Underwater sound source phase undistorted line spectrum enhancement method based on deep learning

Abstract

The invention relates to a phase-undistorted underwater sound source line spectrum enhancement method based on deep learning, and belongs to the technical field of underwater sound detection and sonar. The problems of low signal-to-noise ratio, interruption, Doppler frequency shift and the like of a moving sound source line spectrum and the limitation that a traditional enhancement method is greatly influenced by the signal-to-noise ratio and only nonlinear noise can be processed are solved. Generating training data set pairs and setting labels according to line spectrum model simulation, and verifying generation and setting of a data set; an overall architecture and key components of the FRCRN based on deep learning; designing a loss function of combining the CIRM with a weight scale invariant signal-to-noise ratio (SI-SNR); according to the method, the phase undistorted line spectrum enhancement of the underwater target sound source is successfully realized by utilizing the FRCRN. The method has remarkable advantages in the aspects of distortion-free performance and signal-to-noise ratio gain, and the SNR and the SI-SNR can achieve the gain of 7dB or above under the condition of color noise simulation. After the SwellEx-96 sea test data line spectrum is enhanced, the SNR gain and the SI-SNR gain can be respectively improved by 5.37 dB and 6.18 dB, and the target orientation estimation spatial spectrum gain is improved by 3.79 dB.

Description

Technical Field

[0001] The present invention belongs to the fields of underwater acoustic detection, sonar technology, etc., and relates to a method for enhancing the line spectrum of an underwater sound source without phase distortion by deep learning. Background Art

[0002] The passive sonar detection and recognition of underwater targets is a key research topic in the current underwater acoustic field. However, in the actual scenario of underwater acoustic target recognition, factors such as the complex sound generation mechanism of the target, the spatio-temporal variation of the ocean channel, background and platform noise, etc. affect the signal-to-noise ratio of the radiated noise of the underwater acoustic target to be often very low. At the same time, due to the relative motion between the measured target and the sonar, phenomena such as line spectrum interruption and Doppler frequency shift will occur. Traditional enhancement methods mainly use an Adaptive Line Enhancer (ALE), a Wiener Filter (WF), etc. to achieve signal enhancement. Its principle mainly utilizes the coherent characteristics of the line spectrum, matches the unknown line spectrum signal, and suppresses the additive broadband noise to enhance the signal-to-noise ratio of the line spectrum. Such as the literature "Sparsity-based adaptive line enhancer for passive sonars" (published in the 13th issue of "IET Radar Sonar and Navigation" in 2019, starting page number is 1796), the literature "Underwater target spectral entropy detection based on sparse-driven adaptive line spectrum enhancement" (published in the 6th issue of "Acta Acustica" in 2021, starting page number is 1059), and the literature "Deep-learning-based line enhancer for passive sonar systems" (published in the 16th issue of "IET Radar Sonar and Navigation" in 2022, starting page number is 589) all implement line spectrum enhancement based on traditional methods. However, when implementing line spectrum enhancement based on traditional methods, when the signal-to-noise ratio is lower than a certain threshold, the enhancement effect will deteriorate rapidly. There is literature showing through simulation that when the signal-to-noise ratio is lower than certain conditions, it is difficult for traditional methods to obtain gain and it is impossible to reduce non-linear noise. (Line Spectrum Enhancement Algorithm Based on Unsupervised Deep Learning (published in the 42nd issue of "Ship Science and Technology" in 2020, starting page number is 4) Summary of the Invention

[0003] (I) Object of the Invention

[0004] The object of the present invention is to provide a method for underwater sound source phase distortion-free line spectrum enhancement based on deep learning, which is used to solve the problems that the performance of traditional acoustic signal processing methods degrades under low signal-to-noise ratio conditions; the effect of traditional acoustic signal enhancement methods is not good under colored noise conditions; the ocean acoustic field noise is complex and it is difficult to obtain a pure and sufficient training data set; the enhancement algorithm causes distortion of the phase information of the line spectrum signal, which hinders further target positioning, tracking and recognition, and finally realizes underwater target sound source phase distortion-free line spectrum enhancement.

