Underwater acoustic signal noise suppression method combining bimodal neural network and spectral subtraction

Abstract

The invention discloses a bimodal neural network and spectral subtraction combined underwater acoustic signal noise suppression method. Comprising the following steps: firstly, constructing a simulation underwater sound data set containing self-noise; thirdly, constructing a bimodal neural network, then training the bimodal neural network until training is completed, and obtaining a platform self-noise estimation model; inputting a to-be-processed noise-containing underwater acoustic signal into the platform self-noise estimation model, and outputting estimation platform self-noise by the model; and finally, based on the self noise of the estimation platform, noise reduction processing is carried out on the noise-containing underwater acoustic signal to be processed by using the improved spectral subtraction method, and then a noise-reduced underwater acoustic signal is obtained. The method combines an advanced deep learning method and a classical spectral subtraction algorithm, has the capabilities of autonomous learning and efficient noise reduction, remarkably improves the signal-to-noise ratio of the underwater acoustic signal, is suitable for noise reduction of the underwater acoustic signal in a non-stationary and complex marine environment, remarkably improves the low-frequency noise suppression capability and the target signal fidelity, and has a good application prospect. The method is suitable for underwater acoustic target detection, underwater acoustic communication and other scenes.

Description

Technical Field

[0001] The present invention belongs to an underwater acoustic signal noise suppression method in the technical field of underwater acoustic signal processing, and specifically relates to an underwater acoustic signal noise suppression method combining a dual-modal neural network and spectral subtraction. Background Technique

[0002] Sound waves are currently the only form of energy radiation that can achieve long-distance propagation in the ocean. As an information carrier, they are widely used in tasks such as submarine detection, underwater target recognition, tactical communication, and underwater navigation. However, various forms of self-noise will be generated during the operation of the sound signal receiving platform, including mechanical vibration, electromagnetic interference, and fluid noise. These self-noises seriously affect the reception quality of the sound signal, not only limiting the detection range of the equipment, but also weakening the clarity and reliability of the sound signal, thereby causing trouble for the accurate positioning and identification of enemy targets, and even threatening the accuracy of tactical decisions in severe cases. Therefore, self-noise suppression has become an important research topic in sound signal receiving platforms.

[0003] The self-noise of the receiving platform mainly comes from the cavitation noise generated by the rotation of the propeller, the mechanical vibration noise of the propulsion system, and the hydrodynamic noise caused by the friction between the fluid and the hull. Affected by factors such as the platform's motion state, the power change of the propulsion system, and the vibration of the mechanical structure, this noise exhibits highly non-stationary characteristics in terms of time and frequency.

[0004] Traditional self-noise suppression methods such as spectral subtraction, adaptive filtering, wavelet transform, and empirical mode decomposition have some defects in non-stationary noise environments. First of all, these methods rely on the statistical characteristics of the prior noise, while the self-noise of the receiving platform has time-varying and non-stationary characteristics, making it difficult for traditional methods to dynamically model and adapt to complex noise changes. Moreover, when traditional methods suppress broadband noise, they often damage the line spectrum characteristics of the signal, resulting in the loss of target information. In addition, in a complex underwater acoustic environment, the reverberation effect is coupled with the noise, further reducing the noise separation ability of traditional algorithms.

[0005] Deep learning methods demonstrate strong feature extraction capabilities in underwater acoustic signal processing and can, to a certain extent, separate different components in the signal. According to the different input features, existing deep learning denoising models can be roughly divided into two categories: based on spectrograms and based on waveforms. The former usually converts the audio waveform into a time-frequency spectrogram through the short-time Fourier transform (STFT) to extract frequency structure information, and further refines it into spectrogram masking and spectrogram mapping strategies. However, the time-frequency domain conversion will lose time resolution under the action of the window function, and the modeling of phase information was generally ignored in the early studies using this type of method, resulting in distortion when reconstructing the signal and limiting the denoising performance. The latter directly processes the time-domain waveform, retains the complete phase information and can better capture the instantaneous change characteristics of the signal, but has deficiencies in frequency-domain modeling and is difficult to effectively distinguish noise and target components in terms of frequency.

