A high-fidelity generation method for dynamic radiation noise of a ship

Abstract

This invention discloses a highly realistic method for generating dynamic radiated noise from ships, belonging to the field of underwater acoustic countermeasures technology. The invention proposes a fully simulation-driven solution: First, a two-dimensional core dataset of "speed-noise characteristics" covering all operating conditions—no cavitation, cavitation development, and cavitation saturation—is generated using a ship dynamics simulation system. Then, a CNN-LSTM hybrid model is constructed to accurately predict the core parameters of the line spectrum and continuous spectrum using the target speed as input. Cavitation states are divided based on the critical cavitation speed, and dynamic weights are preset for the line spectrum, continuous spectrum, and modulation envelope. A lightweight GAN is used to generate a frequency-band modulation envelope, and the fluctuation parameters obtained from simulation statistics are combined to simulate the amplitude fluctuations and frequency drift of the line spectrum. Finally, the signal is synthesized according to the actual noise generation mechanism. This invention requires no measured data, has high parameter prediction accuracy, good auditory realism, and can be quickly adapted to different types of ships and cavitation conditions, making it suitable for underwater acoustic countermeasures equipment such as self-propelled acoustic decoys.

Description

Technical Field

[0001] This invention relates to the field of underwater acoustic countermeasures technology, specifically to a highly realistic method for generating dynamic radiated noise of ships based on simulation-generated data, applicable to self-propelled acoustic decoys. Background Technology

[0002] Existing methods for generating dynamic radiated noise from ships suffer from technical problems such as inaccurate parameter prediction, rigid feature fusion, fuzzy modulation envelope modeling, and unrealistic line spectrum simulation. The main issues are as follows:

[0003] High data dependence: Traditional methods rely on a large amount of measured data collection and preprocessing, which is not only cumbersome to operate, but also limited by the measured environment and equipment conditions, making it difficult to quickly adapt to different types of ships.

[0004] Insufficient parameter prediction accuracy: The use of single models such as BP neural networks does not fully capture the correlation between speed and noise characteristics, resulting in large parameter mapping errors at different speeds and poor dynamic simulation effects.

[0005] The feature fusion method is rigid: it uses a fixed ratio to superimpose line spectrum, continuous spectrum and modulation envelope, without considering the impact of cavitation state differences caused by speed on the contribution of each feature, resulting in poor adaptability;

[0006] The modulation envelope modeling is fuzzy: the modulation parameters (modulation degree, pulse amplitude, etc.) are determined by "black box processing" combined with measured statistical characteristics, which lacks precise modeling logic and has limited auditory realism.

[0007] The line spectrum simulation is unrealistic: the simulation of line spectrum instability uses a slow-changing Gaussian process with fixed parameters, without taking into account the fluctuation law related to the speed of the ship, resulting in deviations from the actual noise line spectrum characteristics of ships. Summary of the Invention

[0008] This invention provides a method for generating dynamic radiated noise of ships with high fidelity, based on data generated by a simulation system, adaptable to different air and chemical conditions.

[0009] The specific solution adopted in this invention is as follows:

[0010] A method for generating highly realistic dynamic radiated noise of ships includes the following steps:

[0011] Step 1: Construct a two-dimensional core dataset of "speed-noise characteristics" using a ship dynamics simulation system;

[0012] Step 2: Based on the generated two-dimensional core dataset of "ship speed-noise features", train the CNN-LSTM hybrid model; the CNN-LSTM hybrid model takes the target ship speed parameter as input and outputs the core quantization parameters of the ship's radiated noise. The core quantization parameters include the line spectrum reference amplitude, the line spectrum reference frequency, the continuous spectrum peak frequency, the continuous spectrum peak level, and the coefficient of the spectrum change curve.

[0013] Step 3: Based on the input target speed, combined with the pre-calculated initial cavitation speed and saturation cavitation speed, the target speed is divided into three cavitation states: no cavitation, cavitation development, and cavitation saturation. Based on the cavitation state, a fixed weight ratio is preset for the line spectrum component, continuous spectrum component, and modulation envelope component.

[0014] Step 4: Based on the target speed and cavitation status indicators in Step 3, combined with the propeller shaft frequency, it is used as the input of the lightweight generative adversarial network (GAN). The GAN outputs the core parameters of the modulation envelope in 28 frequency bands through the generator, including modulation degree, pulse amplitude and pulse width, and then synthesizes the modulation envelope signal through the Gaussian pulse train model.

[0015] Step 5: Based on the simulation dataset of "speed-noise characteristics" in Step 1, statistically determine the variance of the line spectrum amplitude fluctuation and the low-pass cutoff frequency and variance of the frequency drift in each speed range according to the cavitation state, and use them as fixed fluctuation parameters; based on the cavitation state determined in Step 3, call the corresponding fixed fluctuation parameters to perform perturbation simulation on the line spectrum reference amplitude and line spectrum reference frequency predicted in Step 2, and generate unstable line spectrum parameters, including the time-varying actual line spectrum amplitude and actual line spectrum frequency.

[0016] Step 6, based on the line spectrum weights determined in Step 3 Continuous spectrum weights and modulation envelope weights The modulation envelope signal generated in step 4 The unstable line spectrum parameters generated in step 5 are used to generate the final ship dynamic radiated noise signal. : in, These are line spectral components constructed based on unstable line spectral parameters. The continuous spectrum components are reconstructed based on the continuous spectrum parameters output in step 2.

[0017] Furthermore, the specific process of step 1 is as follows:

[0018] The ship dynamics simulation system is based on the full-link logic of "noise source modeling - propagation process simulation - time domain signal synthesis" to output time domain signals. Among them, mechanical noise adopts a multi-degree-of-freedom vibration simulation model, propeller noise cavitation state adopts blade fluid impact theory or bubble dynamics equation, and hydrodynamic noise adopts Reynolds number correlation model. The hull conduction simulation is attenuated by 0-15dB, and the seawater medium attenuation is attenuated by frequency domain correlation.

[0019] The simulation input parameters are set using a two-dimensional configuration mode of "basic parameters + operating condition parameters". The speed range is set according to the principle of "no cavitation - cavitation development - cavitation saturation", with a minimum speed interval of 0.1 m / s. The speed range for no cavitation is set to 3-5 levels, the speed range for cavitation development is set to 5-8 levels, and the speed range for cavitation saturation is set to 3-5 levels. The initial speed and saturation speed of cavitation are automatically calculated by the fluid dynamics formula built into the simulation system.

[0020] The built-in feature extraction engine extracts the following from the time-domain signal: line spectrum frequency, line spectrum amplitude, peak frequency, peak level, 1 / 3 octave band sound pressure level, axis frequency and first eight harmonics of the DEMON spectrum, 12th order MFCC coefficients, total loudness peak value and distribution variance of Moore characteristic loudness, forming a structured two-dimensional dataset.

[0021] Furthermore, the CNN-LSTM hybrid model structure in step 2 is an end-to-end structure of "input layer → CNN layer → LSTM layer → fully connected output layer"; wherein:

[0022] The input layer receives the target speed parameters and uses the Min-Max normalization algorithm to map the speed parameters to the [0,1] interval. The normalization algorithm formula is as follows:

[0023]

[0024] in, for ; for ; for ; ;

[0025] Reflecting the normalized speed Shoot into the [0,1] interval, then convert it into a shape of [ , A three-dimensional tensor; where, This represents the number of training samples in a batch. Because the time was short, For feature dimensions;

[0026] The CNN layer is configured with 32 one-dimensional convolutional kernels of size 3, stride 1, padding method 'same', activation function ReLU, and dropout=0.2, used to extract the local nonlinear correlation between speed and noise frequency domain features, and output tensor;

[0027] The LSTM layer uses a two-layer serial structure. The first layer outputs the hidden state at each step and provides timing information to the second layer. The second layer outputs the hidden state at the last step and extracts the global temporal core features. Each layer has 64 hidden units and a dropout rate of 0.2.

