A method and system for hydrophone motion self-noise suppression

Abstract

This invention provides a method and system for suppressing motion-induced self-noise in hydrophones. Using time-frequency data of the hydrophone's displacement velocity as input, it directly establishes a dynamic mapping relationship between the time-frequency characteristics of self-noise, overcoming the limitations of traditional mechanistic models. This model can uncover the deep nonlinear correlation between motion parameters and noise signals, generating high-precision, full-band acoustic features, providing highly adaptive and reliable data support for subsequent noise suppression. This invention, employing the aforementioned method and system for suppressing motion-induced self-noise in hydrophones, aims to address the core problem of insufficient generalization of traditional empirical formulas in random wave environments.

Description

Technical Field

[0001] This invention relates to the fields of acoustic signal processing and neural network prediction technology, and in particular to a method and system for suppressing motion noise in hydrophones. Background Technology

[0002] Sonar buoys, as crucial equipment for underwater target detection, possess the capabilities to search for, locate, and identify underwater targets, making them ideal for large-scale joint detection. However, in practical applications, their detection performance is significantly affected by self-noise interference. Self-noise primarily originates from the vibration of the sonar buoy's own mechanical structure and the motion of the hydrophone caused by sea waves. Traditionally, optimizing the sonar buoy structure to achieve mechanical decoupling between the hydrophone and the surface float reduces vertical motion caused by sea waves, thereby lowering self-noise. However, with the gradual improvement of structural optimization, the complex characteristics of self-noise remain difficult to completely suppress, especially under the excitation of random sea waves, where the motion state of the sonar buoy and the self-noise it generates exhibit significant time-frequency nonlinear characteristics.

[0003] In recent years, deep learning technology, especially neural networks, has demonstrated significant advantages in modeling and predicting complex nonlinear time-series data. Based on this, using neural networks to learn the time-frequency data of hydrophone motion velocity and establish a mapping relationship between it and the corresponding time-frequency characteristics of self-noise has become an effective method for self-noise prediction and suppression. Through real-time monitoring and data-driven modeling of the motion state of sonar buoy hydrophones, accurate prediction of the time-frequency characteristics of self-noise can be achieved, providing a theoretical basis and technical support for subsequent signal denoising and detection accuracy improvement. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for suppressing self-noise in hydrophone motion, aiming to solve the core problem of insufficient generalization of traditional empirical formulas in random wave environments. By using a deep learning algorithm—the PatchMixer neural network—and taking the time-frequency data of the hydrophone's displacement velocity as input, a dynamic mapping relationship of the time-frequency characteristics of self-noise is directly established, overcoming the limitations of traditional mechanistic models. This model can uncover the deep nonlinear correlation between motion parameters and noise signals, generating high-precision, full-band acoustic features, providing highly adaptive and reliable data support for subsequent noise suppression.

[0005] To achieve the above objectives, the present invention provides a method for suppressing motion self-noise in a hydrophone, comprising the following steps:

[0006] Acoustic data of sonar buoys are collected by hydrophones, and feature extraction is performed on the acoustic data to obtain motion wave velocity features and self-noise features. Time-series labels are added to the motion wave velocity features and self-noise features.

[0007] The motion wave velocity features and self-noise features with time-series labels are transformed in the time-frequency domain. Then, the motion wave velocity features and self-noise features with time-series labels are divided into blocks to obtain short-series time-frequency spectrum features under the new timestamp.

[0008] The short-sequence spectral features under the new timestamp are input into the self-noise prediction model for prediction, and the predicted self-noise time spectrum is obtained.

[0009] By performing a differential operation between the predicted self-noise time-frequency plot and the original noisy time-frequency plot, the self-noise component in the hydrophone is removed, resulting in a new signal time-frequency plot.

[0010] Preferably, the motion wave velocity features with time-series labels and the self-noise features with time-series labels are subjected to time-frequency domain transformation, using short-time Fourier transform:

[0011] ;

[0012] In the formula, Represents the short-time Fourier transform of the input signal. For input data, w Represents the window function. m For time-shift index, k Indicates frequency index, n This represents the time-shift index within the window. N Indicates the window length. j Indicates the imaginary part.

[0013] Preferably, the short-sequence spectral features under the new timestamp are input into the self-noise prediction model for prediction, and the predicted self-noise time spectrum is obtained, including the following steps:

[0014] After normalizing the short-sequence spectral features under the new timestamp, the data is segmented to obtain multiple local segments, thereby capturing local features and patterns in the time series.

