041] Furthermore, the aforementioned The activation function is a rectified linear unit function, defined as: .

[0042] An activation function is a common type of activation function; it has a value greater than or equal to 0 in its domain. The value is 0 in the domain where it is less than 0. In this way, the next layer of the network can be deactivated under certain conditions, thus introducing the ability to capture nonlinearity into the model as a whole.

[0043] This implementation method adopts... As an activation function, it avoids the gradient vanishing problem of the traditional Sigmoid function, can accelerate the convergence speed of neural networks, and improve the model's ability to express nonlinear features during training, thereby improving prediction accuracy.

[0044] Furthermore, the method for training the neural network prediction model described in S11 is as follows: First, the mean squared error loss function is used to calculate the loss value for the model: , in, For loss function, The radiated sound power level predicted by the model. For binary variables, This represents the number of data sample groups; This function minimizes the squared loss so that the sample points can better fit the regression curve. The smaller the value of the mean squared error loss function, the higher the accuracy of the prediction model. Secondly, various nonlinear fits to the model are compared, and the fitting effect is evaluated using the following three metrics: , , , in, It is the coefficient of determination. The value ranges from 0 to 1, representing the proportion of variance explained by the model. A value close to 1 means that the model can explain the variation in the data well, that is, the model's fit quality is considered very good; It is the root mean square error. It is the mean absolute error; It is the sum of squared residuals, which is the sum of squared differences between the observed values ​​and the model predictions; It is the total sum of squares, which is the sum of the squares of the differences between the observed values ​​and the mean of the observed values; It is the number of observations. This is the nth actual observation value. It is the first Each model predicts a value; Used to measure the magnitude of absolute error, the measurement of error provides the same units as the actual value. The smaller the value, the better the fit.

[0045] This implementation method optimizes the model by introducing a mean squared error loss function during training, and combines it with... , and The three indicators comprehensively evaluate the model's fit, reflecting the accuracy and robustness of the prediction model from different dimensions, thus ensuring more reliable prediction results.

[0046] Furthermore, the expression for the Dropout strategy described in S12 is: , in, Represents element-wise multiplication. It is a with The proportion of 0s in a random 0-1 vector with the same shape is the Dropout proportion.

[0047] This implementation method introduces the Dropout strategy during training, which randomly deactivates some neurons during iterations. This effectively prevents model overfitting, enhances the model's generalization ability on different datasets, and improves the practical application value of the prediction model.

[0048] Furthermore, the 5-fold cross-validation method described in S12 is as follows: The dataset is divided into 5 subsets. One subset is selected sequentially as the test set, and the rest are used as the training set. This process is repeated 5 times. The average of the 5 test results is then used as the performance evaluation result of the neural network prediction model. Specifically: 1. Dataset partitioning: The entire dataset is randomly and uniformly divided into 5 equal-sized subsets, which are called "folds"; 2. Repeated Training and Testing: Perform 5 rounds of training and testing. In each round, select one fold as the test set, and combine the remaining 4 folds as the training set. Use the training set to train a model and use the test set to evaluate the trained model, recording its performance (e.g., accuracy, F1 score, etc.). 3. Results Summary: After 5 rounds, 5 performance evaluation scores will be obtained (each fold represents the result of one test set). Finally, the average of these 5 scores will be used as the final evaluation result of the model's performance.

[0049] This implementation method employs a 5-fold cross-validation approach, which rationally divides the dataset and rotates it multiple times for training and testing. This avoids accidental biases caused by a single data partitioning, improves the robustness and reliability of the model evaluation results, and thus ensures the stable performance of the prediction model.

[0050] Furthermore, the time-domain signal 1 acquired by the near-field virtual source method in S1 is obtained by placing two hydrophones underwater at a distance of 0.5 meters from the point where the steel ball enters the water; the time-domain signal 2 acquired by the reverberation method is obtained by deploying a hydrophone array at multiple points and using spatial averaging and far-field correction.

[0051] This embodiment achieves complementary acquisition of acoustic information from both the near and far fields by using dual hydrophones to acquire time-domain signal 1 in the near field and by using a hydrophone array in the far field combined with spatial averaging and far-field correction to acquire time-domain signal 2. This effectively avoids the limitations of a single measurement method and improves the reliability and accuracy of sound power level measurement.