[0028] Used to capture the temporal evolution patterns during changes in speed;

[0029] The fully connected output layer has two layers. The first layer has 128 neurons and uses the ReLU activation function. The second layer has the same number of neurons as the core output parameters and uses the linear activation function. The output includes the reference amplitude of the line spectrum, the reference frequency of the line spectrum, the peak frequency of the continuous spectrum, the peak level of the continuous spectrum, and the coefficients of four spectral variation curves.

[0030] The model training adopts supervised learning, using "speed" from the dataset in step 1 as the input sample and "line spectrum parameters + continuous spectrum parameters" as the labels. The training set and validation set are divided in an 8:2 ratio. The batch size of the training parameters is 32, the initial learning rate is 0.001, and the ReduceLROnPlateau decay strategy is adopted. The optimizer is Adam, the loss function is MSE, and the convergence condition is that the MSE of the training set ≤ 1e-4 and the MSE of the validation set ≤ 1.5e-4, or the change rate of the MSE of the validation set ≤ 0.001% for 20 consecutive rounds.

[0031] Furthermore, in step 3:

[0032] Initial speed of cavitation The calculation formula is:

[0033] Cavitation saturation speed The calculation formula is: J P For the operating advance coefficient, H represents the wake fraction, and H represents the navigation depth. For nascent vacuoles;

[0034] The cavitation state determination rule is: v≤ It is a non-vacuolated state. <v< For cavitation development state, v≥ It is in a cavitation saturation state;

[0035] The preset fixed weight ratio is as follows: No vacuolation state: =0.4, =0.3, =0.3; Cavitation development status: =0.3, =0.4, =0.3; Cavitation saturation state: =0.2, =0.45, =0.35.

[0036] Furthermore, the lightweight GAN network in step 4 includes a generator and a discriminator; wherein:

[0037] The generator uses a 3-layer fully connected network. The input layer has 3 neurons, corresponding to the target speed v, propeller shaft frequency f_z, and cavitation state flag. The first hidden layer has 128 neurons, the second hidden layer has 64 neurons, and the activation function is ReLU. The output layer has 84 neurons, the activation function is Sigmoid, and the output has 28 frequency bands, each with its own modulation, pulse amplitude, and pulse width.

[0038] The discriminator is used in the training phase. Its input is the modulation envelope signal synthesized by the Gaussian pulse train model from the modulation parameters output by the generator or the real modulation envelope signal extracted from the simulation dataset in step 1. Its output is the probability of judging whether it is true or false.

[0039] The GAN training data comes from the dataset in step 1, and is divided into training and validation sets in an 8:2 ratio. The batch size is 16, the initial learning rate is 0.0001, the optimizer is Adam, and the loss function is binary cross-entropy. The generator and discriminator are trained alternately. The convergence condition is that the discriminator's accuracy in judging the generated signal and the real signal in the validation set is close to 50% ± 3%, and the loss change rate is ≤ 0.001% for 10 consecutive rounds.

[0040] After training convergence, the generator is retained for inference. During inference, based on the cavitation state flag and target speed input in step 3, the generator outputs modulation parameters, which are then synthesized into a modulated envelope signal using a Gaussian pulse train model. The formula is: Where B is the number of propeller blades and T is the propeller rotation period; the synthesized... The signals are combined into a total modulation envelope signal and output to step 6.

[0041] Furthermore, in step 5,

[0042] The formula for simulating line spectrum amplitude fluctuations is: ,

[0043] in, The reference amplitude of the line spectrum output in step 2. The random variable is uniformly distributed, with a value range of [-√3σ_A, √3σ_A], where σ_A is the variance of amplitude fluctuations obtained statistically from the dataset in step 1 based on the cavitation state determined in step 3; each sampling time is generated independently by a pseudo-random number generator. ;

[0044] The line spectrum frequency drift simulation uses a slowly changing Gaussian process:

[0045] ,

[0046] in, Let be the actual frequency of the line spectrum at time t. The reference frequency of the line spectrum output in step 2. Let be the frequency drift at time t, and the recursive formula is:

[0047] in and These are the low-pass cutoff frequency and frequency drift variance, obtained statistically from the dataset in step 1 based on the cavitation state. Let N(0,1) be the sampling frequency, N(0,1) be a standard normally distributed random variable, and Δ be the initial value. (0)=0.

[0048] Furthermore, after step 6, verification is performed, and the verification process is as follows:

[0049] In step 1, the simulation system calculates the time-domain, frequency-domain, and auditory-domain features of the synthesized signal and compares them with the features of the corresponding working conditions in the dataset of step 1, requiring a relative error of ≤5%. The time-domain features include root mean square and kurtosis, the frequency-domain features include spectral peak frequency and spectral peak level, and the auditory-domain features include MFCC coefficients and Moore loudness.

[0050] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0051] 1. This invention breaks through the dependence on actual measurement data. The entire process uses a simulation system to generate a "speed-noise feature" dataset, which supports rapid adaptation to multiple types of ships and can solve the problems of insufficient actual measurement data and poor adaptability of traditional methods.

[0052] 2. This invention uses a CNN-LSTM hybrid model to replace the traditional BP neural network, while capturing the local correlation and temporal evolution of speed and noise features, avoiding local optima and improving the prediction accuracy of core parameters.

[0053] 3. This invention abandons the fixed ratio fusion mode and dynamically presets the weight ratio based on the cavitation state to accurately match the energy ratio of each noise component under different working conditions, thus solving the problem of insufficient realism caused by the rigidity of traditional fusion.

[0054] 4. This invention uses a GAN network to replace the "black box processing" to generate modulation envelope parameters. Through adversarial training between the generator and the discriminator, the modulation features are made to fit the real distribution more closely, and the realism of the auditory domain is significantly improved.

[0055] 5. This invention uses speed-specific fluctuation parameters based on simulation data statistics to simulate spectrum instability. Compared with a Gaussian process with a single fixed parameter, it better reflects the natural fluctuation pattern of the spectrum at different speeds.

[0056] 6. This invention integrates multi-dimensional feature extraction and fusion in the frequency domain and auditory domain, and the synthesis formula restores the real noise generation mechanism of "modulated continuous spectrum + superimposed line spectrum", ensuring that the signal is highly realistic in the time domain, frequency domain and auditory domain.

[0057] 7. This invention adopts a standardized preprocessing + tensor format conversion input optimization scheme, combined with the Adam optimizer and learning rate decay strategy, to improve the model training convergence efficiency and generalization ability. Attached Figure Description

[0058] Figure 1 This is a flowchart of an embodiment of the present invention. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0060] The specific steps of the method for generating high-fidelity dynamic radiated noise of ships disclosed in this embodiment are as follows:

[0061] Step 1: The simulation system generates the core dataset.