[0015] The segmented local fragments are converted into feature vectors through an embedding layer;

[0016] Two-headed prediction using eigenvectors, including linear and nonlinear prediction;

[0017] The predicted self-noise time spectrum is obtained by concatenating the linear and nonlinear prediction results:

[0018] ;

[0019] This represents the predicted self-noise frequency spectrum. This represents the linear prediction result. This represents the result of a nonlinear prediction.

[0020] Preferably, nonlinear prediction specifically includes:

[0021] The first feature vector is modeled separately by depthwise separable convolution to model the information inside and between frequency components. The modeling results are then subjected to non-linear learning using an activation function. After batch normalization, the results are concatenated with the residual connection to obtain the second feature vector.

[0022] The second feature vector is convolved pointwise to adjust its dimension, and key frequency components are extracted. After nonlinear learning and batch normalization of the key frequency components, the third feature vector is obtained.

[0023] The third eigenvector is flattened using a fully connected process to obtain the nonlinear prediction result.

[0024] Preferably, during the training of the self-noise prediction model, a supervised learning strategy is adopted, using motion wave velocity features with time-series labels as input. X Self-noise features with time-series labels are used as true labels. Y The loss function is the mean squared error, defined as:

[0025] ;

[0026] in, , These are the actual and predicted time-spectral amplitude values, respectively.

[0027] A hydrophone motion self-noise suppression system, comprising:

[0028] The data processing module is used to collect acoustic data of sonar buoys through hydrophones, extract features from the acoustic data to obtain motion wave velocity features and self-noise features, and add time-series labels to the motion wave velocity features and self-noise features.

[0029] The segmentation module is used to perform time-frequency domain transformation on the motion wave velocity features and self-noise features with time-series labels, and then segment the motion wave velocity features and self-noise features with time-series labels into blocks to obtain short-series time-frequency spectrum features under the new timestamp.

[0030] The prediction module is used to input the short-sequence spectral features under the new timestamp into the self-noise prediction model for prediction, and obtain the predicted self-noise time spectrum;

[0031] The self-noise removal module is used to perform a differential operation between the predicted self-noise time-frequency map and the original noisy time-frequency map to remove the self-noise component from the hydrophone and obtain a new signal time-frequency map.

[0032] Therefore, the present invention employs the above-mentioned method and system for suppressing motion noise in a hydrophone, and the technical effects are as follows:

[0033] The key advantages of this invention are: it uses short-time Fourier transform to obtain time-frequency data of hydrophone motion velocity, models each frequency component separately, and uses depthwise separable convolution modules to model the information within and between frequency components, which helps to simultaneously capture high-frequency local disturbances and low-frequency periodic noise patterns, thereby improving prediction accuracy.

[0034] This invention utilizes PatchMixer's ability to capture complex spectral feature changes and a dual-prediction head design to capture complex spectral feature changes. The output structure can identify trends and high-frequency disturbances separately, and predict the time-frequency data of self-noise generated by the forced movement of the hydrophone due to wave excitation. Attached Figure Description

[0035] Figure 1 This is a time-frequency diagram of a hydrophone containing self-noise signals.

[0036] Figure 2 This is an architecture diagram of the self-noise prediction model of the present invention;

[0037] Figure 3 Time-frequency diagram for predicting self-noise of a hydrophone;

[0038] Figure 4 This is a time-frequency diagram of self-noise suppression in a hydrophone based on the PatchMixer neural network. Detailed Implementation

[0039] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0040] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0041] Example 1

[0042] like Figure 2 As shown, a method for suppressing motion self-noise in a hydrophone includes the following steps:

[0043] Step 1: Collect acoustic data of the sonar buoy using a hydrophone, extract features from the acoustic data to obtain motion wave velocity features and self-noise features, and add time-series labels to the motion wave velocity features and self-noise features.

[0044] Excited by sea waves, acoustic data from a sonar buoy is collected via a hydrophone to obtain the raw signal. The signal record includes corresponding time-domain data and the vertical motion wave velocity data of the hydrophone. At this point, the obtained raw signal contains the hydrophone's motion velocity data as well as interference from self-noise.

[0045] Step 2: Perform time-frequency domain transformation on the motion wave velocity features with time-series labels and the self-noise features with time-series labels. Then, cut the motion wave velocity features with time-series labels and the self-noise features with time-series labels into blocks to obtain short-series time-frequency spectrum features under the new timestamp.