[0005] (II) Technical solution

[0006] To achieve the above object and solve the above technical problems, the technical solution of the present invention is as follows:

[0007] A method for underwater sound source phase distortion-free line spectrum enhancement based on deep learning, and the specific steps are as follows:

[0008] Step 1: In order to simulate the physical characteristics of underwater sound source propagation in the underwater channel and solve the problem of insufficient pure training data set for underwater targets, this chapter uses the Bellhop ray model for simulation, studies the influence on the underwater acoustic signal transmission under different parameters, generates sufficient underwater sound source line spectrum simulation experimental data, and further provides a data basis for the deep learning model.

[0009] The radiation source level (SL) of the underwater sound source mainly concentrates on 160 - 200 dB. After the radiation noise propagates and attenuates (TL) through the simulated underwater acoustic channel to obtain the detection threshold (DT), the derivation formula for inversely calculating the noise level (NL) through the passive sonar equation is:

[0010]

[0011] In the formula, P n is the noise power, P ref is the reference sound pressure in water, and SNR pre is the preset noise signal-to-noise ratio.

[0012] The simulated line spectrum consists of two parts: random and deterministic. The deterministic part is the superposition of sine waves with frequencies equal to the random frequencies respectively, and the random part uses Gaussian random white noise to generate a colored noise signal through a band-pass filter. The simulated noise signal can be obtained by passing Gaussian white noise through a filter with an amplitude response of 6 dB per octave attenuation. The result of adding noise to the line spectrum is as Figure 1 shown.

[0013] Step 2: The line spectra of the training dataset of the present invention are randomly distributed between 0 Hz and 1000 Hz, with a resolution of 0.01 Hz. The signal-to-noise ratio distribution of the signals is between -15 dB and 0 dB, with an interval of 5 dB, and the number of line spectra is distributed between 1 and 5. In addition, the enhancement of line spectra below 50 Hz is the frequency band that this study focuses on. To verify the enhancement ability of the model for line spectra in the low signal-to-noise ratio frequency band below 50 Hz, at least one line spectrum of each segment of data is distributed below 50 Hz. The cumulative duration of the dataset is 100,000 s and is divided into three parts: 80% for training (80,000 s), 10% for validation (10,000 s), and the remaining 10% for testing (10,000 s).

[0014] Step 3: The model is based on the learning domain masking method and adopts a feature extraction module, a mask separation module, and a signal reconstruction module. The overall network architecture of FRCRN is as Figure 2 shown. The model fuses the convolutional Encoder-Decoder structure and the convolutional recurrent network (CRN) with a recurrent structure, and enhances the learning ability of the frequency characteristics of line spectrum signals through a frequency recurrence mechanism. The model takes the real and imaginary component masks of the LOFAR spectrum complex value after STFT transformation as input, and the input feature can be denoted as X ∈ R B ×C×F×T ×2, where B represents the batch size, C represents the size of each batch of data, F and T represent the frequency dimension and the time dimension respectively. The encoder uses the Rs module to extract the feature representation of the complex spectrum and inputs it. The Recurrent module inserts two stacked complex feedforward sequential memory network (CFSMN) layers for temporal modeling to learn the temporal dynamic characteristics in the signal. Then, the decoder receives the information from the encoder, reconstructs the complex ratio mask (CIRM) corresponding to each signal, and multiplies it with the original LOFAR complex spectrum. Finally, the denoised spectrum is converted back to the waveform through ISTFT to obtain the denoised time-domain waveform. The model training is implemented based on the Pytorch framework and is trained on an NVIDIA GeForce RTX 4090 GPU with CUDA 11.3 version to improve the training efficiency. The Batch size is set to 4, and the model is trained iteratively 200 times. The optimizer is the Adam algorithm with an initial learning rate of 1.0e -4 The training sequence length is set to 1 second. STFT is calculated using a Hanning window, with a window length of 640, a step of 320, and 512 FFT points.

