Low snr underwater acoustic spectrum enhancement system based on deep learning

Abstract

The present application relates to underwater acoustic signal processing, and in particular to a low signal-to-noise ratio underwater acoustic spectrum enhancement system based on deep learning, a learnable comb filter module, a series of frequency-fixed learnable multi-channel comb filters are constructed through a convolution layer to scan the input signal, enhance the target frequency and attenuate the non-target frequency; a channel fusion module, feature extraction and feature fusion are performed on the output signal of the multi-channel comb filter; a frequency estimation module, the line spectrum component of the target frequency is identified from the feature map output by the channel fusion module, a learnable threshold parameter is introduced to screen the frequency bin, and the frequency index of each frame is generated; a mask enhancement module, the input signal is adaptively filtered using the frequency index sequence output by the frequency estimation module; the technical scheme provided by the present application can effectively overcome the defects of the prior art, such as insufficient line spectrum enhancement capability in a complex marine noise environment.

Description

Technical Field

[0001] This invention relates to underwater acoustic signal processing, and more specifically to a low signal-to-noise ratio underwater acoustic line spectrum enhancement system based on deep learning. Background Technology

[0002] In recent years, with the rapid development of marine exploration technology, underwater acoustic signal processing has been increasingly widely used in target detection, identification, and tracking. Underwater acoustic signals typically contain narrowband line spectrum components, which originate from the radiated noise of targets such as submarines and ships, and possess significant feature identification value. However, in real marine environments, underwater acoustic signals are not only frequently interfered with by background radiated noise, resulting in extremely low signal-to-noise ratios (SNR < -15dB), but also face non-stationary noise fluctuations and Doppler frequency shift (frequency drift). These complex factors intertwine, making the extraction and enhancement of the line spectrum exceptionally difficult, directly limiting the accuracy and real-time performance of subsequent processing.

[0003] Existing line spectrum extraction and enhancement methods mainly include power spectrum analysis, LOFAR graph processing, adaptive line spectrum enhancement (ALE), and deep learning. These methods can achieve certain results under specific conditions, but they generally have significant limitations.

[0004] First, power spectrum analysis methods calculate the power spectrum (periodogram) of the received signal by filtering and analytical transformation, and compare it with a preset threshold to determine the presence of line spectra. To improve the signal-to-noise ratio (SNR), methods such as extending the integration time (averaging across frames) or increasing the threshold are often used. However, in low SNR environments, while extending the integration time can improve the SNR, it sacrifices time resolution, leading to spectral peak broadening during frequency drift and even missed detections; increasing the threshold helps reduce the false alarm rate, but it lowers the detection probability.

[0005] Secondly, LOFAR graph processing methods convert continuous spectrum into a time-frequency plane (LOFER spectrum), utilizing multi-frame data to extract and enhance line spectra. They are often integrated with image processing (e.g., Hough transform for line detection) and target tracking (e.g., Viterbi algorithm for path optimization). The essence of these methods lies in relying on inter-frame line spectrum continuity to highlight features; that is, by associating the spectral values ​​of the current frame with those of several adjacent frames, they enhance continuous trajectories at specific frequencies, thereby suppressing noise and extracting line spectra. However, this core mechanism of frame-to-frame relationships leads to highly intensive computation. Each new frame requires re-traversing and associating the entire time-frequency plane, causing redundant computation and difficulties in parallelization. Simultaneously, they are extremely sensitive to non-stationary signals, as noise fluctuations or sudden interference can disrupt trajectory continuity, resulting in increased false alarms or trajectory breaks. Furthermore, under low signal-to-noise ratios or frequency drift, the algorithm struggles to accurately capture curved trajectories or multi-line spectrum intersections, significantly reducing overall performance and failing to maintain high robustness in complex dynamic environments.

[0006] Secondly, the Adaptive Line Spectrum Enhancement (ALE) method relies on iteratively updating filter coefficients to highlight the line spectrum and suppress noise, but its performance is heavily dependent on the initial parameters and step size settings. At the same time, the algorithm has a slow convergence speed, is highly sensitive to hyperparameters such as delay and step size and requires manual adjustment, and has limited generalization ability. In addition, in low signal-to-noise ratio environments, the improvement in signal-to-noise ratio is insufficient, and the overall effect is difficult to compete with deep learning methods.

[0007] Finally, while existing deep learning methods draw on advancements in speech and image processing, they have significant limitations. For example, methods mimicking ALE (Alternating Lesion Spectrum) use noisy signals as input and delayed signals as labels to simulate ALE-enhanced line spectra end-to-end, but they suffer from unstable convergence in low signal-to-noise ratio environments, are sensitive to parameters, and have limited effectiveness. Most methods (such as CNN / Unet) take LOFAR spectra as input and utilize convolution to enhance frequency features based on intra / inter-frame relationships, but they essentially borrow from image processing paradigms, resulting in poor interpretability (unable to connect to physical mechanisms). Furthermore, the scarcity of existing underwater acoustic datasets leads to insufficient training data, further limiting the generalization performance of the models.

[0008] In general, existing line spectrum extraction and enhancement methods rely on statistical assumptions, fixed thresholds, or frame-to-frame relationships to enhance line spectrum features. For example, they improve the signal-to-noise ratio through multi-frame accumulation or correlation (such as power spectrum integration or LOFAR trajectory search), leading to computational redundancy and inefficiency. Furthermore, each new frame requires re-traversal and correlation of the entire time-frequency plane, making efficient parallelism difficult, and exhibiting poor robustness to frequency drift, non-stationary noise, and complex interference. These shortcomings result in insufficient line spectrum enhancement capabilities in complex marine noise environments, making it difficult to meet the needs of military or civilian underwater acoustic applications. Therefore, an interpretable, efficient, end-to-end deep learning solution is urgently needed. Summary of the Invention

[0009] (a) Technical problems to be solved In view of the above-mentioned shortcomings of the existing technology, the present invention provides a low signal-to-noise ratio underwater acoustic line spectrum enhancement system based on deep learning, which can effectively overcome the shortcomings of the existing technology in terms of insufficient line spectrum enhancement capability in complex marine noise environment.