[0052] Furthermore, the EMD preprocessing method described in S2 is as follows: each time-domain signal is decomposed into an intrinsic mode function, and the intrinsic mode function is reconstructed to obtain a noise-removed near-field virtual source time-domain signal and a reverberation time-domain signal, namely, noise-removed time-domain signal 1 and time-domain signal 2.

[0053] This implementation method uses the EMD method to decompose and reconstruct the acquired time-domain signal, achieving effective separation and suppression of noise components. This allows subsequent sound power level calculations to be based on cleaner signal data, improving the reliability of the sound power level results.

[0054] Furthermore, the number of extreme points and the number of zero-crossing points of the intrinsic mode function are equal or differ by 1 throughout the entire data interval, and the average value of the local maximum envelope and the local minimum envelope at any point is 0.

[0055] This implementation method ensures the physical rationality and stability of the EMD decomposition results by imposing mathematical constraints on the intrinsic mode function, namely, ensuring that the number of extreme points and zero-crossing points are equal or differ by 1, and that the average value of the envelope is zero, thereby further improving the effectiveness of signal denoising and reconstruction.

[0056] Furthermore, the method for calculating the sound power level of the denoised time-domain signal 1 described in S3 is as follows: First, the noise-removed near-field virtual source time-domain signal is substituted into the sound pressure of the dipole near-field virtual source. The expression for particle velocity: , , , in, It is the radial particle velocity. It is the vertical particle vibration velocity. For dipole moments, k For wave number, Angular frequency, It is the imaginary unit, satisfying ; Secondly, the far-field radial sound intensity is obtained based on the sound pressure and particle velocity. and vertical radial sound intensity : , , in, For the length of time, The speed of sound in water. and These are the angular frequency components in the far-field radial and vertical radial directions, respectively; Then obtain the total sound intensity: , Further obtain the radiated acoustic power of the denoised time-domain signal 1. and radiated sound power level : , , in, It is the overall sound strength. It means to integral, It is the radiated sound power level of the sound source. It is the reference sound power level. ; Finally, based on the radiated sound power level Plot the first power spectrum.

[0057] This embodiment calculates the sound power level of the denoised time-domain signal 1 using the near-field virtual source theory of dipoles, and establishes the relationship between sound pressure, particle velocity, sound intensity and radiated sound power level, thereby obtaining a relatively accurate radiated sound power level and its sound power level, and realizing the quantitative measurement of the sound power level of the steel ball entering the water under near-field conditions.

[0058] Furthermore, the method for calculating the sound power level of the denoised time-domain signal 2 described in S4 is as follows: First, using normal mode theory, we assume the distance between the measuring point and the sound source... The radius vector in the water tank under the action of a directional sound source is obtained as Mean square sound pressure at spatial points : , in, Represents the function of direct sound. Represents reverberation sound pressure level. The intensity of a directional sound source, The reverberation radius of a point source equivalent to a directional sound source; Secondly, regarding the mean square sound pressure and radiated sound power level Request for sound level: , , in, It is the pool constant. The average sound absorption coefficient of the pool wall. The area of ​​the pool wall. It is the density of water. It is the speed of sound in water, in the formula The spatial average sound pressure level of the reverberant time-domain signal. The radiated acoustic power level of the reverberant time-domain signal; Then, the spatial average sound pressure level of the denoised time-domain signal 2. and free field sound source level There is a correction amount. : , , Further obtain the radiated acoustic power level of the denoised time-domain signal 2. : , Ultimately based on the radiated sound power level Plot the second power spectrum.

[0059] This implementation calculates the sound power level of the denoised time-domain signal 2 based on normal mode theory and reverberation method. It can comprehensively consider the effects of direct sound and reverberation, and introduces spatial averaging and correction compensation to ensure that the radiated sound power level obtained under water pool conditions is closer to the true value of the free field, thus improving the accuracy of the measurement.