[0062] Using a professional-grade ship dynamics simulation system, a two-dimensional core dataset of "speed-noise characteristics" covering all operating conditions is constructed. The dataset generation process does not rely on any measured data acquisition or preprocessing; the entire process of modeling, calculation, and feature extraction is completed autonomously by the simulation system, specifically in the following three parts:

[0063] 1. Definition and configuration of simulation input parameters

[0064] The simulation input parameters adopt a two-dimensional configuration mode of "basic parameters + operating condition parameters" to ensure that the parameters cover the key influencing factors of ship noise generation and support flexible adaptation to different types of ships.

[0065] (1) Basic parameters of the ship:

[0066] Core parameters of a propeller: propeller diameter, number of blades, operating speed coefficient, and initial cavitation number;

[0067] Ship structural parameters: displacement tonnage, length, draft, wake fraction;

[0068] Basic parameters of noise sources: rated power of main engine, number of auxiliary engines in operation, and surface roughness of hull.

[0069] (2) Operating speed parameters:

[0070] Speed ​​gear design: Based on the principle of full operating condition coverage of "no cavitation - cavitation development - cavitation saturation", multiple continuous speed values ​​are set, with a minimum speed interval of 0.1m / s;

[0071] Among them, the cavitation-free speed range (≤ cavitation initiation speed) Set to levels 3-5, cavitation development range ( <Speed ​​<Cavitation Saturation Speed Set to level 5-8, cavitation saturation range (≥ Set to level 3-5;

[0072] Speed ​​correlation calculation: i and The simulation system automatically calculates cavitation critical speeds using built-in fluid dynamics formulas. These formulas are optimized based on an improved Ross model. After inputting parameters such as propeller diameter, initial cavitation number, wake fraction, and depth, the system outputs the corresponding critical cavitation speeds in real time, ensuring that the speed range accurately covers various cavitation states.

[0073] All input parameters support digital entry and template import. The system has built-in parameter templates for three types of conventional ships (small patrol boats, medium-sized frigates, and large destroyers), which users can directly call or modify based on the templates.

[0074] 2. Data generation logic and time-domain signal output

[0075] The simulation system is based on the physical generation mechanism of ship radiated noise and constructs a full-link generation logic of "noise source modeling - propagation process simulation - time-domain signal synthesis" to ensure that the output signal closely matches the actual generation and propagation laws of noise.

[0076] (1) Multi-source noise modeling:

[0077] Mechanical noise modeling: A multi-degree-of-freedom vibration simulation model is used to simulate the operating vibration of mechanical components such as the main machine, auxiliary machine, and drive shaft. By inputting parameters such as mechanical speed, vibration amplitude, and natural frequency, the time-domain vibration signal of each mechanical component is generated, and then converted into the underwater radiated mechanical noise time-domain signal through the shell conduction coefficient.

[0078] Propeller noise modeling: Modeling is divided into cavitation state. When there is no cavitation, the noise generated by the interaction force between the blade and the seawater is calculated based on the propeller blade fluid impact theory. When cavitation develops and saturates, the bubble dynamics equation is used to simulate the generation, growth and collapse of cavitation bubbles on the back of the blade, and the cavitation noise time domain signal is output. Its amplitude increases nonlinearly with the increase of speed, and the peak frequency of the spectrum decreases with the increase of speed.

[0079] Hydrodynamic noise modeling: Based on the theory of turbulent noise generation and combined with the velocity distribution on the hull surface, the Reynolds number correlation model is used to generate the frictional turbulent noise on the hull surface and the vortex noise generated by the appendages. The amplitude of the time domain signal is proportional to the square of the speed.

[0080] (2) Simulation of noise propagation process:

[0081] Hull conduction simulation: Considering the conduction attenuation of different noise sources in the hull structure, conduction attenuation of 0-15dB is applied according to the location of the noise source and the length of the propagation path;

[0082] Seawater medium attenuation simulation: Based on the formula of seawater acoustic absorption coefficient, frequency domain correlation attenuation is applied to the noise signal. The attenuation in the high frequency band is greater than that in the low frequency band to ensure that the propagation process conforms to the underwater acoustic law.

[0083] (3) Time-domain signal synthesis and output:

[0084] Signal synthesis: The time-domain signals of mechanical noise, propeller noise and hydrodynamic noise are superimposed proportionally, and after superposition, signal abrupt changes are eliminated by smoothing filtering to generate the final ship radiated noise time-domain signal;

[0085] Output parameters: The sampling frequency is fixed at 524288Hz, the signal duration at single speed is 10s, the signal amplitude unit is Pa, the output format is binary raw data, and it supports direct import into the feature extraction module.

[0086] 3. Automatic Feature Extraction and Dataset Construction

[0087] The simulation system has a built-in feature extraction engine that automatically extracts features from the output time-domain signals at various flight speeds, forming a structured two-dimensional dataset of "flight speed-noise features". The extraction process requires no manual intervention, as detailed below:

[0088] (1) Frequency domain feature extraction:

[0089] Line spectrum characteristics: An adaptive line spectrum enhancement algorithm is used to separate the line spectrum from the power spectrum of the time-domain signal, extract the core parameters of the line spectrum: line spectrum frequency and line spectrum amplitude, and distinguish between speed-dependent line spectra and speed-independent line spectra through spectral stability analysis, and store them separately;

[0090] Continuous spectrum characteristics: The Welch power spectrum estimation method is used to calculate the continuous spectrum, extract the peak frequency and peak level, divide the frequency bands according to 1 / 3 octave, calculate the sound pressure level of each frequency band, and form the discretization coefficient of the spectrum variation curve.

[0091] DEMON Spectral Characteristics: Demodulation analysis of time-domain signals is performed to extract the core parameters of the modulation spectrum: shaft frequency, first eight harmonic frequencies, and normalized amplitude, which are used to characterize the propeller modulation characteristics.

[0092] (2) Auditory domain feature extraction:

[0093] Mel frequency cepstral coefficients: The time-domain signal is pre-emphasized, framed, windowed, FFT, Mel filter, logarithmic transform, and discrete cosine transform to extract the 12th order MFCC coefficients as auditory domain pitch features;

[0094] Moore characteristic loudness: Based on the peripheral model of human hearing, the signal is divided into 34 frequency bands according to the equivalent rectangular bandwidth, the loudness value of each frequency band is calculated, and the total loudness curve is obtained through the loudness summation model. The peak value of the total loudness and the variance of the loudness distribution are extracted as loudness features of the auditory domain.

[0095] (3) Dataset construction:

[0096] Data structuring: Using speed as an index, all feature parameters at that speed are associated and stored to form a single speed feature vector;

[0097] Dataset integration: All speed feature vectors are arranged in ascending order of speed to form a complete two-dimensional dataset of "speed-noise features". The dataset is in CSV structured file format, containing metadata descriptions such as feature name, data type, unit, and value range. It can be directly imported into subsequent model training modules without additional data cleaning or format conversion.

[0098] Step 2: Training and core parameter prediction of the CNN-LSTM hybrid model

[0099] 1. Model Input

[0100] The model input is a single-dimensional target speed parameter. However, to adapt to the input requirements and training convergence efficiency of the neural network, the speed parameter needs to be standardized and preprocessed before being converted into a tensor format that the model can recognize. The specific details are as follows:

[0101] 1. Physical definition and range of values ​​for speed parameters

[0102] The input target speed parameter (denoted as) The underwater speed of the ship is expressed in m / s, and its value range is consistent with the speed coverage range of the simulation dataset in step 1. ,in Minimum speed for zero-air mobility, Maximum speed for cavitation saturation 0.1 m / s, matching the speed gear accuracy of the simulation dataset.

