Underwater sound target radiation noise system feature extraction method

Abstract

The invention relates to the related field of underwater acoustic signal processing, in particular to an underwater acoustic target radiation noise system feature extraction method, which comprises the following steps of: firstly, performing ICA (independent component analysis) processing on an original feature vector, extracting main independent components, and forming a dimensionality-reduced feature matrix; then, the most important feature subset is evaluated and selected through a feature selection method, and the feature dimension is further reduced; next, a proper weight is distributed for each feature, and the influence of important features is highlighted; according to the method, the dimensionality of the features is remarkably reduced, and the calculation efficiency and the model training speed are improved. Meanwhile, by removing noise and redundant information, the robustness and classification performance of the model are enhanced. The optimized feature vector is easier to explain and visualize, and is suitable for various underwater acoustic signal processing tasks, such as classification, identification and noise reduction. In general, the invention provides an efficient, flexible and powerful feature optimization scheme, and has wide application prospects and remarkable technical advantages.

Description

Technical Field

[0001] The present invention relates to the field of underwater acoustic signal processing, and specifically to a method for extracting the modulation characteristics of underwater target radiated noise. Background Technique

[0002] Underwater acoustic technology has important applications in the identification, tracking, and classification of underwater targets in military and civilian fields. Underwater targets such as submarines, surface ships, and underwater unmanned vehicles generate various noises when operating underwater. These noise signals contain the physical characteristic information of the targets. Therefore, by analyzing the radiated noise of underwater targets, the purpose of identifying the target type, state, etc. can be achieved. Existing methods for extracting underwater target characteristics usually rely on traditional signal processing techniques such as Fourier transform and wavelet transform. However, these methods perform poorly in processing non-stationary signals, especially in the case of complex noise backgrounds and weak target characteristics, and it is difficult to meet the requirements of high-precision identification. Therefore, it is very necessary to propose a new method for extracting the radiated noise characteristics of underwater targets. Summary of the Invention

[0003] The purpose of the present invention is to provide a method for extracting the modulation characteristics of underwater target radiated noise to solve the problems raised in the above background technique.

[0004] To achieve the above purpose, the present invention provides the following technical solution: A method for extracting the modulation characteristics of underwater target radiated noise, including the following steps:

[0005] Step 1, Signal preprocessing: Use a hydrophone array to collect the radiated noise signal of an underwater target; perform filtering and noise reduction processing on the collected signal to remove background noise and interference signals and improve the signal-to-noise ratio; then perform signal resampling to ensure the consistency and standardization of subsequent processing;

[0006] Step 2, Time-frequency analysis: Use the short-time Fourier transform to perform time-frequency analysis on the signal preprocessed in Step 1 to generate a time-frequency diagram;

[0007] Step 3, Feature extraction: Calculate the power spectral density in the time-frequency diagram in Step 2 to extract the spectral characteristics of the target radiated noise; perform envelope detection on the signal in the time-frequency diagram to extract the envelope characteristics of the target radiated noise; calculate the instantaneous frequency of the signal to extract the instantaneous frequency characteristics of the target radiated noise; use sparse coding technology to represent the extracted spectral characteristics, envelope characteristics, and instantaneous frequency characteristics as a group of sparse basis functions to further compress the feature dimension and improve the robustness of the features;

[0008] Step 4, Feature fusion: Fusion the various features extracted in Step 3 to form a comprehensive feature vector;

[0009] Step 5, Feature Dimensionality Reduction: Use independent component analysis to reduce the dimensionality of the comprehensive feature vector and extract the low-dimensional features that can best reflect the target features;

[0010] Step 6, Feature Optimization: Optimize the features after dimensionality reduction through machine learning methods to further improve the recognition performance of the features;

[0011] Step 7, Target Recognition: Use the optimized feature vector for target recognition.