[0010] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: A deep learning-based low signal-to-noise ratio underwater acoustic line spectrum enhancement system includes a dataset generation module, a learnable comb filter module, a channel fusion module, a frequency estimation module, and a mask enhancement module. The dataset generation module constructs a training dataset for underwater acoustic line spectrum enhancement tasks in low signal-to-noise ratio environments; The learnable comb filter module constructs a series of learnable multi-channel comb filters with fixed frequencies through convolutional layers to scan the input signal, enhance the target frequency, and attenuate non-target frequencies, so as to effectively locate and enhance the line spectrum components of the target frequency in a low signal-to-noise ratio environment. The channel fusion module enhances the expressive power of the target frequency and suppresses background noise by extracting and fusing features from the output signal of the multi-channel comb filter. The frequency estimation module identifies the line spectrum components of the target frequency from the feature map output by the channel fusion module, and filters the frequency bins by introducing a learnable threshold parameter to generate the frequency index for each frame. The mask enhancement module uses the frequency index sequence output by the frequency estimation module to adaptively filter the input signal, enhancing the line spectrum components of the target frequency while attenuating non-target frequencies, thus balancing signal enhancement and noise preservation.

[0011] Preferably, the dataset generation module constructs a training dataset for underwater acoustic line spectrum enhancement tasks in low signal-to-noise ratio environments, including: A clean signal is generated by simulating the Doppler frequency shift and random discontinuous gating of underwater acoustic signals; Noisy signals are generated by adding noise to clean signals; By constructing a soft mask, a negative gain is uniformly applied to the non-linear spectral region of the noisy signal to attenuate the non-linear spectral region. The positive gain of the linear spectral region in the noisy signal is adaptively calculated based on the local signal-to-noise ratio to enhance the linear spectral region, thereby generating a tag signal.

[0012] Preferably, the generation of a clean signal through Doppler frequency shift and random discontinuous gating of simulated underwater acoustic signals includes: S111, Doppler frequency shift and sinusoidal frequency modulation: By simulating frequency shifts caused by water flow, wind speed, or moving objects in a linear or nonlinear manner, and combining this with simulated periodic frequency fluctuations caused by waves, the clean signal frequency more closely approximates the non-stationary changes of underwater acoustic signals in a real marine environment. ; Among them, f n (t) is the instantaneous frequency of the nth line spectrum at time t, f 0,n Let k be the initial frequency of the nth line spectrum. n Let a be the slope of the linear frequency of the nth line spectrum. n Let A be the coefficient of the nonlinear Doppler quadratic term of the nth line spectrum. n , , These represent the modulation amplitude, modulation angular frequency, and initial modulation phase of the nth line spectrum, respectively. S112, Random Discontinuous Gating: This simulates the non-stationary characteristics of underwater acoustic signals in a real marine environment by randomly controlling the appearance and disappearance periods of clean signals, using a random gating function: ; Among them, g n (t) is the random gating function for the nth line spectrum, and on_time and off_time are the randomly set on and off times, respectively; S113. Clean Signal Synthesis: Based on Doppler frequency shift, sinusoidal frequency modulation, and random discontinuous gating, a clean signal (clean_signal) is synthesized. ; in, For the nth line spectrum at time... instantaneous frequency The integral from the initial time to time t, Let n be the initial phase of the nth line spectrum. This represents the total phase of the nth line spectrum at time t, where N is the number of line spectra and A is the amplitude scaling factor, which is adjusted according to a preset signal-to-noise ratio range to ensure signal diversity. The method of generating a noisy signal by adding noise to a clean signal includes: A noise signal is added to the clean signal clean_signal to simulate common interference signals in real marine environments, generating a noisy signal noised_signal: noised_signal=clean_signal+noise; The type of noise signal is determined according to the specific environment to simulate different underwater acoustic conditions; The process involves uniformly applying a negative gain to the non-line spectral regions of the noisy signal using a soft mask to attenuate these regions, while adaptively calculating a positive gain for the line spectral regions of the noisy signal based on the local signal-to-noise ratio to enhance them, thereby generating a tag signal. This includes: S131. Perform a Short Time Fourier Transform (STFT) on the noisy signal noised_signal to obtain the spectrum Y: Y = STFT(noised_signal); S132. Construct a frequency domain masked filter and initialize the frequency domain mask M with wideband uniform attenuation. S133. For each line spectrum, iterate through each frame and determine whether the frequency of that frame is a line spectrum frequency: Line spectrum frequency bin: In the spectrum Y, if the frequency bin belongs to the line spectrum frequency range, then the frequency bin is regarded as the line spectrum frequency bin, and the positive gain is adaptively calculated according to the local signal-to-noise ratio. At the same time, the corresponding frequency domain mask M is adjusted to enhance the line spectrum region. Non-linear frequency bin: In the spectrum Y, if the frequency bin does not belong to the line spectrum frequency range, then the frequency bin is regarded as a non-linear frequency bin, and the corresponding frequency domain mask M is maintained to attenuate the non-linear region; S134. Generate the tag signal target_signal based on the spectrum Y and the frequency domain mask M: ; in, This indicates element-wise multiplication.