[0006] Given that the platform self-noise has complex structural characteristics in the time and frequency dimensions, a single-modal processing method often lacks adaptability when facing diverse noises. Summary of the Invention

[0007] The purpose of the present invention is to overcome the deficiencies of the prior art and provide an underwater acoustic signal noise suppression method that combines a dual-modal neural network and spectral subtraction. Aiming at the characteristics that the underwater acoustic signal is unknown but the platform self-noise can be determined, this method uses a dual-modal neural network to extract and fuse the high time resolution in the time domain and the accurate modeling ability of frequency and phase characteristics in the spectral domain. It can significantly improve the denoising performance while retaining signal details, thereby realizing the estimation of platform self-noise, and combining spectral subtraction to achieve self-noise suppression. This method has the ability of autonomous learning and efficient denoising, is applicable to underwater acoustic signal denoising in non-stationary and complex marine environments, significantly improves the low-frequency noise suppression ability and target signal fidelity, improves the quality of received underwater acoustic signals, reduces the interference of the received platform self-noise on the signal, and is applicable to scenarios such as underwater acoustic target detection and underwater acoustic communication.

[0008] The purpose of the present invention is achieved through the following technical solutions:

[0009] 1. An underwater acoustic signal noise suppression method that combines a dual-modal neural network and spectral subtraction

[0010] Step 1: Construct a simulated underwater acoustic data set containing self-noise;

[0011] Step 2: Construct a dual-modal neural network, and then use the simulated underwater acoustic data set containing self-noise to train the dual-modal neural network until the training is completed to obtain a platform self-noise estimation model; input the noisy underwater acoustic signal to be processed into the platform self-noise estimation model, and the model outputs the estimated platform self-noise;

[0012] Step 3: Based on the estimated platform self-noise, use the improved spectral subtraction method to perform noise reduction on the noisy underwater acoustic signal to be processed, and obtain the denoised underwater acoustic signal.

[0013] The specific steps of Step 1 are as follows:

[0014] First, obtain the self-noise signal of the receiving platform; then generate a simulated underwater acoustic signal with time-varying and non-stationary characteristics; finally, fuse the self-noise signal of the receiving platform with each simulated underwater acoustic signal to generate a simulated underwater acoustic signal with self-noise, so as to obtain a noisy simulated underwater acoustic dataset.

[0015] In Step 2, the dual-modal neural network includes 4 time-domain encoders, 4 time-domain decoders, 4 spectral-domain encoders, 4 spectral-domain decoders, and a cross-domain Transformer encoder. The 4 time-domain encoders are connected in sequence, the 4 time-domain decoders are connected in sequence, the 4 spectral-domain encoders are connected in sequence, and the 4 spectral-domain decoders are connected in sequence. The fourth time-domain encoder and the first time-domain decoder are connected by the cross-domain Transformer encoder, the fourth time-domain encoder and the first time-domain decoder are connected by the cross-domain Transformer encoder, the first time-domain encoder is also connected to the fourth time-domain decoder, the second time-domain encoder is also connected to the third time-domain decoder, the third time-domain encoder is also connected to the second time-domain decoder, and the fourth time-domain encoder is also connected to the first time-domain decoder; the first spectral-domain encoder is also connected to the fourth spectral-domain decoder, the second spectral-domain encoder is also connected to the third spectral-domain decoder, the third spectral-domain encoder is also connected to the second spectral-domain decoder, and the fourth spectral-domain encoder is also connected to the first spectral-domain decoder; the input of the dual-modal neural network is used as the input of the first time-domain encoder. After performing short-time Fourier transform on the input of the dual-modal neural network, it is input into the first spectral-domain encoder. After performing inverse short-time Fourier transform on the output of the fourth spectral-domain decoder, it is superimposed with the output of the fourth time-domain decoder to finally obtain the output of the dual-modal neural network.

[0016] The cross-domain Transformer encoder includes a self-attention layer and a cross-domain attention layer. The first self-attention layer is connected to the third self-attention layer after passing through the first cross-domain attention layer, the second self-attention layer, and the second cross-domain attention layer in sequence; the fourth self-attention layer is connected to the sixth self-attention layer after passing through the third cross-domain attention layer, the fifth self-attention layer, and the fourth cross-domain attention layer in sequence; the output of the first self-attention layer is also used as the input of the third cross-domain attention layer, and the output of the second self-attention layer is also used as the input of the fourth cross-domain attention layer; the output of the fourth self-attention layer is also used as the input of the first cross-domain attention layer, and the output of the fifth self-attention layer is also used as the input of the second cross-domain attention layer. The outputs of the third self-attention layer and the sixth self-attention layer are respectively used as the two outputs of the cross-domain Transformer encoder.