[0012] Preferably, the specific content of the signal preprocessing in step 1 is as follows:

[0013] Signal Acquisition: In an underwater environment, signal acquisition is carried out using a hydrophone array, and the hydrophone array is one or more of one-dimensional, two-dimensional, or three-dimensional;

[0014] Filtering: According to the radiation noise characteristics of the underwater target, determine the low cut-off frequency and high cut-off frequency of a suitable band-pass filter, and then apply the designed band-pass filter to the acquired underwater acoustic signal to remove unwanted frequency components;

[0015] Noise Reduction: Use Kalman filtering to reduce the noise of the radiation noise signal of the underwater target. Specifically, define the state equation and observation equation of the system. For the underwater acoustic signal, assume that the state of the system is the amplitude and phase of the signal, and the observation equation is the acquired signal value; predict the state at the next moment according to the state equation, and update the state estimate according to the observation equation and observation data. Repeat the prediction step and the update step to gradually reduce the influence of noise;

[0016] Resampling: Select a preset target sampling frequency for the acquired radiation noise signal of the underwater target. According to the original sampling frequency and the target sampling frequency, calculate the interpolation coefficient, insert new sampling points between the sampling points of the original signal, and calculate the radiation noise signal value of the new underwater target through linear interpolation.

[0017] Preferably, the specific content of the time-frequency analysis in step 2 is as follows: Short-time Fourier Transform: Select a preset window function. The role of the window function is to segment the signal, that is, slide the window function in time to segment the preprocessed signal; perform Fourier transform on each segmented signal to obtain spectral information; arrange the spectral information of each segment in chronological order to form a time-frequency diagram.

[0018] Preferably, the specific content of the feature extraction in step 3 is as follows:

[0019] Power Spectral Density: Calculate the energy of each frequency component in the time-frequency diagram in step 2, specifically using the squared amplitude value, and then average the energies of all frequency components to obtain the power spectral density;

[0020] Envelope spectrum: Perform envelope detection on the signal of each time window in the time-frequency diagram in step 2. Specifically, use the square detection and Hilbert transform methods to obtain its envelope signal, and then perform Fourier transform on the envelope signal to obtain the envelope spectrum;

[0021] Instantaneous frequency: Perform Hilbert transform on the signal of each time window in the time-frequency diagram in step 2 to obtain its analytic signal, and calculate the phase change rate of the analytic signal to obtain the instantaneous frequency;

[0022] Sparse coding technology: Represent the extracted spectral features, envelope features, and instantaneous frequency features as a set of sparse basis functions, form a dictionary with the selected basis functions, and then solve the sparse representation of the signal on the dictionary through the LASSO optimization algorithm. Use the sparse representation coefficients as feature vectors for subsequent classification and recognition.

[0023] Preferably, the specific implementation steps of the comprehensive features in step 4 are as follows:

[0024] Step 41, Feature extraction: Extract the power spectral density, envelope spectrum, and instantaneous frequency features in step 3;

[0025] Step 42, Feature normalization: Perform normalization processing on the extracted features to ensure the consistency of the dimension and numerical range between different features. Specifically, convert each element of each feature vector in step 41 to the range [0, 1];

[0026] Step 43, Determine weights: Manually set the weights of each feature in step 41 according to domain knowledge and experience;

[0027] Step 44, Feature fusion: Combine the normalized feature vectors in step 42 with the feature weights determined in step 43 for weighted fusion to form a comprehensive feature vector.

[0028] Preferably, the specific steps of feature dimensionality reduction in step 5 are as follows:

[0029] Step 51, Select samples of multiple comprehensive feature vectors from step 4 to form a feature matrix X. Assume the size of X is M×N, where N is the number of samples and M is the dimension of the feature vector;

[0030] Step 52, Normalize the feature matrix X. The formula is: where μ(X) is the mean of the feature matrix, and σ(X) is the standard deviation of the feature matrix; Subsequently, further centralize the normalized feature matrix X′ so that its mean is 0. The specific formula is: X″ = X′ - μ(X′);

[0031] Step 53, Whiten the centralized feature matrix X″ so that its covariance matrix is the identity matrix: Xwhite = WX″, where W is the whitening matrix, obtained by eigenvalue decomposition or singular value decomposition;

[0032] Step 54. Use the ICA algorithm to perform independent component analysis on the whitened feature matrix X white to obtain the independent component matrix S and the mixing matrix A. The specific formula is: X white ≈ AS;

[0033] Step 55. Select the first k independent components from the independent component matrix S to form the feature matrix Y after dimensionality reduction: Y = S[:, :k], where the size of Y is N × k.