[0013] Preferably, the learnable comb filter module constructs a series of learnable multi-channel comb filters with fixed frequencies through convolutional layers to scan the input signal, enhance the target frequency, and attenuate non-target frequencies, so as to effectively locate and enhance the line spectrum components of the target frequency in a low signal-to-noise ratio environment, including: For each output channel in a multi-channel comb filter, the convolution operation is as follows: ; Among them, y i [t] represents the convolution result of the i-th output channel at time t, where i = 1, 2, ..., C, C is the number of output channels, and h i [n] represents the weight value of the convolution kernel at position n for the i-th output channel. For single-channel input signals, This indicates index adjustment, ensuring that the convolution kernel slides symmetrically around time t, where L is the length of the convolution kernel.

[0014] Preferably, the design of the multi-channel comb filter includes: 1) Sampling rate and fundamental frequency settings: Set the sampling rate to f s And define the fundamental frequency set F={f1,f2,…,f C Each base frequency corresponds to an independent output channel; 2) Period: The period for each output channel is: ; Where, p i f is the period of the i-th output channel. i Let be the fundamental frequency of the i-th output channel; 3) Convolution kernel length L: To ensure that each comb filter can cover the lowest frequency f min The period, and the kernel length L are: ; Where s is the half-window size of the output channel; 4) Number of teeth in half-window: For the i-th output channel, the number of teeth K that can be placed within the kernel length L. i for: ; 5) Window function: ; in, Let be the weighting coefficient of the window function of the i-th output channel at the m-th window position. ; 6) Convolution kernel initialization: ; in, Let be the unit impulse function, and c be the geometric center index of the convolution kernel. , This is the floor function.

[0015] Preferably, the channel fusion module enhances the expressive power of the target frequency and suppresses background noise by performing feature extraction and feature fusion on the output signal of the multi-channel comb filter, including: S31. Feature Extraction: Multiple convolutional layers are used to convolve the output signal X of the multi-channel comb filter to capture local features in the frequency domain. Higher-order features are extracted progressively through multiple convolutional layers to reduce noise interference. The feature map Y output by the l-th convolutional layer... l for: ; Among them, W l b l These represent the kernel weights and biases of the l-th convolutional layer, respectively. Represents two-dimensional convolution; S32, Feature Fusion: By gradually reducing the number of channels through multiple convolutions, a single-channel fusion feature map Y' is finally generated for low-dimensional feature representation; Each convolutional layer includes: A two-dimensional convolutional layer performs convolution operations with a kernel size of 3*3 to extract local features. Batch normalization layers accelerate training and suppress overfitting; Activation functions introduce nonlinearity and enhance the model's fitting ability.

[0016] Preferably, the frequency estimation module identifies the line spectral components of the target frequency from the feature map output by the channel fusion module, and filters the frequency bins by introducing a learnable threshold parameter to generate a frequency index for each frame, including: S41. Introduce a learnable threshold parameter. This parameter is optimized during training through backpropagation. For the j-th frame, the feature values ​​of the frequency dimension are compared with the threshold parameter. Compare and filter parameters that are greater than the threshold. The frequency bin is used to output the frequency index corresponding to the line spectrum frequency. ; Among them, I j Let j be the set of frequency indices of the j-th frame, representing the frequency dimension feature values ​​in the j-th frame that are greater than a threshold parameter. The position of the frequency bin is used to locate the target line spectrum frequency. Let f be the feature value of the f-th frequency bin in the j-th frame, where f is the frequency dimension index, f=1,2,…,F', and F' is the frequency dimension size of the fused feature map Y'. S42. Integrate the frequency index set of each frame into a complete frequency index sequence I=[I1,I2,…,I…] T ] indicates the input signal The distribution of line spectrum frequencies in the time dimension, where T is the size of the time dimension of the fused feature map Y'.

[0017] Preferably, the threshold parameter As a learnable parameter, it is optimized together with other parts of the system through the first loss function to ensure that the frequency index can be adaptively selected; The first loss function is designed according to task requirements, with the goal of minimizing frequency estimation error. For the underwater acoustic line spectrum enhancement task, the first loss function Loss1 based on frequency index is used: ; in, This is a reference value for the set of real frequency indices corresponding to the j-th frame.

[0018] Preferably, the mask enhancement module uses the frequency index sequence output by the frequency estimation module to adaptively filter the input signal, enhancing the line spectrum components of the target frequency while attenuating non-target frequencies, balancing signal enhancement and noise preservation, including: S51, Mask Generation: For the j-th frame, based on the frequency index set I... j Generate the corresponding mask: ; Among them, M j (f) is the mask for the f-th frequency bin in the j-th frame. This indicates that the values ​​do not belong to the frequency index set I. jThe frequency bin provides -3dB attenuation, offering smoother noise suppression; S52, Adaptive Filtering Enhancement: Using Mask M j (f) Filter the fused feature map Y' frame by frame to obtain the enhanced signal: ; Among them, X j (f) represents the enhanced signal of the f-th frequency bin in the j-th frame; S53. Result Integration: Integrate the enhanced signals from each frame into a complete enhanced signal, representing the input signal. The enhancement results in the time dimension are used for subsequent signal processing or to output the final enhanced signal.

[0019] Preferably, the mask generation depends on a learnable threshold parameter. By setting a second loss function for backpropagation optimization, it is ensured that the mask can accurately filter the target frequency bin; The second loss function is designed according to task requirements to optimize the enhancement effect. For the goal of maximizing the signal-to-noise ratio of the underwater acoustic line spectrum, the following second loss function, Loss2, is used: ; in, Let f be the frequency component of the real target signal corresponding to the f-th frequency bin in the j-th frame.