[0060] Furthermore, the method by which S5 obtains the radiated sound power level of the entire section when the steel ball enters the water is as follows: S51. Use a low-pass filter to extract the first power spectrum below the cutoff frequency. The power spectrum portion is labeled power spectrum 1; S52. Use a high-pass filter to extract the first power spectrum above the cutoff frequency. The power spectrum portion is labeled power spectrum 2; S53. Use a high-pass filter to extract the second power spectrum above the cutoff frequency. The power spectrum portion is labeled power spectrum 3; S54. Average power spectra 2 and 3 to obtain power spectrum 4; S55. Combine power spectrum 1 and power spectrum 4 to obtain the power spectrum of the entire segment when the steel ball enters the water. S56. Calculate the power spectrum of the entire section when the steel ball enters the water by integration, and obtain the radiated sound power level of the entire section when the steel ball enters the water.

[0061] This embodiment achieves reasonable separation and complementarity of acoustic characteristics in different frequency bands by performing low-pass and high-pass filtering on the first and second power spectra respectively, and then by fusion and integration operations, thereby obtaining a more accurate result of the total radiated sound power level when the steel ball enters the water, effectively avoiding the deviation caused by a single frequency band or a single method.

[0062] Implementation Method 2 This embodiment combines the technical solutions described in the foregoing embodiments, and, in conjunction with actual application scenarios and the usage process of computer program products that implement the method of this invention, further verifies and explains the technical effects of this invention through specific embodiments.

[0063] Step 1: Based on a method for measuring the radiated sound power level of a steel ball entering water and pulsating air bubbles, and combined with laboratory measurement conditions, arrange appropriate measurement conditions, and measure and calculate the radiated sound power levels of the whole section, initial impact sound, and pulsating air bubbles corresponding to different conditions. 1.1 Setting up the experimental measurement plan: Based on the scale effect in fluid dynamics, the size of the sphere has minimal impact on the surrounding fluid. With the sphere's entry velocity and experimental conditions constant, the sphere diameter was varied, and the changes in the initial impact sound and bubble pulsation acoustic characteristics with the sphere's size were observed and calculated. With the steel ball's diameter constant, its entry velocity was varied, and the changes in the initial impact sound and bubble pulsation acoustic characteristics with the drop height were observed and calculated. The sphere's entry velocity can be determined using Newton's second law.

[0064] The operating conditions for this embodiment are shown in Table 1: Table 1 Operating Conditions Table

[0065] 1.2 Pre-processing EMD method: The background noise in the experiment is not constant; it depends on the noise level of the measurement environment. Sometimes the background noise difference between each measurement group is small, but sometimes it can be significantly different when the surroundings are noisy compared to when they are quiet. Measuring and calculating the background noise for each set of experiments would be very tedious, so preprocessing the signal before calculation is more convenient.

[0066] like Figure 2 The figure shows the results of performing a Fast Fourier Transform (FFT) on the acoustic signal obtained from near-field measurements in this embodiment. For clarity, it is displayed on logarithmic scales, and the figure contains three types of data: FFT - Raw Signal, representing the spectral distribution of the unprocessed raw near-field measurement acoustic signal; After FFT-EMD decomposition, the signal reconstructed by processing the original signal through Empirical Mode Decomposition (EMD) has a spectrum that is highly consistent with the original signal, indicating that the EMD method maintains the main frequency characteristics during the denoising or decomposition process. FFT-Background represents the spectral distribution of the background noise signal, with an overall amplitude significantly lower than that of the original signal and the signal after EMD decomposition.

[0067] As can be seen from the figure, in the range of 200Hz–1400Hz, the main spectrum distribution of the original signal and the signal after EMD decomposition basically overlaps, indicating that the EMD method can effectively preserve the signal characteristics; while the background noise amplitude is low and mainly concentrated in the low frequency band, making the energy characteristics of the target signal more prominent in comparison.

[0068] Since EMD decomposition and reconstruction can retain the useful signal characteristics IMF components of non-steady-state signals and remove noise, the signal can be preprocessed before calculating the radiated sound power level. This allows background interference to be eliminated as much as possible while retaining the basic important information of the signal.

[0069] The accuracy of EMD decomposition and signal reconstruction was verified by randomly selecting near-field virtual source method measurement data from eight existing operating conditions. The first set of data was used to calculate the sound power level after removing the background from the original signal. The second set of data was used to calculate the sound power level after EMD decomposition and reconstruction of the original signal. The two sets of data were compared, and Table 2 presents the results.