[0015] Step 4: Each of the Encoder module and the Decoder module in the model contains 10 RS modules. Its function is to extract the two-dimensional features of the input noise signal LOFAR spectrum. A Recurrent module is used to connect between the Encoder module and the Decoder module, and an attention mechanism module is added for connection, as Figure 3 shown, which shows the architectures of the Encoder and Decoder modules, and details the data input and output dimensions in the Encoder architecture. Each Rs module in the Encoder module and the Decoder module consists of a two-dimensional convolutional layer and an FSMN. After all convolutional layers in the encoder and decoder blocks, there are batch normalization (BatchNormalization, BN) and Relu activation functions. The features of each layer will finally pass through the FSMN layer for recursion along the frequency dimension to model broader frequency correlations. The kernel size of each convolutional layer in the Encoder module is (5, 2), and the stride is (2, 1). After dimensionality reduction through 10 RS modules, the final feature representation is R B×Cout×1×T×2 . The features extracted by the encoder block are passed to the corresponding decoder block through the Recurrent connection module and the attention mechanism module. The Decoder module is as Figure 3 shown on the right. Similarly, the kernel size of the convolutional layer is (5, 2), and the stride is (2, 1). After dimensionality increase through 10 RS deconvolution modules, the final feature representation is R B×Cout×F×T×2 , for reconstructing the complex ratio mask (CIRM) corresponding to each signal. Each layer of the Encoder passes through an attention mechanism module (Attention Block), and the attention mechanism module is as Figure 4 shown. After the signal features pass through the FSMN layer, both the real part and the imaginary part are averaged in the time domain and frequency domain through Average Pooling, and then enter the linear layer, Relu activation function, linear layer, and Sigmod activation function in sequence, and are compressed along the C dimension to add different weight coefficients to each channel.

[0016] Step 5: The time-domain loss function SI-SNR is usually used as an evaluation index for noise suppression. SI-SNR uses a single coefficient to explain the scaling difference compared with SNR. By normalizing the signal to zero mean before calculation, scale invariance is ensured, thus more accurately evaluating the quality of the signal. On this basis, this study uses the combined SI-SNR loss function and CIRM to optimize the loss function. The ideal CIRM, SI-SNR, and the final loss function are shown as follows:

[0017]

[0018] In the formula, S target represents the original clean signal, Denoted as the target estimation signal, Denoted as the target estimation complex ratio mask, θ represents the normalization scale, and MSE(·) represents the mean calculation, representing the signal power.

[0019] (III) Effective Yield

[0020] 1. The model of the present invention utilizes the frequency recurrence mechanism and the multi-head attention mechanism to enhance the network's learning ability for the frequency-dimensional features of the line spectrum signal.

[0021] 2. The present invention uses a Feedforward Sequential Memory Networks (FSMN) layer between the encoder and the decoder to replace the Long Short-Term Memory Networks (LSTM) layer to further learn the temporal information of the signal. The model predicts the Complex Ideal Ratio Mask (CIRM) of the line spectrum signal, which can significantly reduce the phase distortion of the line spectrum signal.

[0022] 3. The present invention adopts the combined CIRM and the Weight Scale-Invariant Signal-to-Noise Ratio (SI-SNR) loss function to optimize the training of the model.

[0023] 4. Given sufficient training samples, the present invention can automatically learn the complex relationship between the noisy line spectrum and the clean line spectrum.