[0034] Preferably, the specific steps of feature optimization in step 6 are as follows:

[0035] Step 61. Data preparation: The dimensionality reduction process is performed on the comprehensive feature vector through independent component analysis to obtain a feature matrix Y after dimensionality reduction;

[0036] Step 62. By recursively removing the least important features in the feature matrix Y, gradually reduce the number of features to evaluate the importance of each feature.

[0037] Step 63. According to the evaluation results, select the most important feature subset;

[0038] Step 64. Use the selected feature subset to form a new feature matrix Y opt , that is, form the optimized feature matrix.

[0039] Compared with the prior art, the beneficial effects of the present invention are as follows: By comprehensively using independent component analysis (ICA), machine learning methods, and optimization algorithms, the present invention performs dimensionality reduction and optimization processing on the comprehensive feature vector, significantly improving the quality and efficiency of feature representation. This method can not only reduce the computational complexity and improve the model performance, but also enhance the robustness of the model, facilitate data interpretation and visualization, optimize data storage and transmission, and adapt to various application scenarios. Generally speaking, the present invention provides an efficient, flexible, and powerful feature optimization solution for underwater acoustic signal processing, with broad application prospects and significant technical advantages. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a schematic flowchart of the method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0041] Next, in combination with the accompanying drawings in the embodiments of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0042] Please refer to Figure 1 , the present invention provides a technical solution: a method for extracting the radiated noise signature characteristics of an underwater acoustic target, including the following steps:

[0043] Step 1, signal preprocessing: Use a hydrophone array to collect the radiated noise signal of an underwater target; perform filtering and noise reduction processing on the collected signal to remove background noise and interference signals and improve the signal-to-noise ratio; then perform signal resampling to ensure the consistency and standardization of subsequent processing;

[0044] The specific content of signal preprocessing is as follows:

[0045] Signal acquisition: In an underwater environment, signal acquisition is performed using a hydrophone array, and the hydrophone array is one or more of one-dimensional, two-dimensional, or three-dimensional;

[0046] Filtering: According to the radiated noise characteristics of the underwater acoustic target, determine the low cut-off frequency and high cut-off frequency of a suitable band-pass filter, and then apply the designed band-pass filter to the collected underwater acoustic signal to remove unnecessary frequency components;

[0047] Noise reduction: Use Kalman filtering to reduce the noise of the radiated noise signal of the underwater target. Specifically, define the state equation and observation equation of the system. For the underwater acoustic signal, assume that the state of the system is the amplitude and phase of the signal, and the observation equation is the collected signal value; predict the state at the next moment according to the state equation, and update the state estimate according to the observation equation and observation data, and repeat the prediction step and the update step to gradually reduce the influence of noise;

[0048] Resampling: Select a preset target sampling frequency for the collected radiated noise signal of the underwater target. According to the original sampling frequency and the target sampling frequency, calculate the interpolation coefficient, insert new sampling points between the sampling points of the original signal, and calculate the new radiated noise signal value of the underwater target through linear interpolation.

[0049] Step 2, time-frequency analysis: Perform time-frequency analysis on the preprocessed signal in Step 1 using the short-time Fourier transform to generate a time-frequency diagram;

[0050] The specific content of time-frequency analysis is as follows: Short-time Fourier transform: Select a preset window function. The role of the window function is to segment the signal, that is, slide the window function in time and segment the preprocessed signal; perform Fourier transform on each segmented signal to obtain spectral information; arrange the spectral information of each segment in chronological order to form a time-frequency diagram.

