A deep learning based adaptive line spectral enhancement method

Abstract

The application provides a kind of adaptive line spectrum enhancement method based on deep learning, belong to signal processing field.The adaptive line spectrum enhancement method specifically includes: generating target sound signal sample data;Target sound signal sample data normalization processing;Construction dataset;Construction fully connected neural network model;Training fully connected neural network model;Using fully connected neural network model in actual signal processing Effect verification.The application replaces the iterative process of traditional adaptive line spectrum enhancement by the method of deep learning, realizes the iteration of weight under low signal-to-noise ratio;Join the pretreatment means before fully connected neural network to improve the line spectrum enhancement effect, improve the processing effect of traditional adaptive line spectrum enhancement method under low signal-to-noise ratio.

Description

Technical Field

[0001] This invention relates to the field of signal processing, and more specifically to an adaptive line spectrum enhancement method. Background Technology

[0002] Conventional ALE utilizes the characteristic that the narrowband signal in the input signal and the narrowband signal in the delayed input signal remain correlated while the broadband noise is uncorrelated, to adaptively enhance the narrowband signal and suppress the broadband noise.

[0003] x(nk)=[x(nk),x(nk-1),…x(nk-(L-1))] T

[0004] y(n)=w T (n)x(nk)

[0005] e(n) = x(n) - y(n)

[0006] w(n+1)=w(n)+μe(n)x(nk)

[0007] In each iteration of adaptive line spectrum enhancement, the delayed input signal x(nk) is multiplied by the weighting coefficient w(n) to obtain the output signal at that moment. This output signal is then subtracted from the reference signal at that moment (i.e., the undelayed input signal) to obtain the error signal. The weighting coefficients are then updated using the LMS algorithm. From this, we can see that in each iteration, there is a corresponding delayed input signal x(nk) and an input signal x(n), and each delayed input signal corresponds to an output signal y(n), i.e., y(n) = f w (x(nk)) is then used, followed by the LMS algorithm to minimize the difference between x(n) and y(n) (eliminating noise as much as possible while preserving the signal). Based on this principle, a neural network can be used to replace this process. The neural network is trained using the input signal and a delayed input signal. The neural network can also minimize the difference between the input signal x(n) and the output signal y(n), achieving adaptive enhancement of narrowband signals while suppressing noise. This is the basic idea behind using neural networks for line spectrum enhancement. Furthermore, neural networks have strong nonlinear fitting capabilities, allowing them to better fit the relationship between the input signal and the delayed input signal under low signal-to-noise ratio (SNR) conditions, thus improving the processing performance of conventional ALE under low SNR conditions.

[0008] The conventional ALE method shows a significant deterioration in processing performance when the signal-to-noise ratio decreases. Therefore, it is necessary to further improve the processing gain of the conventional ALE method. Specifically, when the signal is subjected to strong noise interference, the traditional ALE method cannot accurately identify and enhance the target signal, resulting in a decline in signal quality and affecting subsequent signal processing and applications. Summary of the Invention

[0009] To address the aforementioned technical problems, this invention proposes a deep learning-based adaptive line spectrum enhancement method, belonging to the field of signal processing. This invention replaces the iterative process of traditional adaptive line spectrum enhancement with a deep learning method, achieving iterative weighting under low signal-to-noise ratio (SNR) conditions and ultimately improving the SNR of the line spectrum signal. Simultaneously, preprocessing techniques are added before the fully connected neural network to enhance the line spectrum enhancement effect. Even under low SNR conditions, this invention can accurately identify and enhance the target signal, improving the processing performance of traditional adaptive line spectrum enhancement (ALE) methods at low SNR.

[0010] An adaptive line spectrum enhancement method based on deep learning, the adaptive line spectrum enhancement method specifically includes the following steps:

[0011] Step 1: Generate target acoustic signal sample data;

[0012] Step 2: Normalization processing of target acoustic signal sample data;

[0013] Step 3, construct the dataset; the dataset includes a training dataset and a test dataset;

[0014] Step 4: Construct a fully connected neural network model;

[0015] Step 5: Train the fully connected neural network model;

[0016] Step 6: Verify the effectiveness using a fully connected neural network model in actual signal processing.