[0046] The original hydrophone vertical velocity time-domain signal and self-noise data are first converted to the time-frequency domain via Short Time Fourier Transform (STFT). STFT divides the long time series into multiple short time windows, and performs a Fourier Transform on each window to obtain the signal's spectrum at each time point. The time-frequency diagram containing self-noise is shown below. Figure 1 As shown.

[0047] ;

[0048] In the formula, Represents the short-time Fourier transform of the input signal. For input data, w Represents the window function. m For time-shift index, k Indicates frequency index, n This represents the time-shift index within the window. N Indicates the window length. j Indicates the imaginary part.

[0049] The transformation results in a time-frequency spectrum, which reflects the spectral distribution characteristics of the velocity signal as it changes over time.

[0050] Step 3: Input the short-sequence spectral features under the new timestamp into the self-noise prediction model for prediction, and obtain the predicted self-noise time spectrum, including the following steps:

[0051] After normalizing the short-sequence spectral features under the new timestamp, the data is segmented to obtain multiple local segments, thereby capturing local features and patterns in the time series.

[0052] The original time series data is divided into training and test sets, and time series data for the training and test sets are generated. The time series data for the training and test sets are then preprocessed by normalization.

[0053] The resulting time-frequency amplitude spectrum typically has a wide range of values, requiring normalization preprocessing before inputting it into the network. This is achieved by linearly scaling the spectrum values ​​to a fixed interval and then using min-max normalization.

[0054] The segmented local fragments are converted into feature vectors through an embedding layer, including:

[0055] The processed training and test sets are input into the PatchMixer model. The PatchMixer model is used to model the hydrophone's motion velocity data, capturing long-term trends and short-term changes in the data, and outputting prediction results. The PatchMixer network prediction structure diagram is shown below. Figure 2 As shown.

[0056] PatchMixer divides the input time-frequency data into multiple patch segments (e.g., P×F, where P is the time length and F is the number of frequency channels), inputs them into a depthwise separable convolutional module, and performs spatial convolution on each channel individually; the convolutions then fuse the features between the channels. Two prediction heads are used for prediction:

[0057] Two-headed prediction using eigenvectors, including linear and nonlinear prediction;

[0058] Nonlinear prediction, specifically including:

[0059] The first feature vector is modeled separately by depthwise separable convolution to model the information inside and between frequency components. The modeling results are then subjected to non-linear learning using an activation function. After batch normalization, the results are concatenated with the residual connection to obtain the second feature vector.

[0060] The second feature vector is convolved pointwise to adjust its dimension, and key frequency components are extracted. After nonlinear learning and batch normalization of the key frequency components, the third feature vector is obtained.

[0061] The third eigenvector is flattened using a fully connected process to obtain the nonlinear prediction result.

[0062] MLP prediction head:

[0063] ;

[0064] in, h Encode the patch vector; , , , These are trainable parameters; σ ReLU is a non-linear activation function.

[0065] Linear prediction head:

[0066] ;

[0067] in, , These are trainable parameters;

[0068] The predicted self-noise time spectrum is obtained by concatenating the linear and nonlinear prediction results:

[0069] ;

[0070] This represents the predicted self-noise frequency spectrum. This represents the linear prediction result. This represents the result of a nonlinear prediction.

[0071] During the training of the self-noise prediction model, a supervised learning strategy is adopted, using motion wave velocity features with time-series labels as input. X Self-noise features with time-series labels are used as true labels. Y During training, the network outputs... As close as possible to the real label Y The commonly used loss function is the mean squared error (MSE), defined as:

[0072] ;

[0073] in, , These are the actual and predicted time-spectral amplitude values, respectively. i Represents a sequence index. t Indicates a time index. f This represents the frequency index. By optimizing this loss, the parameters of PatchMixer are gradually adjusted during training to optimize the output spectrum. With the target spectrum Minimal error.

[0074] Step 4: Self-Noise Prediction and Suppression. After training, the PatchMixer model can take in the time-spectrum data of the new hydrophone's motion velocity and output the predicted self-noise time-spectrum, such as... Figure 3 The time-frequency plots of the hydrophone's self-noise predictions are shown. These predictions can be used for further research on sonar buoy self-noise estimation and suppression, helping to improve the signal quality and accuracy of sonar buoys.

[0075] Step 5: Perform a difference operation between the self-noise time-frequency plot predicted by the neural network and the original noisy time-frequency plot. This will effectively remove the self-noise component from the hydrophone, resulting in a cleaner signal time-frequency plot, such as... Figure 4 The time-frequency diagram of self-noise suppression in hydrophones based on the PatchMixer neural network is shown. This significantly improves the quality of the target signal and substantially reduces the impact of self-noise on the hydrophone signal, thus providing clearer and more reliable data support for subsequent acoustic analysis or signal processing tasks.