[0024] In view of the low signal-to-noise ratio, interruption, Doppler frequency shift, etc. existing in the line spectrum of the moving sound source and the limitation that the traditional enhancement method is greatly affected by the signal-to-noise ratio and can only process non-linear noise, the present invention proposes a supervised deep learning method under low signal-to-noise ratio conditions, and successfully realizes the phase-undistorted line spectrum enhancement of the underwater target sound source by using FRCRN. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 is the LOFAR diagram of the colored noise under the simulated real ocean environment of the present invention;

[0026] Figure 2 is the overall network architecture block diagram of the line spectrum enhancement based on FRCRN of the present invention;

[0027] Figure 3 is the architecture of the Encoder and Decoder modules designed by the present invention;

[0028] Figure 4 is the architecture of the attention mechanism module of the network model of the present invention;

[0029] Figure 5is the enhanced result of the simulated ocean ambient noise data used in the present invention;

[0030] Figure 6 is the sound speed profile of the SwellEx-96 experimental site;

[0031] Figure 7 is the enhanced result of Channel 1 of the SwellEx-96 S59 experimental data used in the present invention; Figure 8 is the conventional beamforming result of the data before and after enhancement of the SwellEx-96 S5 experiment used in the present invention; Figure 9 is the conventional beamforming result of the data before and after enhancement of the SwellEx-96 S59 experiment used in the present invention; Figure 10 is the slice result of the 150th frame of the beamforming result of the SwellEx-96 S5 and S59 experiments used in the present invention. Detailed implementation manners

[0032] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings, simulation data and SwellEx-96 real experimental data.

[0033] The present invention provides an embodiment of an underwater sound source phase-invariant line spectrum enhancement method based on deep learning, and the process is divided into the following steps:

[0034] Step 1: Under the condition of colored noise simulation data simulating the real ocean environment, in order to conveniently display that the frequency points of the three-line spectrum are 49.4 Hz, 101.8 Hz, and 180.6 Hz respectively, the intuitive enhancement result of the model for the three-line spectrum of -10 dB colored noise is as Figure 5 shown. Under the condition of -10 dB colored noise, the three line spectra are almost submerged in the noise below 200 Hz. FRCRN successfully extracts the three line spectra submerged in the noise and has a good noise reduction effect on the noise floor around the line spectra. The average signal-to-noise ratio gains SNR and SI-SNR can both reach more than 7 dB.

[0035] Step 2: In order to evaluate the generality and noise reduction performance of the model in the real ocean environment, the present invention first uses the SwellEx-96 (http: / / swellex96.ucsd.edu) public dataset for verification. The sound speed profile of the experimental site is as Figure 6 shown. The data selects the data from 0 to 1000 s in the horizontal line array (Horizontal line arrays-North, HLA-N) in the S5 and S59 events for verification. The number of processing channels of the HLA-N array is 27, and the array sampling frequency is 3276.8 Hz. For the convenience of display, Figure 7The line spectrum enhancement results of the FRCRN for the receiving channel 1 of the S59 experiment event are given. For the enhancement of the SwellEx-96 real sea trial data by the FRCRN, the average SNR gain is 5.37 dB and the average SI-SNR gain is 6.18 dB.

[0036] Step 3: To verify the effectiveness and authenticity of the model for array signal processing, after enhancing the full-channel data, the CBF is calculated for the enhanced data, with one frame calculated every 5 s. The comparison results of the normalized azimuth histories of S5 and S59 show that the target trajectories are significantly enhanced from the comparison before and after enhancement. The enhancement effects are mainly reflected in: (1) The original weak target trajectories such as the azimuth angle of 313° in the S5 event are further enhanced; (2) For the targets at azimuth angles of 116° and 132° in the S5 event and the target at azimuth angle of 87° in the S59 event, the target trajectories that are almost unrecognizable are highlighted; (3) For the target at azimuth angle of 245° in the S5 event, two targets with relatively close trajectories are distinguished.

[0037] Step 4: The present invention further quantitatively analyzes the enhancement ability of the model for the full-channel data. In this chapter, the slices of the CBF results before and after enhancement of the 150th frame of the S5 event and the S59 event are extracted for comparison. The energy of the azimuth history diagram after calculating the CBF for each channel before and after enhancement is significantly higher than the result of directly calculating the CBF for the original data. Among them, the spatial spectrum gain corresponding to the target of the 150th frame slice of the S5 event is 4.32 dB, the spatial spectrum gain corresponding to the 150th frame slice of the S59 event is 3.13 dB, and the average spatial spectrum gain is 3.79 dB (S5: 4.51 dB; S59: 3.07 dB).