[0020] (III) Beneficial Effects Compared with existing technologies, the low signal-to-noise ratio underwater acoustic line spectrum enhancement system based on deep learning provided by this invention has the following beneficial effects: 1) The core of this invention lies in the use of a multi-channel FIR filter structure. The convolutional kernel of each output channel is initialized through a pulse train and combined with the multi-layer stacking of CNN (gradually reducing the number of channels), it realizes hierarchical processing from coarse-grained frequency scanning to fine-grained feature fusion. This structure enables the filter to adaptively match the line spectrum frequency of the input signal, avoiding the sensitivity of traditional filters to frequency drift. At the same time, the learnable threshold parameter of the frequency estimation module and the attenuation mechanism of the mask enhancement module form a closed-loop feedback, ensuring that non-target frequencies are effectively processed, improving the robustness of the system and the naturalness of the signal. Compared with the prior art, the inter-module relationship of this invention is closer, realizing end-to-end optimization and improving the overall signal processing efficiency. 2) From the perspective of signal processing theory, this invention utilizes the frequency response of a comb filter to form a passband at the matching frequency, which theoretically can amplify the power of the target line spectrum. The convolution operation of the channel fusion module reduces noise propagation through local feature extraction and activation functions. In addition, the introduction of a learnable threshold parameter conforms to the adaptive detection theory and can dynamically adapt to changes in signal-to-noise ratio. The mask attenuation is based on the amplitude formula, which theoretically balances enhancement and risk and avoids signal distortion caused by hard thresholds. Compared with existing fixed threshold methods, this invention has a higher theoretical enhancement ratio in low signal-to-noise ratio environments, thus improving the accuracy of line spectrum detection. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0022] Figure 1 This is a schematic diagram of the system of the present invention; Figure 2 This is a schematic diagram of parameter initialization for the multi-channel comb filter in this invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0024] The following describes the specific functional modules of the deep learning-based low signal-to-noise ratio underwater acoustic line spectrum enhancement system provided by this invention, using concrete examples (such as...). Figure 1 (As shown). The system's functional modules include: a dataset generation module, a learnable comb filter module, a channel fusion module, a frequency estimation module, and a mask enhancement module; The dataset generation module constructs a training dataset for underwater acoustic line spectrum enhancement tasks in low signal-to-noise ratio environments; The learnable comb filter module constructs a series of learnable multi-channel comb filters with fixed frequencies through convolutional layers to scan the input signal, enhance the target frequency, and attenuate non-target frequencies, so as to effectively locate and enhance the line spectrum components of the target frequency in a low signal-to-noise ratio environment. The channel fusion module enhances the expressive power of the target frequency and suppresses background noise by extracting and fusing features from the output signal of the multi-channel comb filter. The frequency estimation module identifies the line spectrum components of the target frequency from the feature map output by the channel fusion module, and filters the frequency bins by introducing a learnable threshold parameter to generate the frequency index for each frame. The mask enhancement module uses the frequency index sequence output by the frequency estimation module to adaptively filter the input signal, enhancing the line spectrum components of the target frequency while attenuating non-target frequencies, thus balancing signal enhancement and noise preservation.

[0025] I. Dataset Generation Module Due to the scarcity of underwater acoustic datasets and the presence of non-stationary noise (such as Doppler effect, wave modulation, random discontinuity, etc.), a dataset generation module was designed to simulate the noise and signal characteristics of real marine environments. This module aims to construct a training dataset for underwater acoustic line spectrum enhancement tasks in low signal-to-noise ratio environments, including: A clean signal is generated by simulating the Doppler frequency shift and random discontinuous gating of underwater acoustic signals; Noisy signals are generated by adding noise to clean signals; By constructing a soft mask, a negative gain is uniformly applied to the non-linear spectral region of the noisy signal to attenuate the non-linear spectral region. The positive gain of the linear spectral region in the noisy signal is adaptively calculated based on the local signal-to-noise ratio to enhance the linear spectral region, thereby generating a tag signal.

[0026] Generating clean signals through Doppler frequency shifting and random discontinuous gating of simulated underwater acoustic signals, including: S111, Doppler frequency shift and sinusoidal frequency modulation: By simulating frequency shifts caused by water flow, wind speed, or moving objects (such as submarines, ships, etc.) in a linear or nonlinear manner, and combining this with simulated periodic frequency fluctuations caused by waves, the clean signal frequency more closely approximates the non-stationary changes of underwater acoustic signals in a real marine environment. ; Among them, f n (t) is the instantaneous frequency of the nth line spectrum at time t, f 0,n Let k be the initial frequency of the nth line spectrum. n Let a be the slope of the linear frequency of the nth line spectrum. n Let A be the coefficient of the nonlinear Doppler quadratic term of the nth line spectrum. n , , These represent the modulation amplitude, modulation angular frequency, and initial modulation phase of the nth line spectrum, respectively. S112, Random Discontinuous Gating: This simulates the non-stationary characteristics of underwater acoustic signals in a real marine environment by randomly controlling the appearance and disappearance periods of clean signals, using a random gating function: ; Among them, g n (t) is the random gating function for the nth line spectrum, and on_time and off_time are the randomly set on and off times, respectively; S113. Clean Signal Synthesis: Based on Doppler frequency shift, sinusoidal frequency modulation, and random discontinuous gating, a clean signal (clean_signal) is synthesized. ; in, For the nth line spectrum at time... instantaneous frequency The integral from the initial time to time t, Let n be the initial phase of the nth line spectrum. This represents the total phase of the nth line spectrum at time t, where N is the number of line spectra and A is the amplitude scaling factor, which is adjusted according to a preset signal-to-noise ratio range to ensure signal diversity.