[0017] The time-domain encoder includes a first one-dimensional convolutional layer, a first normalization layer, a first gelu activation layer, and a first residual block connected in sequence. The input of the time-domain encoder serves as the input of the first one-dimensional convolutional layer, and the output of the first residual block serves as the output of the time-domain encoder. The time-domain decoder includes a second one-dimensional convolutional layer, a second normalization layer, a second gelu activation layer, and a second residual block connected in sequence. The two inputs of the time-domain decoder are fused and then serve as the input of the second one-dimensional convolutional layer, and the output of the second residual block serves as the output of the time-domain decoder.

[0018] The spectral-domain encoder includes a first two-dimensional convolutional layer, a third normalization layer, a third gelu activation layer, and a third residual block connected in sequence. The input of the spectral-domain encoder serves as the input of the first two-dimensional convolutional layer, and the output of the third residual block serves as the output of the spectral-domain encoder. The spectral-domain decoder includes a second two-dimensional convolutional layer, a fourth normalization layer, a fourth gelu activation layer, and a fourth residual block connected in sequence. The two inputs of the spectral-domain decoder are fused and then serve as the input of the second two-dimensional convolutional layer, and the output of the fourth residual block serves as the output of the spectral-domain decoder.

[0019] Step 3 is specifically as follows:

[0020] Step 3.1: Perform short-time Fourier transforms on the estimated platform self-noise and the noisy underwater acoustic signal to be processed respectively, and obtain the complex time-frequency spectral values corresponding to the estimated platform noise and the noisy underwater acoustic signal respectively.

[0021] Step 3.2: Generate the real-time signal-to-noise ratio of the two on the frequency axis according to the complex time-frequency spectral values corresponding to the estimated platform noise and the noisy underwater acoustic signal, and generate a dynamic subtraction factor matrix α according to the real-time signal-to-noise ratio of the two on the frequency axis.

[0022] Step 3.3: Combine the subtraction factor matrix α and use the following formula to calculate and obtain the complex time-frequency spectral values corresponding to the denoised underwater acoustic signal:

[0023]

[0024] where represents the complex time-frequency spectral amplitude value of the denoised underwater acoustic signal, max(,) represents the maximum value operation, X(f,t) represents the complex time-frequency spectral value of the noisy underwater acoustic signal, represents the complex time-frequency spectral value of the estimated platform noise, β represents the noise threshold; || represents the amplitude operation;

[0025] Step 3.4: After performing the inverse short-time Fourier transform on the complex time-frequency spectral values corresponding to the denoised underwater acoustic signal, obtain the denoised underwater acoustic signal.

[0026] II. An underwater acoustic signal noise suppression device combining a dual-modal neural network and spectral subtraction

[0027] A dataset construction unit for constructing a simulated underwater acoustic dataset containing self-noise based on the self-noise signal of the receiving platform;

[0028] A neural network storage unit for storing a dual-modal neural network;

[0029] A training unit for training the dual-modal neural network using the simulated underwater acoustic dataset containing self-noise until the training is completed to obtain a platform self-noise estimation model;

[0030] A signal noise reduction unit for inputting the noisy underwater acoustic signal to be processed into the platform self-noise estimation model to obtain the estimated platform self-noise; and then based on the estimated platform self-noise, using the improved spectral subtraction method to perform noise reduction processing on the noisy underwater acoustic signal to be processed to obtain the denoised underwater acoustic signal.

[0031] III. A computer device

[0032] The computer device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of the method for suppressing underwater acoustic signal noise by combining a dual-modal neural network and spectral subtraction are implemented.

[0033] IV. A computer-readable storage medium

[0034] The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the method for suppressing underwater acoustic signal noise by combining a dual-modal neural network and spectral subtraction are implemented.

[0035] The beneficial effects of the present invention are:

[0036] 1) The present invention generates a simulated underwater acoustic signal dataset through multi-physical field joint simulation, which is applicable to underwater acoustic signals in the actual environment.