[0051] Step 3: Feature extraction: Calculate the power spectral density in the time-frequency diagram in Step 2 and extract the spectral features of the target radiated noise; extract the envelope features of the target radiated noise by performing envelope detection on the signal in the time-frequency diagram; calculate the instantaneous frequency of the signal and extract the instantaneous frequency features of the target radiated noise; use sparse coding technology to represent the extracted spectral features, envelope features, and instantaneous frequency features as a set of sparse basis functions, further compress the feature dimension, and improve the robustness of the features;

[0052] The specific content of feature extraction is as follows:

[0053] Power spectral density: Calculate the energy of each frequency component in the time-frequency diagram in Step 2. Specifically, use the squared amplitude, and then average the energies of all frequency components to obtain the power spectral density;

[0054] Envelope spectrum: Perform envelope detection on the signal of each time window in the time-frequency diagram in Step 2. Specifically, use the square detection and Hilbert transform methods to obtain its envelope signal, and then perform Fourier transform on the envelope signal to obtain the envelope spectrum;

[0055] Instantaneous frequency: Perform Hilbert transform on the signal of each time window in the time-frequency diagram in Step 2 to obtain its analytic signal, and calculate the phase change rate of the analytic signal to obtain the instantaneous frequency;

[0056] Sparse coding technology: Represent the extracted spectral features, envelope features, and instantaneous frequency features as a set of sparse basis functions, form a dictionary with the selected basis functions, and then solve the sparse representation of the signal on the dictionary through the LASSO optimization algorithm. Use the sparse representation coefficients as feature vectors for subsequent classification and recognition.

[0057] Step 4: Feature fusion: Fuse the multiple features extracted in Step 3 to form a comprehensive feature vector;

[0058] The specific implementation steps of the comprehensive feature are as follows:

[0059] Step 41: Feature extraction: Extract the power spectral density, envelope spectrum, and instantaneous frequency features in Step 3;

[0060] Step 42, Feature normalization: Normalize the extracted features to ensure the same dimension and numerical range among different features. Specifically, convert each element of each feature vector in Step 41 to the range [0, 1].

[0061] Step 43, Determine weights: Manually set the weights of each feature in Step 41 according to domain knowledge and experience.

[0062] Step 44, Feature fusion: Combine the normalized feature vectors in Step 42 with the feature weights determined in Step 43 for weighted fusion to form a comprehensive feature vector.

[0063] Step 5, Feature dimensionality reduction: Use independent component analysis to perform dimensionality reduction on the comprehensive feature vector and extract low-dimensional features that can best reflect the target features.

[0064] The specific steps of feature dimensionality reduction are as follows:

[0065] Step 51, Select samples of multiple comprehensive feature vectors from Step 4 to form a feature matrix X. Assume the size of X is M×N, where N is the number of samples and M is the dimension of the feature vector.

[0066] Step 52, Normalize the feature matrix X. The formula is: where μ(X) is the mean of the feature matrix and σ(X) is the standard deviation of the feature matrix. Subsequently, further centralize the normalized feature matrix X′ so that its mean is 0. The specific formula is: X″ = X′ - μ(X′).

[0067] Step 53, Whiten the centralized feature matrix X″ so that its covariance matrix is the identity matrix: X white = WX″, where W is the whitening matrix obtained through eigenvalue decomposition or singular value decomposition.

[0068] Step 54, Use the ICA algorithm to perform independent component analysis on the whitened feature matrix X white to obtain the independent component matrix S and the mixing matrix A. The specific formula is: X white ≈ AS;

[0069] Step 55, Select the first k independent components from the independent component matrix S to form the dimensionality-reduced feature matrix Y: Y = S[:, :k], where the size of Y is N×k.

[0070] Step 6, Feature optimization: Optimize the dimensionality-reduced features through machine learning methods to further improve the recognition performance of the features.

[0071] The specific steps of feature optimization are as follows:

[0072] Step 61, Data Preparation: The comprehensive feature vector is dimensionally reduced by independent component analysis to obtain a dimensionally reduced feature matrix Y;

[0073] Step 62, Gradually reduce the number of features by recursively removing the least important features in the feature matrix Y to evaluate the importance of each feature.

[0074] Step 63, Select the most important feature subset according to the evaluation results;

[0075] Step 64, Use the selected feature subset to form a new feature matrix Y opt , that is, form an optimized feature matrix.

[0076] Step 7, Target Recognition: Use the optimized feature vector for target recognition.