[0027] Generating noisy signals by adding noise to clean signals includes: Add a noise signal (such as Gaussian white noise with energy of 1) to the clean signal clean_signal to simulate common interference signals in real marine environments, generating a noisy signal noised_signal: noised_signal=clean_signal+noise; The type of noise signal is determined based on the specific environment to simulate different underwater acoustic conditions.

[0028] The method involves constructing a soft mask to uniformly apply a negative gain to the non-line spectral regions of the noisy signal to attenuate these regions. Meanwhile, a positive gain is adaptively calculated based on the local signal-to-noise ratio to enhance the line spectral regions, generating a tag signal (the purpose of tag signal generation is to generate an ideal target signal for each sample for training in underwater acoustic line spectral enhancement). This includes: S131. Perform a Short Time Fourier Transform (STFT) on the noisy signal noised_signal to obtain the spectrum Y: Y = STFT(noised_signal); S132. Construct a frequency domain mask filter and initialize the frequency domain mask M with a wideband uniform attenuation (e.g., give an attenuation of -3dB). S133. For each line spectrum, iterate through each frame and determine whether the frequency of that frame is a line spectrum frequency: Line spectrum frequency bin: In the spectrum Y, if the frequency bin belongs to the line spectrum frequency range, then the frequency bin is regarded as the line spectrum frequency bin, and the positive gain is adaptively calculated according to the local signal-to-noise ratio. At the same time, the corresponding frequency domain mask M is adjusted to enhance the line spectrum region. Non-linear frequency bin: In the spectrum Y, if the frequency bin does not belong to the line spectrum frequency range, then the frequency bin is regarded as a non-linear frequency bin, and the corresponding frequency domain mask M is maintained to attenuate the non-linear region; S134. Generate the tag signal target_signal based on the spectrum Y and the frequency domain mask M: ; in, This indicates element-wise multiplication.

[0029] II. Learnable Comb Filter Module In traditional underwater acoustic signal processing, common methods such as multi-frame data accumulation and inter-frame correlation can effectively enhance signals, but they introduce computational redundancy and difficulties in parallel processing. To address these issues, this invention proposes a learnable comb filter module that can directly enhance signals at specific frequencies, avoiding computational redundancy in traditional methods and improving processing efficiency.

[0030] The learnable comb filter module constructs a series of learnable multi-channel comb filters with fixed frequencies through convolutional layers (each output channel corresponds to an independent FIR filter). It scans the input signal, enhances the target frequency, and attenuates non-target frequencies to effectively locate and enhance the line spectral components of the target frequency in low signal-to-noise ratio environments. This includes: For each output channel in a multi-channel comb filter, the convolution operation is as follows: ; Among them, y i [t] represents the convolution result of the i-th output channel at time t, where i = 1, 2, ..., C, C is the number of output channels, and h i [n] represents the weight value of the convolution kernel at position n for the i-th output channel. For single-channel input signals, This indicates index adjustment, ensuring that the convolution kernel slides symmetrically around time t, where L is the length of the convolution kernel.

[0031] Specifically, the design of a multi-channel comb filter includes: 1) Sampling rate and fundamental frequency settings: Set the sampling rate to fs And define the fundamental frequency set F={f1,f2,…,f C Each base frequency corresponds to an independent output channel; 2) Period: The period for each output channel is: ; Where, p i f is the period of the i-th output channel. i Let be the fundamental frequency of the i-th output channel; 3) Convolution kernel length L: To ensure that each comb filter can cover the lowest frequency f min The period, and the kernel length L are: ; Where s is the half-window size of the output channel; 4) Number of teeth in half-window: For the i-th output channel, the number of teeth K that can be placed within the kernel length L. i for: ; 5) Window function (Hanning window): ; in, Let be the weighting coefficient of the window function of the i-th output channel at the m-th window position. ; 6) Convolution kernel initialization: ; in, Let be the unit impulse function, and c be the geometric center index of the convolution kernel. , This is the floor function.

[0032] III. Channel Fusion Module The channel fusion module enhances the expressive power of the target frequency and suppresses background noise by extracting and fusing features from the output signal of the multi-channel comb filter, including: S31. Feature Extraction: Multiple convolutional layers are used to convolve the output signal X of the multi-channel comb filter to capture local features in the frequency domain. Higher-order features are extracted progressively through multiple convolutional layers to reduce noise interference. The feature map Y output by the l-th convolutional layer... l for: ; Among them, W l b l These represent the kernel weights and biases of the l-th convolutional layer, respectively. Represents two-dimensional convolution; S32, Feature Fusion: Gradually reduce the number of channels through multiple convolutional layers (the number of channels changes by...) Finally, a single-channel fusion feature map Y' is generated for low-dimensional feature representation; Each convolutional layer includes: A two-dimensional convolutional layer performs convolution operations with a kernel size of 3*3 to extract local features. Batch normalization layers accelerate training and suppress overfitting; Activation functions introduce nonlinearity and enhance the model's fitting ability.

[0033] If the above technical solutions do not use CNN, methods such as Transformer or Graph Neural Network (GNN) can be considered to capture the complex relationships between frequency channels through self-attention mechanisms or graph convolution.