[0037] 2) The present invention takes into account that the type of platform self-noise is clear and can be collected in advance, while the target signal is dynamically changing and unknown in actual applications. Therefore, by using a dual-modal neural network, the time-domain waveform features of the signal are extracted to enhance the ability to capture transient noise. The time-frequency branch obtains complex time-frequency spectrum features based on STFT to capture frequency and phase features, and uses a cross-modal attention mechanism to fuse dual-modal information to realize the modeling of self-noise features, which can better enhance the adaptability and generalization ability of the model in multiple scenarios.

[0038] 3) The present invention introduces a dynamic spectral subtraction factor based on real-time signal-to-noise ratio estimation through spectral subtraction according to the real-time signal-to-noise ratio dynamic spectral subtraction coefficient, and combines power spectral density curve analysis, and the measured signal-to-noise ratio can be increased by more than 30 dB.

[0039] 4) The results of the phase gain analysis of the present invention show that this method has significant advantages in phase information reconstruction. Among them, 95.77% of the samples show positive phase gain, and the maximum phase gain reaches 83.80%, which reflects the effectiveness of the model in phase recovery. Brief Description of the Drawings

[0040] Figure 1 is a schematic flowchart of an underwater acoustic signal noise suppression method combining a dual-modal neural network and spectral subtraction proposed by the present invention;

[0041] Figure 2 is a schematic diagram of the network structure of the dual-modal neural network adopted by the present invention;

[0042] Figure 3 is a flowchart of step 3 in the present invention;

[0043] Figure 4 is a schematic diagram of using the combined dual-modal model and spectral subtraction of the present invention to achieve self-noise suppression of an autonomous underwater vehicle; where (a) is the noisy underwater acoustic signal during sea trial, (b) is the denoised underwater acoustic signal, (c) is the estimated self-noise signal of the autonomous underwater vehicle; (d) is the corresponding power spectral density curve graph. Detailed Embodiment

[0044] To make the purpose, technical solution and advantages of the implementation of the present invention clearer, the technical solutions in the embodiments of the present invention will be described in more detail below with reference to the accompanying drawings in the embodiments of the present invention.

[0045] As Figure 1 shown, an underwater acoustic signal noise suppression method combining a dual-modal neural network and spectral subtraction proposed by the present invention includes the following steps:

[0046] Step 1: Construct a simulated underwater acoustic data set containing self-noise;

[0047] Step 1 is specifically as follows:

[0048] First, sonar is used to collect and process the self-noise signal of the receiving platform. The processing process includes preprocessing, data segmentation, and annotation to obtain the self-noise signal of the receiving platform. Then, the Matlab underwater acoustic toolbox is used to generate a simulated underwater acoustic signal with time-varying and non-stationary characteristics. Specifically, the simulated underwater acoustic signal is generated by first calculating the modal attenuation coefficient of the signal in the 0-500 Hz frequency band through the normal mode model, and then combining the ray theory equation to solve the acoustic ray trajectory of the signal in the 500-5000 Hz frequency band. Signals below 500 Hz are propagated and simulated through the normal mode model, while signals above 500 Hz are calculated using the ray theory equation to obtain the simulated underwater acoustic signal after attenuation through the underwater acoustic channel. The simulated signal includes periodic / aperiodic single-frequency signals and periodic / aperiodic linear frequency modulation (LFM) signals. Finally, the self-noise signal of the receiving platform is fused with each simulated underwater acoustic signal (i.e., the simulated underwater acoustic signal is superimposed on the self-noise signal) to generate a simulated underwater acoustic signal x with self-noise, x = s clean + n, where s clean is the simulated underwater acoustic signal and n is the self-noise signal of the receiving platform, which includes mechanical vibration noise, propeller vibration and cavitation noise, hydrodynamic noise, etc., to obtain a simulated underwater acoustic data set with noise. The training set and test set are established according to a preset ratio.