[0103] 2. Standardization preprocessing of speed parameters

[0104] Because the speed parameter has different numerical ranges, direct input will lead to slow model training convergence and gradient explosion. Therefore, the Min-Max normalization algorithm is used to map the speed parameter to the [0,1] interval. The normalization formula is as follows:

[0105]

[0106] in:

[0107]

[0108] The normalization process is completed automatically by the model's built-in preprocessing module without manual intervention, and there is no need to reverse normalize the speed parameters after the prediction is completed.

[0109] 3. Tensor format of model input

[0110] To adapt to the input requirements of the CNN-LSTM hybrid model, the normalized speed parameters were... Converted to a 3D tensor format, the tensor shape is defined as follows: ,in:

[0111] The number of training samples in a batch is preset to 32.

[0112] The time step is set to 1 here;

[0113] Feature dimension, set to 1 here.

[0114] The final input model has a speed tensor of [32,1,1] or [1,1,1].

[0115] 2. Model Structure

[0116] An end-to-end hybrid model structure is constructed, consisting of an input layer, a CNN layer, an LSTM layer, and a fully connected output layer. The CNN layer is responsible for extracting the local correlation between the frequency domain features of speed and noise, while the LSTM layer is responsible for capturing the temporal evolution of noise features during speed changes. The two layers complement each other to maximize the accuracy of parameter prediction. The detailed structure of each layer is as follows:

[0117] 1. Input layer

[0118] The core function of the input layer is to receive the standardized speed tensor and pass it to the subsequent network layers. There are no additional computational operations; it only performs data format verification and dimension matching.

[0119] Layer structure parameters: The input dimension has the same shape as the speed tensor, i.e. (None supports input adaptation for any batch size);

[0120] Output: The velocity tensor is directly passed to the CNN layer, and the output shape is consistent with the input shape. .

[0121] 2. CNN layer

[0122] A single-hidden-layer CNN structure is adopted, configured with 1D convolutional kernels (adapted to single speed feature input). The core function is to extract the local nonlinear correlation between speed and noise frequency domain features, avoiding the loss of feature information. Specific configuration and core formula are as follows:

[0123] 1. Layer structure parameter configuration

[0124] Number of convolutional kernels: 32 (to ensure sufficient extraction of local features);

[0125] Kernel size: (1D convolution kernel, covering feature associations of adjacent speed levels);

[0126] Step size: (No skipping sampling, preserving complete feature information);

[0127] Fill method: (Zero padding ensures that the dimensions of the output feature map after convolution are consistent with the input).

[0128] Activation function: ReLU function (avoids gradient vanishing and accelerates model convergence);

[0129] Dropout layer: Configure dropout=0.2 (randomly drop 20% of neurons to prevent model overfitting).

[0130] 2. Core Calculation Formula

[0131] 1D convolution operation formula:

[0132]

[0133] in:

[0134] After convolution, the first Feature values ​​at each position;

[0135] No. The weight parameters of each convolutional kernel;

[0136] Input feature map Input values ​​at each position;

[0137] Bias parameters of convolutional layers;

[0138] The summation range covers the entire kernel size (here)

[0139]

[0140] 3. Layer output shape

[0141] After convolution, activation, and dropout processing, the output tensor of the CNN layer has the shape of [None, 1, 32], where 32 is the number of convolution kernels. This tensor is then reshaped to [None, 1, 32] and directly passed to the LSTM layer.

[0142] 3. LSTM layer

[0143] 1. Structural parameter configuration:

[0144] (1) Number of LSTM layers: Set up 2 serial LSTM layers. The first layer is in "return sequence" mode (outputs the hidden state of each step to provide complete timing information for the second layer), and the second layer is in "do not return sequence" mode (only outputs the hidden state of the last step to extract the global timing core features).

[0145] (2) Number of hidden units: The number of hidden units in each LSTM layer is set to 64, which can fully capture the dependencies of temporal features, while avoiding the surge in computation and overfitting caused by too many hidden units;

[0146] (3) Regularization configuration: A Dropout layer is introduced with a dropout rate of 0.2 (i.e., randomly discarding 20% ​​of the hidden unit outputs), which effectively prevents the model from overfitting on the simulation dataset and improves the model's generalization ability;

[0147] (4) Activation function: The LSTM unit uses the tanh activation function (used to calculate candidate hidden states, with a value range of [-1,1]), and the gating unit (input gate, forget gate, output gate) uses the sigmoid activation function (used to control the transmission and discard of information, with a value range of [0,1]).

[0148] 2. Workflow: Receive the feature tensor output by the CNN layer, treat it as a temporal sequence of length 1, extract the temporal correlation of local features through the first LSTM layer, and then integrate the global temporal information through the second LSTM layer, finally outputting a fused feature vector containing "local correlation + global temporal sequence".

[0149] 3. Output format: The output fused feature vector is a two-dimensional tensor with a shape of (1,64) or (N,64) (number of samples × number of hidden unit features). This vector contains all the correlation information between the core parameters of speed and noise, and is directly passed to the fully connected output layer for parameter mapping.

[0150] 4. Fully Connected Output Layer

[0151] Core function: The fused feature vector output by the LSTM layer is converted into specific core parameters of ship noise through linear mapping and nonlinear transformation, realizing the final mapping from "feature vector" to "quantization parameter" and providing directly usable quantization data for subsequent steps;

[0152] Structural parameter configuration:

[0153] Number of fully connected layers: Set 2 fully connected layers, the first layer is a hidden fully connected layer, and the second layer is an output fully connected layer;

[0154] Number of neurons: The number of neurons in the first hidden fully connected layer is 128, which is used to further integrate features and perform nonlinear transformations on the fused feature vector; the number of neurons in the second output fully connected layer is consistent with the dimension of the core parameters of the model output (generally 20~30, corresponding to all core parameters of the line spectrum and the continuous spectrum).

[0155] Activation function: The first hidden fully connected layer uses the ReLU activation function to introduce nonlinear transformation and improve the accuracy of parameter mapping; the second output fully connected layer uses the linear activation function (Identity activation function) because the line spectrum frequency, amplitude and continuous spectrum parameters of the model output are all continuous values, and no nonlinear transformation is required, ensuring the authenticity and quantization accuracy of the output parameters;

[0156] Output format: The core parameters of the output are two-dimensional tensors with a shape of (1,M) or (N,M) (number of samples × number of core parameter dimensions M). Each element in the tensor corresponds to a specific noise core parameter and has been denormalized. The physical quantity (with unit) is directly output without any subsequent conversion.

[0157] 3. Model Output

[0158] The model output contains the core quantification parameters of ship radiated noise, including only the key parameters required for subsequent steps, and eliminating redundant features.

[0159] The line spectrum parameters focus on the "speed-related line spectrum", including the line spectrum frequency and the line spectrum reference amplitude: the line spectrum frequency is the center frequency value of the speed-related line spectrum, in Hz, with an accuracy of 0.1Hz. The frequency of the first 10 main speed-related line spectra is output, reflecting the correlation between the ship's propeller speed, number of blades and speed.