[0034] IV. Frequency Estimation Module The frequency estimation module identifies the line spectral components of the target frequency from the feature map output by the channel fusion module. It then filters the frequency bins by introducing a learnable threshold parameter, generating a frequency index for each frame, including: S41. Introduce a learnable threshold parameter. This parameter is optimized during training through backpropagation. For the j-th frame, the feature values ​​of the frequency dimension are compared with the threshold parameter. Compare and filter parameters that are greater than the threshold. The frequency bin is used to output the frequency index corresponding to the line spectrum frequency. ; Among them, I j Let j be the set of frequency indices of the j-th frame, representing the frequency dimension feature values ​​in the j-th frame that are greater than a threshold parameter. The position of the frequency bin is used to locate the target line spectrum frequency. Let f be the feature value of the f-th frequency bin in the j-th frame, where f is the frequency dimension index, f=1,2,…,F', and F' is the frequency dimension size of the fused feature map Y'. S42. Integrate the frequency index set of each frame into a complete frequency index sequence I=[I1,I2,…,I…] T ] indicates the input signal The distribution of line spectrum frequencies in the time dimension, where T is the size of the time dimension of the fused feature map Y'.

[0035] In the technical solution of this application, the threshold parameter As a learnable parameter, it is optimized together with other parts of the system through the first loss function to ensure that the frequency index can be adaptively selected; The first loss function is designed based on task requirements, with the goal of minimizing the frequency estimation error. For the underwater acoustic line spectrum enhancement task, the first loss function Loss1 based on the frequency index is used: ; in, This is a reference value for the set of real frequency indices corresponding to the j-th frame.

[0036] V. Mask Enhancement Module The mask enhancement module uses the frequency index sequence output by the frequency estimation module to adaptively filter the input signal, enhancing the line spectrum components of the target frequency while attenuating non-target frequencies, balancing signal enhancement and noise preservation, including: S51, Mask Generation: For the j-th frame, based on the frequency index set I... j Generate the corresponding mask: ; Among them, M j (f) is the mask for the f-th frequency bin in the j-th frame. This indicates that the values ​​do not belong to the frequency index set I. j The frequency bin provides -3dB attenuation (incomplete suppression), offering smoother noise suppression; S52, Adaptive Filtering Enhancement: Using Mask M j (f) Filter the fused feature map Y' frame by frame to obtain the enhanced signal: ; Among them, X j (f) represents the enhanced signal of the f-th frequency bin in the j-th frame; S53. Result Integration: Integrate the enhanced signals from each frame into a complete enhanced signal, representing the input signal. The enhancement results in the time dimension are used for subsequent signal processing or to output the final enhanced signal.

[0037] In the technical solution of this application, mask generation relies on a learnable threshold parameter. By setting a second loss function for backpropagation optimization, it is ensured that the mask can accurately filter the target frequency bin; The second loss function is designed according to task requirements to optimize the enhancement effect. For the goal of maximizing the signal-to-noise ratio of the underwater acoustic line spectrum, the following second loss function, Loss2, is used: ; in, Let f be the frequency component of the real target signal corresponding to the f-th frequency bin in the j-th frame.

[0038] The embodiments of this invention are based on an end-to-end neural network architecture to achieve line spectrum enhancement of low signal-to-noise ratio underwater acoustic signals. The following details the system deployment, parameter settings, and implementation steps, which are consistent with the line spectrum localization and enhancement problem in low signal-to-noise ratio environments addressed by this invention. The embodiments utilize a Python / PyTorch environment, suitable for signal processing devices (such as DSP processors, GPU servers, etc.), and each embodiment focuses on the beneficial effects of adaptive filtering and noise suppression.

[0039] The system deployment of the technical solution in this application adopts a pipeline architecture of "data input - module processing - output enhancement", and its specific components are as follows: 1) Data Input Layer Responsible for loading the original signal, supporting sampling rate f s A time-domain signal x[n] (length 1024~8192) with a frequency of 1000~32000Hz is preprocessed (such as maximum value normalization or zero mean normalization) to generate an input time-domain signal tensor, which supports batch processing to adapt to real-time input. 2) Neural network layer (core processing module) Deployed on compute nodes (such as NVIDIA GPUs), it includes a learnable comb filter module, a channel fusion module, and a frequency estimation module (with learnable threshold parameters). Initialized to 0.5) and with a mask enhancement module, the model has approximately 500K total parameters and supports end-to-end training; 3) Output layer It generates an enhanced signal and converts it to a time-domain output via inverse STFT, supporting real-time applications (such as audio device interfaces).

[0040] The specific implementation steps of the technical solution of this application include: Step 1: Environment Modeling and Initialization Load the original signal and initialize parameters: Set the sampling rate f s =200Hz, target fundamental frequency set F={50Hz,100Hz,150Hz,…,500Hz}, calculate the period of each output channel, and initialize the convolution kernel; Step 2: Train the model The first stage is offline training. An end-to-end model is trained on a GPU using a simulated dataset (containing low signal-to-noise ratio line spectrum signals, SNR = [-25dB:15dB]). The input is the original signal x[n], and the total loss function is... The optimizer Adam (learning rate 0.001) is trained for 200 epochs and supports online fine-tuning. In practical signal processing, the module incrementally updates the learnable threshold parameters using new data (such as real-time acquired signals). and masking to ensure adaptation to dynamic noise; Step 3: Real-time signal processing Based on the input signal, processing is performed once per frame (256 samples per frame, 128 steps): multi-channel comb filter scans the frequency, channel fusion reduces the dimensionality features, frequency estimation filters the index, and finally enhances the signal. The enhanced signal is then used for downstream tasks (such as frequency tracking).