[0049] Step 2: Construct a dual-modal neural network, and then use the simulated underwater acoustic data set with self-noise to train the dual-modal neural network until the training is completed to obtain a platform self-noise estimation model;

[0050] Such as Figure 2As shown in (a) of, in step 2, the bimodal neural network includes 4 time-domain encoders, 4 time-domain decoders, 4 spectral-domain encoders, 4 spectral-domain decoders, and a cross-domain Transformer encoder. The 4 time-domain encoders are connected in sequence, the 4 time-domain decoders are connected in sequence, the 4 spectral-domain encoders are connected in sequence, and the 4 spectral-domain decoders are connected in sequence. The fourth time-domain encoder and the first time-domain decoder are connected by the cross-domain Transformer encoder. That is, the outputs of the fourth time-domain encoder and the fourth spectral-domain encoder are both used as the inputs of the cross-domain Transformer encoder. The two outputs of the cross-domain Transformer encoder are respectively used as the inputs of the first time-domain decoder and the first spectral-domain decoder. The 4 time-domain encoders, the 4 time-domain decoders, and the first self-attention layer, the first cross-domain attention layer, the second self-attention layer, the second cross-domain attention layer, and the third self-attention layer in the cross-domain Transformer encoder form the time-domain branch. The 4 spectral-domain encoders, the 4 spectral-domain decoders, and the fourth self-attention layer, the third cross-domain attention layer, the fifth self-attention layer, the fourth cross-domain attention layer, and the sixth self-attention layer in the cross-domain Transformer encoder form the spectral-domain branch. Therefore, the bimodal neural network is a hybrid time-domain - spectral-domain double U-Net architecture. The first time-domain encoder is also connected to the fourth time-domain decoder, the second time-domain encoder is also connected to the third time-domain decoder, the third time-domain encoder is also connected to the second time-domain decoder, and the fourth time-domain encoder is also connected to the first time-domain decoder; the first spectral-domain encoder is also connected to the fourth spectral-domain decoder, the second spectral-domain encoder is also connected to the third spectral-domain decoder, the third spectral-domain encoder is also connected to the second spectral-domain decoder, and the fourth spectral-domain encoder is also connected to the first spectral-domain decoder. Low-level detailed features can be retained in information transmission through skip connections. The input of the bimodal neural network (i.e., the simulated underwater acoustic signal with self-noise or the noisy underwater acoustic signal to be processed that has been normalized by the average noise level) is used as the input of the first time-domain encoder. After performing a short-time Fourier transform on the input of the bimodal neural network, a complex time-frequency spectrum is obtained, and then the complex time-frequency spectrum is input into the first spectral-domain encoder. After performing an inverse short-time Fourier transform on the output of the fourth spectral-domain decoder and then superimposing it with the output of the fourth time-domain decoder, the output of the bimodal neural network is finally obtained.

[0051] As Figure 2As shown in (b), the cross-domain Transformer encoder includes a self-attention layer and a cross-domain attention layer. The first self-attention layer is connected to the third self-attention layer after passing through the first cross-domain attention layer, the second self-attention layer, and the second cross-domain attention layer in sequence, thus forming a part of the time-domain branch; the fourth self-attention layer is connected to the sixth self-attention layer after passing through the third cross-domain attention layer, the fifth self-attention layer, and the fourth cross-domain attention layer in sequence, thus forming a part of the spectral-domain branch; the output of the first self-attention layer also serves as the input of the third cross-domain attention layer, and the output of the second self-attention layer also serves as the input of the fourth cross-domain attention layer; the output of the fourth self-attention layer also serves as the input of the first cross-domain attention layer, and the output of the fifth self-attention layer also serves as the input of the second cross-domain attention layer. The outputs of the third self-attention layer and the sixth self-attention layer serve as the two outputs of the cross-domain Transformer encoder respectively. The time-domain features and frequency-domain features extract deep context features through their respective self-attention layers, and then both are input into the cross-domain attention layer to achieve information interaction and fusion between the two branches.

[0052] The time-domain encoder includes a first one-dimensional convolutional layer, a first normalization layer, a first gelu activation layer, and a first residual block connected in sequence. The input of the time-domain encoder serves as the input of the first one-dimensional convolutional layer, and the output of the first residual block serves as the output of the time-domain encoder, which is used to gradually extract the instantaneous features of the original waveform.

[0053] The time-domain decoder includes a second one-dimensional convolutional layer, a second normalization layer, a second gelu activation layer, and a second residual block connected in sequence. The two inputs of the time-domain decoder are fused and then serve as the input of the second one-dimensional convolutional layer, and the output of the second residual block serves as the output of the time-domain decoder. The time-domain decoder restores the features to a time-domain signal to achieve fine reconstruction.

[0054] The spectral-domain encoder includes a first two-dimensional convolutional layer, a third normalization layer, a third gated gelu activation layer, and a third residual block connected in sequence. The input of the spectral-domain encoder serves as the input of the first two-dimensional convolutional layer, and the output of the third residual block serves as the output of the spectral-domain encoder, which is used to extract the spatial and temporal features of the spectral domain.