[0077] The present invention dimensionally reduces the comprehensive feature vector by using independent component analysis (ICA), and optimizes the dimensionally reduced features through machine learning methods and optimization algorithms, having the following remarkable beneficial effects:

[0078] Reduce computational complexity

[0079] Dimensionality reduction processing: By dimensionally reducing the high-dimensional feature vector, for example, reducing the 50-dimensional feature to 10 dimensions, the computational complexity in subsequent data processing and modeling can be significantly reduced. This is particularly important for large-scale data sets and can effectively improve the computational efficiency and model training speed; Feature optimization: Further feature optimization steps, such as feature selection and feature weighting, can further reduce the dimension of the features, making the model more concise and efficient.

[0080] Improve model performance

[0081] Feature selection: By selecting the most important feature subset and removing irrelevant or redundant features, the classification and recognition performance of the model can be improved. This helps to reduce overfitting and improve the generalization ability of the model; Feature weighting: By weighting the features to highlight the influence of important features, the representation of the feature vector can be further optimized, improving the accuracy and stability of the model;

[0082] Enhance robustness

[0083] Reduce the influence of noise: Dimensionality reduction processing and feature optimization can help remove noise and redundant information, reduce the sensitivity of the model to noise, and improve robustness; Improve data quality: Through feature selection and weighting, it can be ensured that the feature vector input to the model is more concentrated on important information, thereby improving the overall quality of the data;

[0084] Improve data storage and transmission efficiency

[0085] Reduce data volume: Through dimensionality reduction processing and feature optimization, the storage and transmission volume of data is significantly reduced, lowering the cost of data management; Optimize resource utilization: Lower feature dimensions mean that computing resources and storage space can be utilized more efficiently, especially in resource-constrained environments (such as embedded systems or mobile devices).

[0086] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for extracting characteristics of underwater acoustic target radiation noise, characterized in that: The following steps are involved: Step 1: Signal preprocessing: Use a hydrophone array to collect the radiation noise signal of the underwater target; filter and denoise the collected signal to remove background noise and interference signals and improve the signal-to-noise ratio; then resample the signal to ensure consistency and standardization of subsequent processing; Step 2, time-frequency analysis: Use short-time Fourier transform to perform time-frequency analysis on the signal preprocessed in step 1 to generate a time-frequency graph; Step 3, feature extraction: Calculate the power spectrum density in the time-frequency diagram in step 2 to extract the spectrum characteristics of the target radiation noise; Extract the envelope characteristics of the target radiation noise by performing envelope detection on the signal in the time-frequency diagram; Calculate the instantaneous frequency of the signal and extract the instantaneous frequency characteristics of the target radiation noise; Using sparse coding technology, the extracted spectrum features, envelope features and instantaneous frequency features are represented as a set of sparse basis functions to further compress the feature dimension and improve the robustness of the features. Step 4: Feature fusion: fuse the multiple features extracted in step 3 to form a comprehensive feature vector; Step 5: Feature dimensionality reduction: Use independent component analysis to reduce the dimensionality of the comprehensive feature vector and extract the low-dimensional features that best reflect the target features; Step 6: Feature optimization: Optimize the features after dimensionality reduction through machine learning methods to further improve the recognition performance of the features; Step 7: Target recognition: Use the optimized feature vector for target recognition.

2. The method for extracting characteristics of underwater acoustic target radiation noise according to claim 1, characterized in that: The specific contents of the signal preprocessing in step 1 are as follows: Signal acquisition: In an underwater environment, signal acquisition is performed using a hydrophone array, which is one or more of one, two or three dimensions; Filtering: According to the radiation noise characteristics of the underwater acoustic target, determine the appropriate low cutoff frequency and high cutoff frequency of the bandpass filter, and then apply the designed bandpass filter to the collected underwater acoustic signal to remove unnecessary frequency components; Noise reduction: Use Kalman filtering to reduce the radiated noise signal of underwater targets. Specifically, define the state equation and observation equation of the system. For underwater acoustic signals, assume that the state of the system is the amplitude and phase of the signal, and the observation equation is the collected signal value; predict the state at the next moment based on the state equation, update the state estimate based on the observation equation and observation data, repeat the prediction step and update step, and gradually reduce the impact of noise; Resampling: Select a preset target sampling frequency for the collected radiation noise signal of the underwater target, calculate the interpolation coefficient based on the original sampling frequency and the target sampling frequency, insert new sampling points between the sampling points of the original signal, and calculate the new radiation noise signal value of the underwater target through linear interpolation.