[0041] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A low signal-to-noise ratio underwater acoustic line spectrum enhancement system based on deep learning, characterized in that: It includes a dataset generation module, a learnable comb filter module, a channel fusion module, a frequency estimation module, and a mask enhancement module; The dataset generation module constructs a training dataset for underwater acoustic line spectrum enhancement tasks in low signal-to-noise ratio environments; The learnable comb filter module constructs a series of learnable multi-channel comb filters with fixed frequencies through convolutional layers to scan the input signal, enhance the target frequency, and attenuate non-target frequencies, so as to effectively locate and enhance the line spectrum components of the target frequency in a low signal-to-noise ratio environment. The channel fusion module enhances the expressive power of the target frequency and suppresses background noise by extracting and fusing features from the output signal of the multi-channel comb filter. The frequency estimation module identifies the line spectrum components of the target frequency from the feature map output by the channel fusion module, and filters the frequency bins by introducing a learnable threshold parameter to generate the frequency index for each frame. The mask enhancement module uses the frequency index sequence output by the frequency estimation module to adaptively filter the input signal, enhance the line spectrum components of the target frequency, and attenuate the non-target frequencies, thus balancing signal enhancement and noise preservation. The learnable comb filter module constructs a series of learnable multi-channel comb filters with fixed frequencies through convolutional layers to scan the input signal, enhance the target frequency, and attenuate non-target frequencies, so as to effectively locate and enhance the line spectrum components of the target frequency in a low signal-to-noise ratio environment, including: For each output channel in a multi-channel comb filter, the convolution operation is as follows: ； Among them, y i [t] represents the convolution result of the i-th output channel at time t, where i = 1, 2, ..., C, C is the number of output channels, and h i [n] represents the weight value of the convolution kernel at position n for the i-th output channel. For single-channel input signals, This indicates index adjustment, ensuring that the convolution kernel slides symmetrically around time t, where L is the length of the convolution kernel; The design of the multi-channel comb filter includes: 1) Sampling rate and fundamental frequency settings: Set the sampling rate to f s And define the fundamental frequency set F={f1,f2,…,f C Each base frequency corresponds to an independent output channel; 2) Period: The period for each output channel is: ； Where, p i For the period of the i-th output channel, f i Let be the fundamental frequency of the i-th output channel; 3) Convolution kernel length L: To ensure that each comb filter can cover the lowest frequency f min The period, and the kernel length L are: ； Where s is the half-window size of the output channel; 4) Number of teeth in half-window: For the i-th output channel, the number of teeth K that can be placed within the kernel length L. i for: ； 5) Window function: ； in, Let be the weighting coefficient of the window function of the i-th output channel at the m-th window position. ; 6) Convolution kernel initialization: ； in, Let be the unit impulse function, and c be the geometric center index of the convolution kernel. , It is a rounding function; The frequency estimation module identifies the line spectral components of the target frequency from the feature map output by the channel fusion module, and filters the frequency bins by introducing a learnable threshold parameter to generate a frequency index for each frame, including: S41. Introduce a learnable threshold parameter. This parameter is optimized during training through backpropagation. For the j-th frame, the feature values ​​of the frequency dimension are compared with the threshold parameter. Compare and filter parameters that are greater than the threshold. The frequency bin is used to output the frequency index corresponding to the line spectrum frequency. ； Among them, I j Let j be the set of frequency indices of the j-th frame, representing the frequency dimension feature values ​​in the j-th frame that are greater than a threshold parameter. The position of the frequency bin is used to locate the target line spectrum frequency. Let f be the feature value of the f-th frequency bin in the j-th frame, where f is the frequency dimension index, f=1,2,…,F', and F' is the frequency dimension size of the fused feature map Y'. S42. Integrate the frequency index set of each frame into a complete frequency index sequence I=[I1,I2,…,I…] T ] indicates the input signal The distribution of line spectrum frequencies in the time dimension, where T is the size of the time dimension of the fused feature map Y'.

2. The low signal-to-noise ratio underwater acoustic line spectrum enhancement system based on deep learning according to claim 1, characterized in that: The dataset generation module constructs a training dataset for underwater acoustic line spectrum enhancement tasks in low signal-to-noise ratio environments, including: A clean signal is generated by simulating the Doppler frequency shift and random discontinuous gating of underwater acoustic signals; Noisy signals are generated by adding noise to clean signals; By constructing a soft mask, a negative gain is uniformly applied to the non-linear spectral region of the noisy signal to attenuate the non-linear spectral region. The positive gain of the linear spectral region in the noisy signal is adaptively calculated based on the local signal-to-noise ratio to enhance the linear spectral region, thereby generating a tag signal.