[0055] The spectral-domain decoder includes a second two-dimensional convolutional layer, a fourth normalization layer, a fourth gelu activation layer, and a fourth residual block connected in sequence. The two inputs of the spectral-domain decoder are fused and then serve as the input of the second two-dimensional convolutional layer, and the output of the fourth residual block serves as the output of the spectral-domain decoder. The spectral-domain decoder is used to recover the complex time-frequency spectrum and restore it to a time-domain signal through iSTFT.

[0056] During the training process of the dual-modal neural network, the loss is calculated based on the predicted self-noise time-domain waveform and the true self-noise time-domain waveform, so as to continuously iterate and train to optimize the neural network.

[0057] The noisy underwater acoustic signal during sea trial is as follows Figure 4 shown in (a) of . Input the noisy underwater acoustic signal during this sea trial into the platform self-noise estimation model, and the estimated platform self-noise Figure 4 is output by the model as shown in (c) of

[0058] Step 3: Based on the estimated platform self-noise After using the improved spectral subtraction method to denoise the to-be-processed noisy underwater acoustic signal, a denoised underwater acoustic signal is obtained, as shown in (b) of Figure 4 .

[0059] As shown in Figure 3 , Step 3 is specifically as follows

[0060] Step 3.1: Perform short-time Fourier transform (STFT) on the estimated platform self-noise and the to-be-processed noisy underwater acoustic signal respectively to obtain the complex time-frequency spectrum values corresponding to the estimated platform noise and the noisy underwater acoustic signal

[0061] Step 3.2: In the traditional spectral subtraction method, it is assumed that the noise is static and the starting segment is regarded as pure noise. However, in the present invention, a real-time signal-to-noise ratio on the frequency axis of the two is generated according to the complex time-frequency spectrum values corresponding to the estimated platform noise and the noisy underwater acoustic signal, and a dynamic subtraction factor matrix α is generated according to the real-time signal-to-noise ratio of the two on the frequency axis

[0062] Step 3.3: Combine the subtraction factor matrix α and use the following formula to calculate the complex time-frequency spectrum value corresponding to the denoised underwater acoustic signal

[0063]

[0064] where represents the complex time-frequency spectrum amplitude value of the denoised underwater acoustic signal, max(,) represents the maximum value operation, X(f,t) represents the complex time-frequency spectrum value of the noisy underwater acoustic signal represents the complex time-frequency spectrum value of the estimated platform noise, β represents the noise threshold, which is used to narrow the gap between the residual noise introduced in the case of imperfect and inaccurate noise estimation and the minimum value; || represents the amplitude operation

[0065] Step 3.4: Perform inverse short-time Fourier transform on the complex time-frequency spectrum value corresponding to the denoised underwater acoustic signal to obtain the denoised underwater acoustic signal

[0066] As can be seen from (a) of Figure 4 , in the actually collected noisy signal, the self-noise of the autonomous underwater vehicle will damage the signal quality Figure 4In (b), the AUV self-noise is estimated by the model. After using spectral subtraction for suppression, the denoised underwater acoustic signal can be obtained, that is Figure 4 In (c), it can be intuitively seen that the denoised signal is significantly cleaner. Comparing the power spectral density curves of these three, as shown in Figure 4 In (d), it is obvious that the signal-to-noise ratio of the target signal can be increased by more than 30 dB at 500 Hz.

[0067] The above steps have described in detail how to start from obtaining the self-noise of the receiving platform and the simulated underwater acoustic signal dataset, through constructing and training and optimizing the dual-modal neural network, and combining spectral subtraction, finally realizing the noise suppression of underwater acoustic signals. This process involves multiple links such as data processing, feature extraction, network optimization, model training, and information fusion. Each link is for improving the accuracy and efficiency of underwater acoustic signal denoising.