[0160] The baseline amplitude of the line spectrum is the reference sound pressure amplitude value of each speed-related line spectrum under undisturbed conditions, in Pa, with an accuracy of 0.01 Pa, serving as the basic reference value for subsequent line spectrum instability simulations. The continuous spectrum parameters focus on the overall characteristics of the continuous spectrum of ship radiated noise, including peak frequencies, peak levels, and coefficients of the spectral variation curve: the peak frequency is the center frequency corresponding to the energy peak in the continuous spectrum power spectrum, in Hz, with an accuracy of 1 Hz, reflecting the energy concentration band of ship noise; the peak level is the sound pressure level at the peak frequency of the continuous spectrum, in dB, with an accuracy of 0.1 dB, reflecting the energy intensity of the continuous spectrum.

[0161] The coefficients of the spectral variation curve are obtained by fitting the shape of the continuous spectrum using a cubic polynomial. They consist of four fitting coefficients, are dimensionless, and have a precision of 1e-6. These coefficients are used to reconstruct the complete shape of the continuous spectrum. All output parameters are stored in a structured format, allowing direct import into subsequent feature fusion modules without additional formatting.

[0162] 4. Model Training Process

[0163] The entire model training process is based on the "speed-noise feature" simulation dataset generated in step 1. A supervised learning method is employed, aiming to minimize the deviation between predicted and simulated values, and iterative training and optimization are completed. First, training samples are preprocessed and the dataset is divided. The CSV format dataset is imported into the model training platform, and "speed parameters" are automatically extracted as input samples, while "line spectrum parameters + continuous spectrum parameters" are extracted as output labels. A stratified sampling method is used to divide the dataset into training and validation sets in an 8:2 ratio, ensuring that the proportion of samples in each speed range is consistent across the two sets, avoiding uneven sample distribution that could lead to insufficient model generalization ability.

[0164] In terms of training parameter configuration, the batch size is set to 32 to balance training efficiency and convergence stability; the maximum preset training epochs are 200 epochs, supporting an early stopping mechanism; the initial learning rate is set to 0.001, employing a "ReduceLROnPlateau" learning rate decay strategy. When the validation set loss does not decrease for 10 consecutive epochs, the learning rate decays to its original value of 0.5, and continues to decay until it drops to 1e-6. The optimizer is Adam, configured with a first-order moment estimation decay coefficient (β1) of 0.9, a second-order moment estimation decay coefficient (β2) of 0.999, and a numerical stability term (ε) of 1e-8. During backpropagation, the gradient is automatically calculated and parameters are updated to minimize the loss function value.

[0165] The loss function uses mean squared error (MSE), and its core objective is to minimize the mean squared error between the model's predicted values ​​and the true values ​​in the simulation dataset. The specific formula is as follows:

[0166]

[0167] Where n is the number of training samples, and M is the number of dimensions of the output core parameters. To simulate the true value, The smaller the loss function value, the higher the prediction accuracy.

[0168] The iterative training process follows a cyclical logic of "forward propagation - back propagation - parameter update - batch iteration - validation and evaluation". After each training round, the loss value is calculated using the validation set samples and the changes are recorded. Convergence is determined when either of the following conditions is met: the MSE loss value of the training set ≤ 1e-4 and the MSE loss value of the validation set ≤ 1.5e-4; the rate of change of the MSE loss value of the validation set ≤ 0.001% over 20 consecutive training rounds. After the model converges, the complete model file (including model structure, weight parameters, optimizer state, and standardized / destandardized calibration values) is automatically saved. The saved format is compatible with mainstream deep learning frameworks and supports direct loading for parameter prediction later.

[0169] After the model training is completed, the prediction effect is verified using a validation set. The core verification index is the relative error of each core parameter. The requirements are: relative error of line spectrum frequency prediction ≤2%, relative error of line spectrum reference amplitude prediction ≤3%, relative error of continuous spectrum peak frequency prediction ≤1.5%, and relative error of continuous spectrum peak level prediction ≤0.5dB. After the verification is passed, it can be put into formal use. By inputting any target speed parameter (within the speed coverage of step 1), the model can quickly output accurate core noise parameters, providing reliable support for the high-fidelity generation of subsequent ship dynamic radiated noise.

[0170] Step 3: Setting feature fusion weights based on cavitation state

[0171] 1. Cavitation critical speed ( , Pre-calculation

[0172] The classification of cavitation states is based on two key speed thresholds: cavitation initiation speed and cavitation velocity. (The critical speed at which propeller cavitation begins) and the cavitation saturation speed (The critical speed at which propeller cavitation saturates and noise characteristics stabilize). These two thresholds are pre-calculated using ship basic parameters combined with optimized hydrodynamic simulation formulas. The calculation process is automatically completed by the simulation system without manual intervention. The input parameters are all from the ship basic parameters configured in step 1, including propeller diameter D (unit: m), number of blades B, operating advance coefficient J_P, initial cavitation number σ_i (preset range 0.001-0.01), wake fraction ω (preset range 0.1-0.3), ship design depth H (unit: m), and displacement tonnage DT (unit: t), ensuring that the calculation results are consistent with the parameter system for subsequent noise generation.

[0173] The formula for calculating the initial velocity of cavitation v_Si (unit: m / s) is:

[0174] ,

[0175] The formula logic is that the navigation depth H determines the water pressure (the higher the pressure, the less likely cavitation will occur, and the higher v_Si will be), the initial cavitation number σ_i reflects the propeller's anti-cavitation capability (the smaller σ_i is, the easier cavitation will occur, and the lower v_Si will be), and the wake fraction ω affects the actual effective rotational speed of the propeller. The larger the value, the higher the value of v_Si).

[0176] The formula for calculating the cavitation saturation speed v_SY (unit: m / s) is:

[0177] ,

[0178] Compared to v_Si, v_SY does not depend on the initial cavitation number σ_i, but is determined only by the depth and advance coefficient, reflecting the limiting speed at which cavitation develops to saturation. When the speed exceeds v_SY, the cavitation coverage area on the propeller surface no longer increases, and the continuous spectrum noise energy tends to stabilize.

[0179] 2. Cavitation state determination rules based on flight speed

[0180] The simulation system automatically determines the cavitation state based on the numerical relationship between the target speed v and v_Si and v_SY: v≤v_Si indicates no cavitation (no cavitation is generated on the propeller surface, and mechanical noise dominates); v_Si<v<v_SY indicates cavitation development (the propeller cavitation gradually diffuses from its generation, and the continuous spectrum energy continuously increases); v≥v_SY indicates cavitation saturation (the propeller surface is completely covered by cavitation, and the continuous spectrum energy reaches its peak and stabilizes). The division rules are clear and unique, with no ambiguity.

[0181] 3. Weighting of different cavitation states

[0182] For the three cavitation states, a fixed weight ratio is preset (W_l is the line spectrum weight, W_g is the continuous spectrum weight, and W_m is the modulation envelope weight). The sum of all weight values ​​is 1 to ensure energy conservation. The setting of each weight is based on the energy ratio of the noise component in the corresponding state.

[0183] In the non-cavitation state (v ≤ v_Si), W_l = 0.4, W_g = 0.3, and W_m = 0.3 are set. In this state, the discrete line spectrum generated by mechanical vibration is the core noise source and is given the highest weight to highlight its energy dominance. The continuous spectrum is mainly composed of hydrodynamic noise with weak energy and no obvious modulation effect. Both are given the same medium weight. In the cavitation development state (v_Si < v < v_SY), W_l = 0.3, W_g = 0.4, and W_m = 0.3 are set. At this time, the cavitation noise energy rises rapidly, the weight of the continuous spectrum exceeds that of the line spectrum, and the weight of the line spectrum decreases accordingly. The enhancement of the modulation effect is limited, and the original weight is maintained. In the cavitation saturation state (v ≥ v_SY), W_l = 0.2, W_g = 0.45, and W_m = 0.35 are set. The energy of the continuous spectrum reaches its maximum value and becomes absolutely dominant, so it is given the highest weight. The relative proportion of the line spectrum drops to the minimum. The modulation effect of the high-speed rotation of the propeller is more significant, and the modulation envelope weight is appropriately increased to make the rhythm of the signal more in line with the actual auditory experience during high-speed navigation.