[0070] Table 2 Applicability verification data for EMD decomposition

[0071] As can be seen from the table, the radiated acoustic power level obtained by processing the data using EMD decomposition and reconstruction is almost identical to the radiated acoustic power level of the original signal after removing background noise. The difference between the two is less than 1 dB. Therefore, EMD decomposition and reconstruction is very suitable for signal preprocessing obtained by both measurement methods.

[0072] Step 2: Plot the radiated sound power level curves of the entire segment, the initial impact sound, and the bubble pulsation sound as a function of particle size and drop height, and analyze the variation law.

[0073] 2.1 Analysis of the overall trend of the results like Figure 3 As shown, these data were plotted with particle size on the horizontal axis and radiated sound power level on the vertical axis, at drop heights of 2m, 4m, 8m, and 16m. The four colors in the graphs correspond to the variation of radiated sound power level with particle size across the entire near-field measurement at drop heights of 2m, 4m, 8m, and 16m. It can be seen that although the radiated sound power level increases with increasing particle size, the specific trend is not clearly predictable. Furthermore, the broken lines formed by these data intersect at certain particle size values, indicating that at certain particle size values, there is no clear correlation between radiated sound power level and the release height of the ball.

[0074] 2.2 Analysis of initial impact sound and bubble pulsation sound The acoustic signal of an object entering water mainly consists of two parts: bubble pulsation sound and initial impact sound. For the initial impact sound, the time-domain waveform of the near-field measurement signal shows that the duration of the impact sound is very short. Furthermore, the time-domain waveform of the reverberation method measurement signal clearly shows that a portion of the reflected sound signal directly overlaps with the bubble pulsation sound. Therefore, the impact sound signal should be directly extracted from the time-domain waveform of the near-field measurement signal. For the bubble pulsation sound, this embodiment shows that its duration is greater than the interval between the signal echo and the original signal. Therefore, the sound power level of the bubble pulsation sound is obtained by subtracting the previously measured impact sound power level from the total sound power level.

[0075] For initial impact sound, due to its distinct signal characteristics and short duration, it is easier to extract than bubble pulsation sound. Therefore, as long as there is no external noise interference, it can be easily separated from the whole segment. These signals can then be individually calculated for radiated sound power level and reduced to sound level, and curves showing their variation with particle size can be plotted. Figure 4 As shown, the relationship between the radiated sound power level of the initial impact sound and the particle size is relatively clear, exhibiting an approximately logarithmic trend. However, the relationship with the release height is not obvious.

[0076] For bubble pulsating sound, due to the difficulty of separation, the variation curve is obtained by subtracting the initial impact sound radiated sound power level from the total radiated sound power level, as shown in the figure. Figure 5 As shown in the figure, in this embodiment, the overall pulsating sound shows an increasing trend with the change in particle size. However, according to the intersecting parts in the broken line graph, it can also be seen that the change in radiated sound power level with release height is uncertain.

[0077] In summary, the patterns of change in these curves are difficult to derive using simple linear and nonlinear fitting methods. Therefore, if we want to achieve the predictive function based on the existing data, we still need to establish a predictive model that combines the two variables, particle size and release point height, as independent variables.

[0078] Step 3: Based on the changing patterns in Step 2, a preliminary prediction model based on a neural network was designed and established. The effectiveness of the model in this implementation method was verified by comparing the indices of various nonlinear fitting methods. A typical neural network node is shown below. Figure 6 As shown.

[0079] This implementation constructs a deep neural network model with particle size and release height as independent variables, and the power levels of the initial impact sound and bubble pulsation sound as dependent variables, to fit the relationship between the independent and dependent variables. The radiated sound power level corresponding to each working condition obtained in step 2 is used as the dependent variable, and particle size and release height are used as independent variables. From 34 sets of self-measured particle size and release height-radiated sound power level data, 80% of the data (28 sets) is selected as training data for fitting, and 20% of the data (6 sets) is used as data to verify accuracy. This network model consists of a three-layer perceptron model, with an input layer, a hidden layer, and an output layer. A ReLU activation function is set between each two layers to introduce nonlinear fitting capability into the model.