[0068] The present invention also proposes an underwater acoustic signal noise suppression device combining a dual-modal neural network and spectral subtraction, including:

[0069] A dataset construction unit, configured to construct a simulated underwater acoustic dataset containing self-noise according to the self-noise signal of the receiving platform;

[0070] A neural network storage unit, configured to store the dual-modal neural network;

[0071] A training unit, configured to use the simulated underwater acoustic dataset containing self-noise to train the dual-modal neural network until the training is completed to obtain a platform self-noise estimation model;

[0072] A signal denoising unit, configured to input the noisy underwater acoustic signal to be processed into the platform self-noise estimation model to obtain the estimated platform self-noise Then based on the estimated platform self-noise Use the improved spectral subtraction to perform denoising processing on the noisy underwater acoustic signal to be processed, and obtain the denoised underwater acoustic signal.

[0073] The present invention also proposes a computer device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the steps of a method for suppressing underwater acoustic signal noise by combining a dual-modal neural network and spectral subtraction.

[0074] The content described in the embodiments of this specification is only an enumeration of the implementation forms of the inventive concept. The protection scope of the present invention should not be regarded as limited to the specific forms stated in the embodiments. The protection scope of the present invention also extends to equivalent technical means that those skilled in the art can think of according to the inventive concept.

Claims

1. An underwater acoustic signal noise suppression method combining a dual-modal neural network and spectral subtraction, characterized in that It includes the following steps: Step 1: Construct a simulated underwater acoustic dataset with self-noise; Step 2: Construct a dual-modal neural network, and then use the simulated underwater acoustic dataset with self-noise to train the dual-modal neural network until the training is completed to obtain a platform self-noise estimation model; input the noisy underwater acoustic signal to be processed into the platform self-noise estimation model, and the model outputs the estimated platform self-noise; Step 3: Based on the estimated platform self-noise, use the improved spectral subtraction method to perform noise reduction processing on the noisy underwater acoustic signal to be processed, and obtain the denoised underwater acoustic signal.

2. The underwater acoustic signal noise suppression method combining a dual-modal neural network and spectral subtraction according to claim 1, characterized in that, The specific content of Step 1 is as follows: First, obtain the self-noise signal of the receiving platform; then generate simulated underwater acoustic signals with time-varying and non-stationary characteristics; finally, fuse the self-noise signal of the receiving platform with each simulated underwater acoustic signal to generate a simulated underwater acoustic dataset with self-noise, so as to obtain a noisy simulated underwater acoustic dataset.

3. A method for suppressing underwater acoustic signal noise by combining a dual-modal neural network and spectral subtraction according to claim 1, characterized in that In Step 2, the dual-modal neural network includes 4 time-domain encoders, 4 time-domain decoders, 4 spectral-domain encoders, 4 spectral-domain decoders and a cross-domain Transformer encoder. The 4 time-domain encoders are connected in sequence, the 4 time-domain decoders are connected in sequence, the 4 spectral-domain encoders are connected in sequence, and the 4 spectral-domain decoders are connected in sequence. The fourth time-domain encoder and the first time-domain decoder are connected by a cross-domain Transformer encoder, the fourth time-domain encoder and the first time-domain decoder are connected by a cross-domain Transformer encoder, the first time-domain encoder is also connected to the fourth time-domain decoder, the second time-domain encoder is also connected to the third time-domain decoder, the third time-domain encoder is also connected to the second time-domain decoder, and the fourth time-domain encoder is also connected to the first time-domain decoder; the first spectral-domain encoder is also connected to the fourth spectral-domain decoder, the second spectral-domain encoder is also connected to the third spectral-domain decoder, the third spectral-domain encoder is also connected to the second spectral-domain decoder, and the fourth spectral-domain encoder is also connected to the first spectral-domain decoder; the input of the dual-modal neural network is used as the input of the first time-domain encoder, and after performing short-time Fourier transform on the input of the dual-modal neural network, it is input into the first spectral-domain encoder. After performing inverse short-time Fourier transform on the output of the fourth spectral-domain decoder, it is superimposed with the output of the fourth time-domain decoder to finally obtain the output of the dual-modal neural network.

4. A method for suppressing underwater acoustic signal noise by combining a dual-modal neural network and spectral subtraction according to claim 3, characterized in that, The cross-domain Transformer encoder includes a self-attention layer and a cross-domain attention layer. The first self-attention layer is connected to the third self-attention layer after passing through the first cross-domain attention layer, the second self-attention layer, and the second cross-domain attention layer in sequence; the fourth self-attention layer is connected to the sixth self-attention layer after passing through the third cross-domain attention layer, the fifth self-attention layer, and the fourth cross-domain attention layer in sequence; the output of the first self-attention layer also serves as the input of the third cross-domain attention layer, and the output of the second self-attention layer also serves as the input of the fourth cross-domain attention layer; the output of the fourth self-attention layer also serves as the input of the first cross-domain attention layer, and the output of the fifth self-attention layer also serves as the input of the second cross-domain attention layer. The outputs of the third self-attention layer and the sixth self-attention layer respectively serve as the two outputs of the cross-domain Transformer encoder.