[0038] Step 5: The present invention proposes to apply the FRCRN model to denoise underwater acoustic signals under low signal-to-noise ratio conditions. In terms of feature extraction, FRCRN introduces a frequency recurrence mechanism and an attention mechanism after each convolutional layer to improve the learning ability of frequency dimension features. Between the Encoder and the Decoder, two stacked complex feedforward sequential memory network (CFSMN) layers are used to replace the traditional complex long short-term memory neural network layer (CLSTM) to further learn the temporal features of supervised samples and noise samples. The color noise simulation dataset and the SwellEx-96 dataset are used to further verify the generality and feasibility of the model. For the simulation dataset, the line spectrum signal can achieve an average SNR gain of 7.77 dB and an SI-SNR gain of 7.91 dB. For the HLV-N data in the SwellEx-96 sea trial events S5 and S59, an SNR gain of 5.37 dB and an SI-SNR gain of 6.18 dB can be achieved; the average spatial spectrum gain is 3.79 dB (the average spatial spectrum gain for event S5 is 4.51 dB, and the average spatial spectrum gain for event S59 is 3.07 dB). It is verified that the FRCRN model has good distortionless property after enhancing the underwater target signal.

[0039] The above content is a further detailed description of the present invention in combination with specific implementation manners, and it cannot be determined that the specific implementation of the present invention is only limited to these descriptions. For those of ordinary skill in the technical field to which the present invention belongs, without departing from the concept of the present invention, several simple deductions or substitutions can be made, and all should be regarded as belonging to the protection scope of the present invention.

Claims

1. A method for enhancing underwater sound source phase without distortion based on deep learning, characterized in that: The steps include: Step 1: Acquisition of underwater sound source line spectrum simulation experimental data Use Bellhop acoustic line model simulation to study the impact of different parameters on underwater acoustic signal transmission, generate sufficient underwater sound source line spectrum simulation experimental data sets, and provide a data basis for deep learning models; Step 2: The line spectrum of the training data set is randomly distributed between 0Hz and 1000Hz, with a resolution of 0.01Hz, a signal-to-noise ratio distribution between -15dB and 0dB, an interval of 5dB, and the number of line spectra is distributed between 1 and 5; ensure that at least one line spectrum of each segment of data is distributed below 50Hz. The cumulative duration of the data set is 100,000s and is divided into three parts, of which 80% is used for training, 10% is used for verification, and the remaining 10% is used for testing; Step 3: The Bellhop sound line model is based on the learning domain mask method, and adopts a feature extraction module, a mask separation module, and a signal reconstruction module. The Bellhop sound line model integrates a convolutional encoder-decoder structure and a convolutional recursive network with a loop structure, and enhances the learning ability of the frequency characteristics of the line spectrum signal through a frequency recursive mechanism; The overall network architecture of FRCRN is described as follows: The Bellhop ray model takes the real and imaginary component mask of the complex value of the LOFAR spectrum after STFT transformation as input, and the input feature is recorded as X∈R B×C×F×T×2 , where B represents the batch size, C represents the size of each batch of data, F and T represent the frequency dimension and time dimension respectively; the encoder uses the Rs module to extract the feature representation of the complex spectrum and inputs it, and the Recurrent module inserts two stacked complex feedforward sequential memory network layers for time series modeling and learning the temporal dynamic features in the signal; The decoder receives information from the encoder, reconstructs the complex ratio mask corresponding to each signal and multiplies it with the original LOFAR complex spectrum; finally, the denoised spectrum is converted back to a waveform through ISTFT to obtain a denoised time domain waveform; Step 4: In the Bellhop acoustic line model, the Encoder module and the Decoder module each contain 10 layers of RS modules, which are used to extract the two-dimensional features of the LOFAR spectrum of the input noise signal. The Recurrent module is used between the Encoder module and the Decoder module and the attention mechanism module is added for connection; each RS module of the Encoder module and the Decoder module consists of a two-dimensional convolution layer and a FSMN. All convolution layers in the encoder and decoder blocks are followed by batch normalization and Relu activation function; each layer of features will eventually pass through the FSMN layer for recursion along the frequency dimension to model a wider range of frequency correlations; the features extracted by the encoder block are passed to the corresponding decoder block through the Recurrent connection module and the attention mechanism module; each layer of the Encoder passes through an attention mechanism module. After the signal features pass through the FSMN layer, the real part and the imaginary part are averaged in the time domain and frequency domain through Average Pooling, and then enter the linear layer, Relu activation function and linear layer, Sigmod activation function in turn, and are compressed along the C dimension, adding different weight coefficients to each channel; Step 5: Use the joint SI-SNR loss function and CIRM to optimize the loss function. The training of the Bellhop acoustic model, the ideal CIRM, SI-SNR and the final loss function are shown in the following formula: Where S target Represented as the original clean signal, Denoted as the target estimation signal, It is represented as the target estimated complex ratio mask, θ represents the normalized scale, MSE(·) represents the mean calculation, Represents the signal power.