[0076] Therefore, this invention employs the aforementioned method and system for suppressing self-noise in hydrophone motion, aiming to address the core problem of insufficient generalization of traditional empirical formulas in random wave environments. Through a deep learning algorithm—the PatchMixer neural network—it directly establishes a dynamic mapping relationship of self-noise time-frequency characteristics using time-frequency data of hydrophone displacement velocity as input, overcoming the limitations of traditional mechanistic models. This model can uncover the deep nonlinear correlation between motion parameters and noise signals, generating high-precision, full-band acoustic features, providing highly adaptive and reliable data support for subsequent noise suppression.

[0077] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method of hydrophone motion self-noise suppression, characterized by, Includes the following steps: Acoustic data of sonar buoys are collected by hydrophones, and feature extraction is performed on the acoustic data to obtain motion wave velocity features and self-noise features. Time-series labels are added to the motion wave velocity features and self-noise features. The motion wave velocity features and self-noise features with time-series labels are transformed in the time-frequency domain. Then, the motion wave velocity features and self-noise features with time-series labels are divided into blocks to obtain short-series time-frequency spectrum features under the new timestamp. The short-sequence spectral features under the new timestamp are input into the self-noise prediction model for prediction, and the predicted self-noise time spectrum is obtained. By performing a difference operation between the predicted self-noise time-frequency plot and the original noisy time-frequency plot, the self-noise component in the hydrophone is removed, resulting in a new signal time-frequency plot. The short-series time-series spectral features under the new timestamp are input into the self-noise prediction model for prediction, and the predicted self-noise time-series spectrum is obtained, including the following steps: After normalizing the short-sequence spectral features under the new timestamp, the data is segmented to obtain multiple local segments, thereby capturing local features and patterns in the time series. The segmented local fragments are converted into feature vectors through an embedding layer; Two-headed prediction using eigenvectors, including linear and nonlinear prediction; The predicted self-noise time spectrum is obtained by concatenating the linear and nonlinear prediction results: ； represents a predicted self-noise time-frequency spectrum, represents a linear prediction result, represents a non-linear prediction result; Nonlinear prediction, specifically including: The first feature vector is modeled by depthwise separable convolution to model the information within and between frequency components. The modeling results are then subjected to non-linear learning using an activation function. After batch normalization, the results are concatenated with the residual connection to obtain the second feature vector. The second feature vector is convolved pointwise to adjust its dimension, and key frequency components are extracted. After nonlinear learning and batch normalization of the key frequency components, the third feature vector is obtained. The third eigenvector is flattened using a fully connected process to obtain the nonlinear prediction result.

2. A method of hydrophone motion self-noise suppression according to claim 1, wherein, The time-labeled motion wave velocity characteristics and time-labeled self-noise characteristics are subjected to time-frequency domain transformation using the short-time Fourier transform: ； In the formula, Represents the short-time Fourier transform of the input signal. For input data, w Represents the window function. m For time-shift index, k Indicates frequency index, n This represents the time-shift index within the window. N Indicates the window length. j Indicates the imaginary part.

3. A method of hydrophone motion self-noise suppression according to claim 2, wherein, During the training of the self-noise prediction model, a supervised learning strategy is adopted, using motion wave velocity features with time-series labels as input. X Self-noise features with time-series labels are used as true labels. Y The loss function is the mean squared error, defined as: ； in, , These are the actual and predicted time-spectral amplitude values, respectively. i This represents a sequence index.

4. A hydrophone motion self-noise suppression system, characterized by, A method for suppressing motion-induced self-noise in a hydrophone as described in claim 1 includes: The data processing module is used to collect acoustic data of sonar buoys through hydrophones, extract features from the acoustic data to obtain motion wave velocity features and self-noise features, and add time-series labels to the motion wave velocity features and self-noise features. The segmentation module is used to perform time-frequency domain transformation on the motion wave velocity features and self-noise features with time-series labels, and then segment the motion wave velocity features and self-noise features with time-series labels into blocks to obtain short-series time-frequency spectrum features under the new timestamp. The prediction module is used to input the short-sequence spectral features under the new timestamp into the self-noise prediction model for prediction, and obtain the predicted self-noise time spectrum; The self-noise removal module is used to perform a differential operation between the predicted self-noise time-frequency map and the original noisy time-frequency map to remove the self-noise component from the hydrophone and obtain a new signal time-frequency map.

Images