[0080] The input layer contains two feature variables: the diameter of the sphere and the height of its impact point. Both input features are standardized using Z-scores, with the standardization parameters being the training set mean (diameter: 0.084m; height: 7.24m) and standard deviation (diameter: 0.062m; height: 5.11m). The hidden layers employ a two-level progressive structure: the first hidden layer has 32 neurons, introducing non-linear feature mapping through the ReLU activation function; the second hidden layer has 16 neurons, also using the ReLU activation function. This design is based on preliminary trial-and-error experiments, where overfitting occurred when the total number of hidden layer neurons was greater than or equal to 64 (validation set loss increased by 26%), hence the adoption of a 32+16 level configuration. The output layer contains two variables: the sound power level of the initial impact sound (L1) and the sound power level of the bubble pulsation sound (L2), ultimately determined by the coefficient of determination. Synthesize the overall sound power level.

[0081] 3.1 Experimental Setup: In this embodiment, the AdamW algorithm is used during model training, and its initial learning rate is set to... The weight decay coefficient is preferably set to . ; The selection of the above parameters is based on the experimental results of controlling variables: When the learning rate is greater than At times, the model oscillates during training and fails to converge; When the learning rate is lower than At that time, the model's convergence speed decreased by approximately [percentage missing] compared to the baseline. ; In the weight decay coefficient arrive Within the test range, the weight decay coefficient is equal to At that time, it can achieve a good balance between model complexity and generalization ability; Regarding learning rate scheduling, this embodiment preferably employs an exponential decay strategy, and sets the decay factor to... ,when At that time, the learning rate decreased too quickly, causing the model to fail to converge, therefore, the following was chosen: This ensures the stability of the training process; Regarding the training termination condition, this embodiment adopts a dual control strategy: on the one hand, it sets a maximum number of iterations. On the other hand, an early stopping mechanism is enabled, that is, when the validation set loss exceeds the threshold for continuous termination, the early stopping mechanism is activated. The process will automatically terminate if no decrease is observed within the specified timeframe to avoid overfitting; the batch size is set to 8.

[0082] With the above parameter configuration, the model converged in approximately 20,000 training epochs, meeting the training requirements.

[0083] 3.2 Experimental Results: Meanwhile, this implementation method compares several classic nonlinear regression models, such as multinomial regression, decision tree regression, and K-nearest neighbor regression, and examines the determination coefficients corresponding to the training results of each model. Root mean square error and mean absolute error As shown in Table 3.

[0084] Table 3 Comparison of Experimental Indicators for Nonlinear Regression Models

[0085] According to the indicators in the table above, among the aforementioned nonlinear regression methods, the indicator using a neural network model for fitting has the highest coefficient of determination. and root mean square error All of these parameters have reached their optimal levels, meeting the requirements for model fitting ability.

[0086] Step 4: Optimize the prediction model proposed in Step 3 to avoid overfitting due to the limited number of working conditions caused by experimental limitations, thereby improving the model's generalization ability and obtaining the final prediction model.

[0087] 4.1 Setting the Dropout strategy: like Figure 7 As shown, in order to alleviate the overfitting problem during training caused by the small amount of measurement data, this implementation uses the Dropout strategy to reduce the complexity of the model and improve its generalization ability.

[0088] In the strategy, the Dropout layer is set between the first hidden layer and the second hidden layer, and the dropout probability is set to 0.5. This was determined after comparative experiments with probabilities of 0.3, 0.5 and 0.7. When the probability is 0.5, the model performs best on the validation set, and the overfitting phenomenon is effectively alleviated.

[0089] 4.25-fold cross-validation: This is a widely used model performance evaluation technique in machine learning. Its core purpose is to more reliably estimate the model's generalization ability, that is, the model's ability to predict new, unseen data, while making more efficient use of limited datasets.

[0090] The 28 samples were randomly divided into 5 subsets (6+6+6+6+5). Four subsets were used as the training set and one as the validation set in turn. The final model performance was the average of the five trials. The cross-validation results are shown in Table 5, which displays the results of each trial. The values ​​fluctuated between 0.9899 and 0.9947 (standard deviation 0.0024), demonstrating the stability of the model.