[0184] The preset weighting ratios will be directly used as fixed coefficients for subsequent signal synthesis, integrated into the signal synthesis formula in step 6. The line spectrum component l(t), continuous spectrum component g(t), and modulation envelope component m(t) are multiplied by their corresponding weights and then fused according to the formula. The application of weights is automatically executed by the simulation system throughout the process, without manual intervention from the user. After determining the cavitation state of the target speed, the system directly calls the corresponding weighting and recombination, ensuring that the fusion process is efficient and accurate. Furthermore, the fixed weighting design avoids complex real-time calculations and reduces the computing power requirements for embedded deployment of self-propelled acoustic decoys.

[0185] Step 4: GAN-driven modulation envelope generation

[0186] The modulation envelope of ship radiated noise is essentially the periodic modulation of the continuous spectrum by propeller rotation. Its characteristics (modulation degree, pulse amplitude, etc.) are strongly correlated with ship speed and cavitation state: the higher the ship speed, the higher the modulation frequency (related to the propeller shaft frequency); different cavitation states also result in differences in modulation depth and pulse distribution patterns. Traditional methods, relying on statistical characteristics as a "black box," cannot accurately capture this correlation. However, GANs, through adversarial training between the generator and discriminator, can automatically learn the inherent patterns of "ship speed-cavitation state-modulation features" in simulation data. The generated modulation parameters are more consistent with the real distribution, and no manual intervention in feature design is required, resulting in stronger adaptability.

[0187] 1. Core Design Logic of GAN Networks

[0188] The GAN network consists of a generator and a discriminator, both of which adopt a lightweight fully connected network structure, balancing prediction accuracy and engineering deployment efficiency. The core function of the generator is to receive key input parameters and output the core parameters of the modulation envelope for each frequency band. The input parameters focus on key factors affecting the modulation characteristics, including the target speed v (unit: m / s, the value after standardization in step 2), the propeller shaft frequency f_z (unit: Hz, calculated by relating the target speed and the propeller speed, f_z=v×B / (2πR), where B is the number of blades and R is the propeller radius), and cavitation status indicators (0=no cavitation, 1=cavitation development, 2=cavitation saturation, directly input from the determination result in step 3). Its network structure adopts a 3-layer fully connected network. The number of neurons in the input layer is 3 (consistent with the dimension of the input parameters), the number of neurons in the first hidden layer is 128, the number of neurons in the second hidden layer is 64, and the activation function is ReLU. The number of neurons in the output layer is 84 (28 frequency bands × 3 parameters: modulation, pulse amplitude, and pulse width), and the activation function is Sigmoid. The output parameters are mapped to the [0,1] interval for easy normalization and restoration. The output parameters are output one by one according to the 28 frequency bands of the auditory filter bank in step 1. Each frequency band corresponds to a set of independent parameters to ensure fine modulation of frequency bands.

[0189] 2. Discriminator Structure and True / False Determination Logic

[0190] The core function of the discriminator is to distinguish between the "modulation envelope signal output by the generator" and the "real modulation envelope signal generated by the simulation system". The generator is optimized through reverse feedback. The input signal is the modulation envelope signal m_gen(t) synthesized by the modulation parameters output by the generator through a Gaussian pulse train model, or the real modulation envelope signal m_real(t) extracted by demodulating the noise time-domain signal generated by the simulation system. Both types of signals are input in 28 frequency bands. The length of a single frequency band signal is 10s × 524288Hz = 5.24288 × 10^6 sampling points. Its network structure adopts a 3-layer fully connected network, which is symmetrical with the generator structure to balance training efficiency. The number of neurons in the input layer is 5.24288×10^6 (number of sampling points of single-band signal), the number of neurons in the first hidden layer is 64, the number of neurons in the second hidden layer is 32, and the activation function is LeakyReLU (to avoid gradient vanishing in the negative region of ReLU). The number of neurons in the output layer is 1, the activation function is Sigmoid, and the output value is ∈[0,1]. It is used to determine whether the input signal is a "real signal" (output≈1) or a "generated signal" (output≈0). By "determining the authenticity" of the generated signal, gradient information is passed to the generator, forcing the generator to continuously optimize the parameter generation logic.

[0191] 3. GAN Network Adversarial Training Process and Convergence Criterion

[0192] The entire GAN training process is based on the simulated dataset of "speed-noise features" generated in step 1. A supervised adversarial training mode is used. Real modulation envelope signals m_real(t) for each speed and cavitation state are extracted from the simulated dataset and paired with the corresponding input parameters to form a training sample set of "input parameters-real modulation signals," which is divided into training and validation sets in an 8:2 ratio. Training parameters are configured as follows: batch size = 16, maximum training epochs = 150, initial learning rate = 0.0001 (consistent for generator and discriminator). The optimizer is Adam (β1 = 0.5, β2 = 0.999), and the loss function is binary cross-entropy (BCE). The generator loss function focuses on "minimizing the discriminator's recognition accuracy," while the discriminator loss function focuses on "maximizing the distinguishability between real and generated signals." The iterative training adopts an alternating pattern of "one round of generator training → one round of discriminator training". After each round of training, the performance is evaluated using a validation set: the generator receives batch input parameters, outputs modulation parameters, and synthesizes m_gen(t). The discriminator receives m_gen(t) and m_real(t) respectively, outputs the judgment result, calculates the loss value, updates the discriminator weights through backpropagation, and then updates the generator weights by fixing the discriminator to reduce the probability of being recognized. When the discriminator's accuracy in judging both generated and real signals in the validation set is close to 50% (±3%), and the generator loss value and discriminator loss value tend to stabilize (change rate ≤ 0.001%) in 10 consecutive rounds of training, the model training is considered converged, and the generator network parameters are saved (the discriminator is only used for training and does not need to be loaded during the inference phase).

[0193] 4. Implementation of Modulation Envelope Signal Synthesis

[0194] After model training converges, the generator parameters are loaded, and the target operating condition's speed, shaft frequency, and cavitation status indicators are input to generate modulation parameters for 28 frequency bands. Then, the modulation envelope signal m_i(t) for each frequency band (i=1~28, corresponding to the 28 frequency bands) is synthesized using a Gaussian pulse train model. The modulation envelope signal for each frequency band is a superposition of multiple Gaussian pulses; the core formula is:

[0195]

[0196] Where B is the number of propeller blades (the ship's basic parameters preset in step 1), T is the propeller rotation period (T=1 / f_z, f_z is the propeller shaft frequency), and μ_ξj(t) is a single Gaussian pulse signal ( ξ_j is the pulse amplitude output by the generator, σ is the pulse width output by the generator, and m_p-i is the modulation index of the generator output for that frequency band. The synthesis process is performed band by band. The pulse triggering time of each frequency band is determined by the propeller rotation period and the number of blades (k is the number of rotations, j is the blade index), ensuring that the periodicity of the modulation envelope is consistent with the rotation law of the ship's propeller. The synthesized m_i(t) is a continuous time-domain signal with a length of 10s and a sampling frequency of 524288Hz, which can be directly used for subsequent signal synthesis.