2. According to the method for enhancing underwater sound source phase without distortion based on deep learning in claim 1, it is characterized in that: The specific implementation steps of step 1 are as follows: the underwater sound source radiation source level is concentrated in 160-200dB, the radiation noise is attenuated by the simulated underwater acoustic channel propagation (TL) to obtain the detection threshold (DT), and the derivation formula of the noise level (NL) is inversely deduced by the passive sonar equation: Where P n is the noise power, P ref is the reference sound pressure in water, SNR pre is the preset signal-to-noise ratio. The simulated line spectrum consists of two parts, random and deterministic. The deterministic part is the superposition of sine waves with frequencies equal to the random frequencies, and the random part uses Gaussian random white noise to generate a colored noise signal through a bandpass filter. The simulated noise signal is obtained by passing the Gaussian white noise through a filter with an amplitude response of 6dB per octave attenuation.

3. According to the method for enhancing underwater sound source phase without distortion based on deep learning in claim 1, it is characterized in that: In step 3, the Bellhop acoustic line model training is implemented based on the Pytorch framework and is trained on the NVIDIA GeForceRTX 4090GPU, CUDA11.3 version to improve training efficiency.

4. According to claim 3, a method for enhancing underwater sound source phase without distortion based on deep learning is characterized in that: In step 3, in the Bellhop acoustic line model training, the batch size is set to 4, the model training iterations are 200 times; the optimizer has an initial learning rate of 1.0e -4 Adam algorithm; The training sequence length is set to 1 second; STFT is calculated using a Hanning window with a window length of 640, a step size of 320, and an FFT point number of 512.

5. According to the method for enhancing underwater sound source phase without distortion based on deep learning in claim 1, it is characterized in that: In step 4, the kernel size of each convolutional layer in the Encoder module is (5,2), the stride is (2,1), and the final feature representation is R after dimensionality reduction through 10 layers of RS modules. B×Cout ×1×T×2.

6. According to the method of claim 1, the method is characterized in that: In step 4, the kernel size of the convolution layer of the decoder module is (5, 2), the stride is (2, 1), and the final dimension-upgraded feature representation is R after 10 layers of RS deconvolution modules. B×Cout×F×T×2 , to reconstruct the complex ratio mask corresponding to each signal.

7. A method for enhancing underwater sound source phase without distortion based on deep learning according to any one of claims 1 to 6, characterized in that: The underwater sound source phase-undistorted line spectrum enhancement method can achieve a gain of more than 7dB for both SNR and SI-SNR under colored noise simulation conditions; after line spectrum enhancement of SwellEx-96 sea trial data, the SNR and SI-SNR gains are increased by 5.37dB and 6.18dB respectively, and the spatial spectrum gain of target azimuth estimation is increased by 3.79dB.

Images