[0091] Table 4. 5-fold cross-validation metrics

[0092] 4.3 Set up an ablation experiment to verify effectiveness: Ablation experiments are a crucial model analysis method in machine learning and deep learning. Their core idea is similar to a surgical excision, where removing a component or function from the model is used to observe its impact on overall performance, thereby determining the component's effectiveness and necessity. This implementation primarily analyzes the advantages of a double-hidden-layer structure and the effectiveness of the Dropout layer to determine the coefficients. , and The results are shown in Table 5 as a metric. Table 5 Comparison Indicators of Ablation Experiments

[0093] Ablation experiments showed that removing the Dropout layer and reducing one hidden layer... decline, and All metrics showed significant improvements, validating the effectiveness of Dropout and the advantages of its two-layer structure.

[0094] 4.4 Final Acoustic Prediction Model: In this embodiment, the sound power levels of the two fundamental components of the water entry sound—initial impact sound and bubble pulsation sound—are used as the main fitting variables. The total radiated sound power level is obtained by adding the fitting surfaces generated by the two variables point by point, and the resulting fitting surface is the final model established in this embodiment. The fitting diagrams of the bubble pulsation sound power level and the initial impact sound power level are shown below. Figure 8 , 9 As shown, the final radiated sound power level fitting diagram for the entire segment is as follows: Figure 10 As shown, this embodiment provides an error bar based on multiple sets of independent repeated experiments. Specifically, five sets of independent repeated experiments were conducted under each operating condition, and the errors of all test data are within [a certain range]. Within this range, the stability of this implementation method is verified. In the fitting graph, green represents training sample points, red represents test sample points, and the surface is the fitted surface. It can be seen from the fitting graph that the surface fitted by the model fits the sample point distribution well, and also performs well on the test samples, serving as a regular distribution image of particle size and drop height on near-field radiated sound power level. This is a predictive model; using this type of predictive model, accurate predictions can be achieved within a small range near the test conditions. For conditions that deviate significantly from the experimental range, predictions cannot be made at this time due to the lack of sample size and insufficient consideration of influencing factors.

[0095] The above detailed description of the technical solution provided by the present invention is intended to highlight the advantages and benefits of the technical solution provided by the present invention. However, the above detailed embodiments are not intended to limit the scope of protection of the present invention. Any reasonable modifications and improvements to the present invention, recombination of embodiments, and equivalent substitutions based on the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0096] Those skilled in the art will understand that the above description is merely a preferred embodiment of the present invention, and the features described in the various embodiments and / or claims disclosed in the present invention can be combined or combined in various ways, even if such combinations or combinations are not explicitly described in the disclosure of the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle scope of the present invention should be considered to fall within the protection scope of the present invention.

[0097] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if these modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include these modifications and modifications.

Claims

1. A method for constructing an acoustic prediction model for steel ball entering water, characterized in that, The method includes: S1. The time-domain signal 1 of the steel ball entering the water is acquired by the near-field virtual source method, and the time-domain signal 2 of the steel ball entering the water is acquired by the reverberation method. S2. Perform EMD preprocessing on both time-domain signals described in S1 to obtain denoised time-domain signal 1 and time-domain signal 2. S3. Calculate the sound power level of the denoised time-domain signal 1 and plot the first power spectrum; S4. Calculate the sound power level of the denoised time-domain signal 2 and plot the second power spectrum; S5. Filter and fuse the first power spectrum and the second power spectrum, and use integral calculation to obtain the radiated sound power level of the whole segment when the steel ball enters the water. S6. Extract the initial impact sound of the noise-reduced time-domain signal 1 and calculate the sound power level to obtain the radiated sound power level of the initial impact sound when the steel ball enters the water. S7. Subtract the radiated sound power level of the initial impact sound from the radiated sound power level of the entire segment to obtain the radiated sound power level of the bubble pulsation sound; S8. Repeat steps S1-S7 to obtain the radiated sound power levels of the whole section, initial impact sound and bubble pulsation sound corresponding to different working conditions. S9. Plot the curves of the radiated sound power level of the whole section, the initial impact sound, and the bubble pulsation sound under different working conditions as a function of the steel ball particle size and the release height to obtain the variation law. S10. Based on the variation law, a neural network prediction model is constructed with steel ball diameter and release height as independent variables and the radiated sound power level of initial impact sound and bubble pulsation sound as dependent variables. S11. The neural network prediction model is trained using the mean squared error loss function. During the training process, the Dropout strategy is introduced, and the neural network prediction model is trained using 5-fold cross-validation. S12. Obtain the acoustic prediction model for the steel ball entering the water.