[0017] Furthermore, in step 1, target acoustic signal sample data is generated using simulation or sonar methods. The target acoustic signals acquired using simulation or sonar methods can ensure the diversity and authenticity of the target acoustic signals.

[0018] The simulation method uses MATLAB as the data tool, and the specific simulation settings for MATLAB are as follows:

[0019] First, the frequency, duration, and power of the target sound signal are preset; then, a white noise background is preset to simulate the real environment; next, the bandwidth and power of the noise signal are set.

[0020] The target acoustic signal is denoted as x(n), which includes the line spectrum signal S(n) and the noise signal N. noise (n);

[0021]

[0022] Where A is the amplitude of the line spectrum signal; f0 is the frequency of the line spectrum signal; T is the total duration of the line spectrum signal; f s The sampling frequency is denoted by n; n represents the nth sample. Let U be the random phase of the line spectrum signal, and let U be the random phase that satisfies the distribution U(0,2π).

[0023] Furthermore, in step 2, the normalization processing method for the target acoustic signal sample data is as follows:

[0024] To ensure a consistent convergence speed during training of the sample data, a minimum-maximum normalization method is used to normalize the target acoustic signal sample data. The normalized target acoustic signal... as follows:

[0025]

[0026] Where, min x and max x These represent the minimum and maximum values ​​of the target acoustic signal sequence, respectively.

[0027] Furthermore, in step 3, the dataset is constructed as follows:

[0028] The dataset will be the normalized target acoustic signal sample data, which includes a training set and a test set; the top T values ​​of the normalized target acoustic signal will be taken. s Time, i.e., T s *f s The signal of the number of sampling points is used as the training set data x T (n), the signal of the remaining sampling points is the test set data x. E (n).

[0029] Furthermore, in step 4, the fully connected neural network model includes an input layer, a hidden layer, and an output layer; the fully connected neural network model is used to replace the adaptive weight iteration process of ALE.

[0030] Furthermore, the input layer is the delay vector x of the training set data. T (nk):

[0031] x T (nk)=[x T (nk),x T (nk-1),…,x T (nk-(N-1)] T ,

[0032] Where k represents the pre-delay point, and the pre-delay is set to 1 second, i.e., 1*f s The number of digital sampling points, [·] T is the transpose of the matrix, and N is the order of the input layer;

[0033] x training set data T (n) and the delay vector x of the training set data T(nk) serves as the input to the fully connected neural network model; the data at the corresponding sampling points in the training set are the output signals of the fully connected neural network model.

[0034] Furthermore, in step 5, the training process of the fully connected neural network model includes:

[0035] Step 5.1: Perform initial training on the fully connected neural network model:

[0036] x training set data T If (n) is directly input into the fully connected neural network model, then the output signal y(n) of the fully connected neural network model is:

[0037] y(n)=f θ (x T (n))

[0038] Among them, f θ (·) represents the function mapping relationship of a fully connected neural network model;

[0039] Obtain the parameters of the trained fully connected neural network model

[0040]

[0041] Where N represents the order of the input layer; i is the sampling sequence number; x(i) is the input acoustic signal sampled at the i-th time; y(i) is the output acoustic signal sampled at the i-th time; L is the loss function; θ is the parameters of the fully connected neural network model;

[0042] The loss function L is:

[0043] L(x,y)=||xy|| 2

[0044] Where, |||·||| represents the squared error operator; x and y are the sampled input acoustic signal and the sampled output acoustic signal, respectively;

[0045] Step 5.2: Use the Adam algorithm to process the fully connected neural network model parameters θ obtained in Step 5.1. * Optimize:

[0046]

[0047] In the Adam algorithm, the gradient descent method is used to solve for the learning rate α, which is set to 0.001. β′ and β″ are the first and second moments of the initial estimated gradient, respectively, with the first and second moments set to 0.9 and 0.999. The Adam algorithm can adaptively adjust the learning rate without manual adjustment, effectively solve the gradient noise problem, and converge quickly, achieving good results.