[0197] Step 5: Simulation of line spectrum instability based on simulation patterns

[0198] The fluctuation parameters (amplitude fluctuation variance, low-pass cutoff frequency and variance of frequency drift) are all statistically determined based on the "speed-noise characteristics" simulation dataset generated in step 1. The speed is divided into three intervals according to the cavitation state (consistent with step 3). For the simulation line spectrum data of each interval, the standard deviation of the line spectrum amplitude (as amplitude fluctuation variance ξ), the mean of the low-pass cutoff frequency of the line spectrum frequency drift (as f0) and the mean of the variance (as σ²) are extracted. The statistically obtained mean is used as the fixed fluctuation parameter of the interval and stored in the parameter library of the simulation system. It can be directly called according to the interval to which the target speed belongs in the future, without the need for real-time statistics.

[0199] 1. Simulation of line spectrum amplitude fluctuations

[0200] Based on preset fluctuation parameters, the baseline amplitude A_n' and baseline frequency f_n' predicted in step 2 are perturbed to generate unstable line spectrum parameters. The core of the line spectrum amplitude fluctuation simulation is to make the line spectrum amplitude fluctuate randomly around the baseline value to simulate the small changes in mechanical vibration. The core formula is:

[0201]

[0202] Where A_n(t) is the actual amplitude of the line spectrum at time t (unit: Pa), A_n' is the baseline amplitude of the line spectrum predicted by the CNN-LSTM model in step 2 (unit: Pa), and ξ is a uniformly distributed random quantity with a value range of [-√3σ_A, √3σ_A] (σ_A is the variance of amplitude fluctuation in this speed range, which is determined by statistical analysis of simulation data). The uniform distribution can ensure the randomness and stability of amplitude fluctuations and avoid signal distortion caused by too many extreme values. In implementation, at each sampling time, a random quantity ξ that conforms to the above uniform distribution is generated by the pseudo-random number generator built into the simulation system (with a fixed seed to ensure reproducibility), and superimposed with the baseline amplitude A_n' to obtain the actual amplitude at that time.

[0203] 2. Simulation of line spectrum frequency drift

[0204] The core of line spectrum frequency drift simulation is to allow the line spectrum frequency to drift slowly, simulating the minute fluctuations in propeller speed. This is achieved using a slowly changing Gaussian process, and the core formula is: Where f_n(t) is the actual frequency of the line spectrum at time t (in Hz), f_n' is the reference frequency of the line spectrum predicted by the CNN-LSTM model in step 2 (in Hz), and Δf_n(t) is the frequency shift at time t (in Hz), derived from the recursive formula.

[0205]

[0206] Among the parameters, Δf_n(t-1) is the frequency drift at time t-1 (initial value Δf_n(0)=0), f0 is the low-pass cutoff frequency (unit: Hz, determined by simulation data statistics, controlling the speed of frequency drift), f_s is the sampling frequency (fixed at 524288Hz, consistent with step 1), σ² is the variance of frequency drift (unit: Hz², determined by simulation data statistics, controlling the drift amplitude), and N(0,1) is a standard normal distribution random quantity; in implementation, Δf_n(t) is calculated point by point according to the sampling time, and superimposed with the reference frequency f_n' to obtain the actual frequency.

[0207] Step Six: Signal Synthesis and Output

[0208] Signal synthesis employs formulas consistent with the noise generation mechanism:

[0209]

[0210] Where W_l×l(t) is the weighting term for the line spectrum component, W_l is the line spectrum weight set in step 3, used to adjust the energy proportion of the line spectrum in the total signal. The larger the weight, the more obvious the line spectrum characteristics. W_g×g(t) is the weighting term for the continuous spectrum component, W_g is the continuous spectrum weight, which adjusts the basic energy proportion of the continuous spectrum. (1+W_m×m(t)) is the modulation term, used to perform amplitude modulation on the continuous spectrum. m(t) is the total modulation envelope signal, W_m is the modulation envelope weight that controls the modulation intensity. "1+" ensures that the modulated continuous spectrum will not have a negative amplitude (avoiding signal distortion). The modulation depth is determined by W_m×m(t). The overall logic is to first change the amplitude of the continuous spectrum through the modulation term, and then superimpose the weighted line spectrum with the modulated continuous spectrum to form the final signal, which perfectly matches the composition law of real ship noise.

[0211] The synthesized ship dynamic radiated noise signal must meet the hardware output requirements of the self-propelled acoustic decoy, covering a frequency range of 20Hz~40kHz, a duration of 10s, a sampling frequency of 524288Hz, an amplitude unit of Pa, and an output format of binary raw data (.raw format) or WAV format (supporting mainstream audio processing software to open), which can be directly transmitted to the sound-generating device of the self-propelled acoustic decoy. For each output signal, the system automatically calculates the time domain characteristics (root mean square, kurtosis), frequency domain characteristics (spectral peak frequency, spectral peak level), and auditory domain characteristics (MFCC coefficients, Moore loudness), and compares them with the simulation data characteristics of the corresponding operating conditions, requiring a relative error ≤5%.

Claims

1. A method for generating highly realistic dynamic radiated noise of ships, characterized in that, Includes the following steps: Step 1: Construct a two-dimensional core dataset of "speed-noise characteristics" using a ship dynamics simulation system; Step 2: Train a CNN-LSTM hybrid model based on the generated two-dimensional core dataset of "speed-noise features"; The CNN-LSTM hybrid model takes the target speed parameter as input and outputs the core quantization parameters of the ship's radiated noise. The core quantization parameters include the reference amplitude of the line spectrum, the reference frequency of the line spectrum, the peak frequency of the continuous spectrum, the peak level of the continuous spectrum, and the coefficients of the spectrum variation curve. Step 3: Based on the input target speed, combined with the pre-calculated initial cavitation speed and saturation cavitation speed, the target speed is divided into three cavitation states: no cavitation, cavitation development, and cavitation saturation. Based on the cavitation state, a fixed weight ratio is preset for the line spectrum component, continuous spectrum component, and modulation envelope component. Step 4: Based on the target speed and cavitation status indicators in Step 3, combined with the propeller shaft frequency, it is used as the input of the lightweight generative adversarial network (GAN). The GAN outputs the core parameters of the modulation envelope in 28 frequency bands through the generator, including modulation degree, pulse amplitude and pulse width, and then synthesizes the modulation envelope signal through the Gaussian pulse train model. Step 5: Based on the "speed-noise characteristics" simulation dataset in Step 1, statistically determine the variance of the line spectrum amplitude fluctuation and the low-pass cutoff frequency and variance of the frequency drift within each speed range according to the cavitation state, and use them as fixed fluctuation parameters; based on the cavitation state determined in Step 3, call the corresponding fixed fluctuation parameters to perform perturbation simulation on the line spectrum reference amplitude and line spectrum reference frequency predicted in Step 2, and generate unstable line spectrum parameters, including the time-varying actual line spectrum amplitude and actual line spectrum frequency. Step 6, based on the line spectrum weights determined in Step 3 Continuous spectrum weights and modulation envelope weights The modulation envelope signal generated in step 4 The unstable line spectrum parameters generated in step 5 are used to generate the final ship dynamic radiated noise signal. : in, These are line spectral components constructed based on unstable line spectral parameters. The continuous spectrum components are reconstructed based on the continuous spectrum parameters output in step 2.