2. The method according to claim 1, characterized in that, The neural network prediction model described in S10 is a three-layer perceptron model, including an input layer, a hidden layer, and an output layer. The input layer contains two input features: the steel ball diameter and the height from which it falls into the water. The hidden layer has a two-level structure: the first hidden layer has 32 neurons, and the second hidden layer has 16 neurons. Activation function; the output layer includes two output variables, namely the initial impact sound power level and the bubble pulsation sound power level.

3. The method according to claim 2, characterized in that, The neural network prediction model can be represented as: ， in, For input variables, For output variables, This represents the input variable of the i-th multilayer perceptron. Its function.

4. The method according to claim 2, characterized in that, The The activation function is a rectified linear unit function, defined as: 。 5. The method according to claim 1, characterized in that, The method for training the neural network prediction model described in S11 is as follows: First, the mean squared error loss function is used to calculate the loss value for the model: ， in, For loss function, The radiated sound power level predicted by the model. For binary variables, This represents the number of data sample groups; Secondly, various nonlinear fits to the model are compared, and the fitting effect is evaluated using the following three metrics: ， ， ， in, It is the coefficient of determination. It is the root mean square error. It is the mean absolute error. It is the sum of squared residuals. It is the total sum of squares. It is the number of observations. This is the nth actual observation value. It is the first Each model predicts a value.

6. The method according to claim 1, characterized in that, The expression for the Dropout strategy described in S12 is: ， in, Represents element-wise multiplication. and It is a random 0-1 vector.

7. The method according to claim 1, characterized in that, The 5-fold cross-validation method described in S12 is as follows: The dataset was divided into 5 subsets. One subset was selected as the test set and the rest were used as the training set. This process was repeated 5 times. The average of the 5 test results was then used as the performance evaluation result of the neural network prediction model.

8. An apparatus for constructing an acoustic prediction model of a steel ball entering water, characterized in that, The device includes: Acquisition module: Acquires time-domain signal 1 of steel ball entering water using near-field virtual source method, and acquires time-domain signal 2 of steel ball entering water using reverberation method; EMD module: Used to perform EMD preprocessing on both types of time-domain signals in the acquisition module to obtain noise-reduced time-domain signal 1 and time-domain signal 2; Power Spectrum Module 1: Used to calculate the sound power level of the denoised time-domain signal 1 and plot the first power spectrum; Power Spectrum Module 2: Used to calculate the sound power level of the denoised time-domain signal 2 and plot the second power spectrum; Fusion module: used to filter and fuse the first power spectrum and the second power spectrum, and use integral calculation to obtain the radiated sound power level of the whole segment when the steel ball enters the water; The initial impact sound module is used to extract the initial impact sound of the noise-reduced time-domain signal 1 and calculate the sound power level to obtain the radiated sound power level of the initial impact sound when the steel ball enters the water. The bubble pulsation module is used to obtain the radiated sound power level of the bubble pulsation sound by subtracting the radiated sound power level of the initial impact sound from the radiated sound power level of the entire segment. Multi-condition module: used to repeatedly collect data from the module to the bubble pulsation module to obtain the radiated sound power levels of the whole section, initial impact sound and bubble pulsation sound corresponding to different conditions; Dynamic Curve Module: Used to plot the radiated sound power level of the whole section, initial impact sound and bubble pulsation sound under different working conditions as a function of steel ball particle size and release height, and to obtain the variation law; Model building module: used to build a neural network prediction model based on the change pattern, with steel ball diameter and release height as independent variables and the radiated sound power level of initial impact sound and bubble pulsation sound as dependent variables; Model training module: used to train the neural network prediction model using the mean squared error loss function. During the training process, the Dropout strategy is introduced, and the neural network prediction model is adopted with 5-fold cross-validation. Output module: Used to obtain the acoustic prediction model of the steel ball entering the water.

9. A computer program product, as a computer program, is characterized by: When the computer program is read, the method described in any one of claims 1-7 is implemented.

10. A computer storage medium for storing computer programs, characterized in that, When the computer program is read by the computer, the computer executes the method according to any one of claims 1-7.

Images