5. A method for suppressing underwater acoustic signal noise by combining a dual-modal neural network and spectral subtraction according to claim 3, characterized in that The time-domain encoder includes a first one-dimensional convolutional layer, a first normalization layer, a first gelu activation layer, and a first residual block connected in sequence. The input of the time-domain encoder serves as the input of the first one-dimensional convolutional layer, and the output of the first residual block serves as the output of the time-domain encoder; the time-domain decoder includes a second one-dimensional convolutional layer, a second normalization layer, a second gelu activation layer, and a second residual block connected in sequence. The two inputs of the time-domain decoder are fused and then serve as the input of the second one-dimensional convolutional layer, and the output of the second residual block serves as the output of the time-domain decoder.

6. A method for suppressing underwater acoustic signal noise by combining a dual-modal neural network and spectral subtraction according to claim 3, characterized in that The spectral-domain encoder includes a first two-dimensional convolutional layer, a third normalization layer, a third gelu activation layer, and a third residual block connected in sequence. The input of the spectral-domain encoder serves as the input of the first two-dimensional convolutional layer, and the output of the third residual block serves as the output of the spectral-domain encoder; the spectral-domain decoder includes a second two-dimensional convolutional layer, a fourth normalization layer, a fourth gelu activation layer, and a fourth residual block connected in sequence. The two inputs of the spectral-domain decoder are fused and then serve as the input of the second two-dimensional convolutional layer, and the output of the fourth residual block serves as the output of the spectral-domain decoder.

7. A method for suppressing underwater acoustic signal noise by combining a dual-modal neural network and spectral subtraction according to claim 1, characterized in that, The specific content of step 3 is as follows: Step 3.1: Perform short-time Fourier transform on the estimated platform self-noise and the noisy underwater acoustic signal to be processed respectively, and obtain the complex time-frequency spectral values corresponding to the estimated platform noise and the noisy underwater acoustic signal respectively; Step 3.2: Generate the real-time signal-to-noise ratio of the two on the frequency axis according to the complex time-frequency spectral values corresponding to the estimated platform noise and the noisy underwater acoustic signal, and generate a dynamic subtraction factor matrix α according to the real-time signal-to-noise ratio of the two on the frequency axis; Step 3.3: Combine the subtraction factor matrix α and use the following formula to calculate and obtain the complex time-frequency spectral values corresponding to the denoised underwater acoustic signal: Among them, represents the complex time-frequency spectrum amplitude value of the denoised underwater acoustic signal, max(,) represents the maximum value operation, and X(f,t) represents the complex time-frequency spectrum value of the noisy underwater acoustic signal, represents the complex time-frequency spectrum value of the estimated platform noise, β represents the noise threshold; || represents the amplitude operation; Step 3.4: The complex time-frequency spectrum values corresponding to the denoised underwater acoustic signal After performing the inverse short-time Fourier transform, the denoised underwater acoustic signal is obtained.

8. An underwater acoustic signal noise suppression device combining a dual-modal neural network and spectral subtraction, characterized in that, including: A dataset construction unit for constructing a simulated underwater acoustic dataset with self-noise according to the self-noise signal of the receiving platform; A neural network storage unit for storing a dual-modal neural network; A training unit for training the dual-modal neural network using the simulated underwater acoustic dataset with self-noise until the training is completed to obtain a platform self-noise estimation model; A signal denoising unit for inputting the noisy underwater acoustic signal to be processed into the platform self-noise estimation model to obtain the estimated platform self-noise; After estimating the platform self-noise and using the improved spectral subtraction method to denoise the noisy underwater acoustic signal to be processed, a denoised underwater acoustic signal is obtained.

9. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method for suppressing underwater acoustic signal noise by combining a dual-modal neural network and spectral subtraction according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the method for suppressing underwater acoustic signal noise by combining a dual-modal neural network and spectral subtraction according to any one of claims 1 to 7.

Images