3. The low signal-to-noise ratio underwater acoustic line spectrum enhancement system based on deep learning according to claim 2, characterized in that: The method of generating a clean signal by simulating the Doppler frequency shift and random discontinuous gating of underwater acoustic signals includes: S111, Doppler frequency shift and sinusoidal frequency modulation: By simulating frequency shifts caused by water flow, wind speed, or moving objects in a linear or nonlinear manner, and combining this with simulated periodic frequency fluctuations caused by waves, the clean signal frequency more closely approximates the non-stationary changes of underwater acoustic signals in a real marine environment. ； Among them, f n (t) is the instantaneous frequency of the nth line spectrum at time t, f 0,n Let k be the initial frequency of the nth line spectrum. n Let a be the slope of the linear frequency of the nth line spectrum. n Let A be the coefficient of the nonlinear Doppler quadratic term of the nth line spectrum. n , , These represent the modulation amplitude, modulation angular frequency, and initial modulation phase of the nth line spectrum, respectively. S112, Random Discontinuous Gating: This simulates the non-stationary characteristics of underwater acoustic signals in a real marine environment by randomly controlling the appearance and disappearance periods of clean signals, using a random gating function: ； Among them, g n (t) is the random gating function for the nth line spectrum, and on_time and off_time are the randomly set on and off times, respectively; S113. Clean Signal Synthesis: Based on Doppler frequency shift, sinusoidal frequency modulation, and random discontinuous gating, a clean signal (clean_signal) is synthesized. ； in, For the nth line spectrum at time... instantaneous frequency The integral from the initial time to time t, Let n be the initial phase of the nth line spectrum. This represents the total phase of the nth line spectrum at time t, where N is the number of line spectra and A is the amplitude scaling factor, which is adjusted according to a preset signal-to-noise ratio range to ensure signal diversity. The method of generating a noisy signal by adding noise to a clean signal includes: A noise signal is added to the clean signal clean_signal to simulate common interference signals in real marine environments, generating a noisy signal noised_signal: noised_signal=clean_signal+noise; The type of noise signal is determined according to the specific environment to simulate different underwater acoustic conditions; The process involves uniformly applying a negative gain to the non-line spectral regions of the noisy signal using a soft mask to attenuate these regions, while adaptively calculating a positive gain for the line spectral regions of the noisy signal based on the local signal-to-noise ratio to enhance them, thereby generating a tag signal. This includes: S131. Perform a Short Time Fourier Transform (STFT) on the noisy signal noised_signal to obtain the spectrum Y: Y = STFT(noised_signal); S132. Construct a frequency domain masked filter and initialize the frequency domain mask M with wideband uniform attenuation. S133. For each line spectrum, iterate through each frame and determine whether the frequency of that frame is a line spectrum frequency: Line spectrum frequency bin: In the spectrum Y, if the frequency bin belongs to the line spectrum frequency range, then the frequency bin is regarded as the line spectrum frequency bin, and the positive gain is adaptively calculated according to the local signal-to-noise ratio. At the same time, the corresponding frequency domain mask M is adjusted to enhance the line spectrum region. Non-linear frequency bin: In the spectrum Y, if the frequency bin does not belong to the line spectrum frequency range, then the frequency bin is regarded as a non-linear frequency bin, and the corresponding frequency domain mask M is maintained to attenuate the non-linear region; S134. Generate the tag signal target_signal based on the spectrum Y and the frequency domain mask M: ； in, This indicates element-wise multiplication.

4. The low signal-to-noise ratio underwater acoustic line spectrum enhancement system based on deep learning according to claim 1, characterized in that: The channel fusion module enhances the expressive power of the target frequency and suppresses background noise by extracting and fusing features from the output signal of the multi-channel comb filter, including: S31. Feature Extraction: Multiple convolutional layers are used to convolve the output signal X of the multi-channel comb filter to capture local features in the frequency domain. Higher-order features are extracted progressively through multiple convolutional layers to reduce noise interference. The feature map Y output by the l-th convolutional layer... l for: ； Among them, W l b l These represent the kernel weights and biases of the l-th convolutional layer, respectively. Represents two-dimensional convolution; S32, Feature Fusion: By gradually reducing the number of channels through multiple convolutions, a single-channel fusion feature map Y' is finally generated for low-dimensional feature representation; Each convolutional layer includes: A two-dimensional convolutional layer performs convolution operations with a kernel size of 3*3 to extract local features. Batch normalization layers accelerate training and suppress overfitting; Activation functions introduce nonlinearity and enhance the model's fitting ability.

5. The low signal-to-noise ratio underwater acoustic line spectrum enhancement system based on deep learning according to claim 1, characterized in that: The threshold parameter As a learnable parameter, it is optimized together with other parts of the system through the first loss function to ensure that the frequency index can be adaptively selected; The first loss function is designed based on task requirements, with the goal of minimizing frequency estimation error. For the underwater acoustic line spectrum enhancement task, the first loss function Loss1 based on frequency index is used: ； in, This is a reference value for the set of real frequency indices corresponding to the j-th frame.

6. The low signal-to-noise ratio underwater acoustic line spectrum enhancement system based on deep learning according to claim 1, characterized in that: The mask enhancement module uses the frequency index sequence output by the frequency estimation module to adaptively filter the input signal, enhancing the line spectrum components of the target frequency while attenuating non-target frequencies, balancing signal enhancement and noise preservation, including: S51, Mask Generation: For the j-th frame, based on the frequency index set I... j Generate the corresponding mask: ； Among them, M j (f) is the mask for the f-th frequency bin in the j-th frame. This indicates that the values ​​do not belong to the frequency index set I. j The frequency bin provides -3dB attenuation, offering smoother noise suppression; S52, Adaptive Filtering Enhancement: Using Mask M j (f) Filter the fused feature map Y' frame by frame to obtain the enhanced signal: ； Among them, X j (f) represents the enhanced signal of the f-th frequency bin in the j-th frame; S53. Result Integration: Integrate the enhanced signals from each frame into a complete enhanced signal, representing the input signal. The enhancement results in the time dimension are used for subsequent signal processing or to output the final enhanced signal.

7. The low signal-to-noise ratio underwater acoustic line spectrum enhancement system based on deep learning according to claim 6, characterized in that: The mask generation depends on a learnable threshold parameter. By setting a second loss function for backpropagation optimization, it is ensured that the mask can accurately filter the target frequency bin; The second loss function is designed according to task requirements to optimize the enhancement effect. For the goal of maximizing the signal-to-noise ratio of the underwater acoustic line spectrum, the following second loss function, Loss2, is used: ； in, Let f be the frequency component of the real target signal corresponding to the f-th frequency bin in the j-th frame.

Images