2. The method for generating high-fidelity dynamic radiated noise of ships according to claim 1, characterized in that... The ship dynamics simulation system outputs time-domain signals based on the full-link logic of "noise source modeling - propagation process simulation - time-domain signal synthesis". Among them, mechanical noise adopts a multi-degree-of-freedom vibration simulation model, propeller noise cavitation state adopts blade fluid impact theory or bubble dynamics equation, and hydrodynamic noise adopts Reynolds number correlation model. The hull conduction simulation is attenuated by 0-15dB, and the seawater medium attenuation is attenuated by frequency domain correlation. The simulation input parameters are set using a two-dimensional configuration mode of "basic parameters + operating condition parameters". The speed range is set according to the principle of "no cavitation - cavitation development - cavitation saturation" with a minimum speed interval of 0.1 m / s. The speed range with no cavitation is set to 3-5 levels, the speed range with cavitation development is set to 5-8 levels, and the speed range with cavitation saturation is set to 3-5 levels. The initial speed of cavitation and the speed of cavitation saturation are automatically calculated by the fluid dynamics formula built into the simulation system. The built-in feature extraction engine extracts the following from the time-domain signal: line spectrum frequency, line spectrum amplitude, peak frequency, peak level, 1 / 3 octave band sound pressure level, axis frequency and first eight harmonics of the DEMON spectrum, 12th order MFCC coefficients, total loudness peak value and distribution variance of Moore characteristic loudness, forming a structured two-dimensional dataset.

3. The method for generating high-fidelity dynamic radiated noise of a ship according to claim 1, characterized in that, The CNN-LSTM hybrid model structure in step 2 is an end-to-end structure of "input layer → CNN layer → LSTM layer → fully connected output layer"; wherein: The input layer receives the target speed parameters and uses the Min-Max normalization algorithm to map the speed parameters to the [0,1] interval. The normalization algorithm formula is as follows: ， in, for ; for ; for ; ; Reflecting the normalized speed Shoot into the [0,1] interval, then convert it into a shape of [ , A three-dimensional tensor; where, This represents the number of training samples in a batch. Because the time was short, For feature dimensions; The CNN layer is configured with 32 one-dimensional convolutional kernels of size 3, stride 1, padding method 'same', activation function ReLU, and dropout=0.2, used to extract the local nonlinear correlation between speed and noise frequency domain features, and output tensor; The LSTM layer is set up with a two-layer serial structure. The first layer outputs the hidden state of each step and provides timing information for the second layer. The second layer outputs the hidden state of the last step and extracts the global temporal core features. Each layer has 64 hidden units and a dropout rate of 0.2, which is used to capture the temporal evolution pattern during the speed change process. The fully connected output layer has two layers. The first layer has 128 neurons and uses the ReLU activation function. The second layer has the same number of neurons as the core output parameters and uses the linear activation function. The output includes the reference amplitude of the line spectrum, the reference frequency of the line spectrum, the peak frequency of the continuous spectrum, the peak level of the continuous spectrum, and the coefficients of four spectral variation curves. The model training adopts supervised learning, using "speed" from the dataset in step 1 as the input sample and "line spectrum parameters + continuous spectrum parameters" as the label. The training set and validation set are divided in an 8:2 ratio. The batch size of the training parameters is 32, the initial learning rate is 0.001 and the ReduceLROnPlateau decay strategy is adopted. The optimizer is Adam, the loss function is MSE, and the convergence condition is that the MSE of the training set ≤ 1e-4 and the MSE of the validation set ≤ 1.5e-4, or the change rate of the MSE of the validation set ≤ 0.001% for 20 consecutive rounds.

4. The method for generating high-fidelity dynamic radiated noise of ships according to claim 1, characterized in that, In step 3: Initial speed of cavitation The calculation formula is: , Cavitation saturation speed The calculation formula is: J P For the operating advance coefficient, H represents the wake fraction, and H represents the navigation depth. For nascent vacuoles; The cavitation state determination rule is: v≤ It is a non-vacuolated state. <v< For cavitation development state, v≥ It is in a cavitation saturation state; The preset fixed weight ratio is as follows: No vacuolation state: =0.4, =0.3, =0.3; Cavitation development status: =0.3, =0.4, =0.3; cavitation saturation state: =0.2, =0.45, =0.

35.

5. The method for generating high-fidelity dynamic radiated noise of a ship according to claim 1, characterized in that, The lightweight GAN network in step 4 includes a generator and a discriminator; wherein: The generator uses a 3-layer fully connected network. The input layer has 3 neurons, corresponding to the target speed v, propeller shaft frequency f_z, and cavitation state flag. The first hidden layer has 128 neurons, the second hidden layer has 64 neurons, and the activation function is ReLU. The output layer has 84 neurons, the activation function is Sigmoid, and the output has 28 frequency bands, each with its own modulation, pulse amplitude, and pulse width. The discriminator is used in the training phase. Its input is the modulation envelope signal m_gen(t) synthesized by the modulation parameters output by the generator through the Gaussian pulse train model or the real modulation envelope signal m_real(t) extracted from the simulation dataset in step 1. Its output is the probability of judging whether it is true or false. The GAN training data comes from the dataset in step 1, and is divided into training and validation sets in an 8:2 ratio. The batch size is 16, the initial learning rate is 0.0001, the optimizer is Adam, and the loss function is binary cross-entropy. The generator and discriminator are trained alternately. The convergence condition is that the discriminator's accuracy in judging the generated signal and the real signal in the validation set is close to 50% ± 3%, and the loss change rate is ≤ 0.001% for 10 consecutive rounds. After training convergence, the generator is retained for inference. During inference, based on the cavitation state flag and target speed input in step 3, the generator outputs modulation parameters, which are then synthesized into a modulated envelope signal using a Gaussian pulse train model. The formula is: Where B is the number of propeller blades and T is the propeller rotation period; the synthesized... The signals are combined into a total modulation envelope signal and output to step 6.

6. The method for generating high-fidelity dynamic radiated noise of a ship according to claim 1, characterized in that, In step 5 The formula for simulating line spectrum amplitude fluctuations is: , in, The reference amplitude of the line spectrum output in step 2. The random variable is uniformly distributed, with a value range of [-√3σ_A, √3σ_A], where σ_A is the variance of amplitude fluctuations obtained statistically from the dataset in step 1 based on the cavitation state determined in step 3; each sampling time is generated independently by a pseudo-random number generator. ; The line spectrum frequency drift simulation uses a slowly changing Gaussian process: ， in, Let be the actual frequency of the line spectrum at time t. The reference frequency of the line spectrum output in step 2. Let be the frequency drift at time t, and the recursive formula is: , in and These are the low-pass cutoff frequency and frequency drift variance, obtained statistically from the dataset in step 1 based on the cavitation state. Let N(0,1) be the sampling frequency, N(0,1) be a standard normally distributed random variable, and Δ be the initial value. (0)=0.

7. The method for generating high-fidelity dynamic radiated noise of a ship according to claim 1, characterized in that, After step 6, verification is performed, and the verification process is as follows: In step 1, the simulation system calculates the time-domain, frequency-domain, and auditory-domain features of the synthesized signal and compares them with the features of the corresponding working conditions in the dataset of step 1, requiring a relative error of ≤5%. The time-domain features include root mean square and kurtosis, the frequency-domain features include spectral peak frequency and spectral peak level, and the auditory-domain features include MFCC coefficients and Moore loudness.

Images