3. The method for extracting characteristics of underwater acoustic target radiation noise according to claim 1, characterized in that: The specific contents of the time-frequency analysis of step 2 are as follows: short-time Fourier transform: select a preset window function, the function of the window function is to segment the signal, that is, slide the window function in time to segment the preprocessed signal; perform Fourier transform on each segmented signal to obtain spectrum information; arrange the spectrum information of each segment in chronological order to form a time-frequency diagram.

4. The method for extracting characteristics of underwater acoustic target radiation noise according to claim 1, characterized in that: The specific content of the feature extraction in step 3 is as follows: Power spectral density: Calculate the energy of each frequency component in the time-frequency diagram in step 2, using the squared amplitude, and then average the energy of all frequency components to obtain the power spectral density; Envelope spectrum: Envelope detection is performed on each time window signal in the time-frequency diagram in step 2, specifically using square detection and Hilbert transform method to obtain its envelope signal, and then Fourier transform is performed on the envelope signal to obtain the envelope spectrum; Instantaneous frequency: Perform Hilbert transform on each time window signal in the time-frequency diagram in step 2 to obtain its analytical signal, calculate the phase change rate of the analytical signal to obtain the instantaneous frequency; Sparse coding technology: The extracted spectral features, envelope features and instantaneous frequency features are represented as a set of sparse basis functions. The selected basis functions are combined into a dictionary. The sparse representation of the signal in the dictionary is then solved by the LASSO optimization algorithm. The sparse representation coefficients are used as feature vectors for subsequent classification and recognition.

5. The method for extracting characteristics of underwater acoustic target radiation noise according to claim 1, characterized in that: The specific implementation steps of the comprehensive features in step 4 are as follows: Step 41, feature extraction: extracting the power spectrum density, envelope spectrum and instantaneous frequency features in step 3; Step 42, feature normalization: normalize the extracted features to ensure that the dimensions and value ranges of different features are consistent. Specifically, convert each element of each feature vector in step 41 to the range of [0, 1]. Step 43: Determine weights: Manually set the weights of each feature in step 41 based on domain knowledge and experience; Step 44, feature fusion: The feature vector normalized in step 42 is combined with the feature weight determined in step 43 for weighted fusion to form a comprehensive feature vector.

6. The method for extracting characteristics of underwater acoustic target radiation noise according to claim 1, characterized in that: The specific steps of feature dimensionality reduction in step 5 are as follows: Step 51, select multiple samples of comprehensive feature vectors from step 4 to form a feature matrix X, assuming that the size of X is M×N, where N is the number of samples and M is the dimension of the feature vector; Step 52: normalize the feature matrix X. The formula is: Where μ(X) is the mean of the feature matrix, σ(X) is the standard deviation of the feature matrix; then the normalized feature matrix X′ is further centered to make its mean 0, the specific formula is: X″=X′-μ(X′); Step 53: whiten the centralized feature matrix X″ so that its covariance matrix is ​​the unit matrix: X white =WX″, where W is the whitening matrix obtained by eigenvalue decomposition or singular value decomposition; Step 54: Use the ICA algorithm to whiten the feature matrix X white Perform independent component analysis to obtain the independent component matrix S and the mixing matrix A. The specific formula is: white ≈AS; Step 55: Select the first k independent components from the independent component matrix S to form a reduced-dimensional feature matrix Y: Y = S[:,:k], where the size of Y is N×k.

7. The method for extracting characteristics of underwater acoustic target radiation noise according to claim 1, characterized in that: The specific steps of feature optimization in step 6 are as follows: Step 61, data preparation: the comprehensive feature vector is reduced in dimension by independent component analysis to obtain a feature matrix Y after dimension reduction; Step 62: Recursively remove the least important features in the feature matrix Y and gradually reduce the number of features to evaluate the importance of each feature. Step 63: Select the most important feature subset based on the evaluation results; Step 64: Use the selected feature subset to form a new feature matrix Y opt , that is, the optimized feature matrix is ​​formed.

Images