[0048] The Adam algorithm applies the parameters θ of a fully connected neural network model. * The optimization process is as follows:

[0049] During the optimization process, letter subscripts all represent the current iteration number;

[0050] Step 5.2.1, given initial parameters: given learning rate α; given parameters β′ and β″; parameters β′ and β″ are used to estimate the first and second moments of the gradient; defined the optimization objective function f(θ) and the initial values ​​of the optimization parameters θ0;

[0051] Step 5.2.2: Set the initial values ​​t = 0, m0 = 0, v0 = 0, where t is the number of iterations, m is the first moment estimate of the gradient, and v is the second moment estimate;

[0052] Step 5.2.3, update time t+1→t, calculate the gradient g of the objective function. t Gradient of objective function g t Represented as: Among them, the first moment estimate m of the updated gradient t As follows: m t =β′m t-1 +(1-β′)g t Calculate the first-order gradient moment estimate after bias correction. Update the second moment estimate of the gradient and calculate the bias-corrected second moment estimate of the gradient.

[0053] v t =β″v t-1 +(1-β″)g t +g t

[0054]

[0055] Update parameters:

[0056]

[0057] Where ε = 10 -8 It is an infinitesimal quantity to ensure that the denominator is not zero;

[0058] Step 5.2.4: Repeat steps 5.2.1 to 5.2.4 iteratively. When the Adam algorithm has completely traversed the database more than 40 times, the optimal θ is finally obtained. * ;

[0059] Step 5.3: In the training of the fully connected neural network model, activation functions are added between the input layer, hidden layer, and output layer of the fully connected neural network model to ensure the non-linear transformation of the fully connected neural network model.

[0060] The modified linear activation function ReLU(z) is used as the activation function for the fully connected neural network model.

[0061]

[0062] Where z is the calculation result of the fully connected neural network model during the forward propagation process;

[0063] When the input of the modified linear activation function is less than 0, the output of the fully connected neural network model is 0; the ReLU function obtains a sparse representation of the data, enabling the fully connected neural network model to accurately extract the relevant line spectrum frequencies in the acoustic signal; the ReLU function is more conducive to the fast and accurate descent of the gradient.

[0064] Furthermore, in step 6, the verification process of the fully connected neural network model in actual signal processing tasks is as follows:

[0065] Step 6.1: Test the trained fully connected neural network model:

[0066] The test set data is input into a fully connected neural network model for adaptive line spectrum enhancement. The method for adaptive line spectrum enhancement using a fully connected neural network model is called DLE (Deep Learning Line Enhancement). The enhancement result of DLE is the output result x of the fully connected neural network model. DLE (n) is:

[0067]

[0068] Step 6.2: Preprocess the tested fully connected neural network model to improve the signal-to-noise ratio:

[0069] The noise suppression gate (NSG) method is used as a preprocessing technique before inputting into a fully connected neural network model. The noise suppression gate method is as follows:

[0070] Test set data signal x E The autocorrelation function T of (n) x (n) is:

[0071] T x (n)=x E (-n)*x E (n)

[0072] The noise suppression gate I(n) is:

[0073]

[0074] Where, αT The zeroing time range is defined as n, where n is the nth sample, and N is the nth time interval. s This represents the total number of sampling points;

[0075] The amplitude equalization window E(n) is:

[0076]

[0077] Where A is the amplitude of the test set data signal x(n);

[0078] Preprocessed test set data x NSG (n) is:

[0079] x NSG (n)=T x (n)·I(n)·E(n)

[0080] Step 6.3: Apply the preprocessed fully connected neural network model to actual signal processing.

[0081] The beneficial effects of this invention are as follows:

[0082] (1) A fully connected neural network model was used to replace the iterative process of ALE. The powerful nonlinear fitting ability of the neural network can better handle complex acoustic environment changes.

[0083] (2) For different types of noise and signal characteristics, targeted optimization is carried out through a large amount of data training, rather than being limited to ALE, which requires reading the parameters in a new environment;

[0084] (3) It better realizes the iteration of weights under low signal-to-noise ratio. The processing gain of the traditional adaptive line spectrum enhancement method is only 0.31dB, while the processing gain of the deep learning-based adaptive line spectrum enhancement method is 9.38dB, which greatly improves the signal processing gain.

[0085] (4) By training with a large amount of data, the environmental features are iterated, unlike ALE which relies on manually designed algorithms and feature extraction methods;

[0086] (5) For weights with a specific low signal-to-noise ratio, the processing gain is significantly improved under the same input conditions. Attached Figure Description

[0087] Figure 1 This is a flowchart of a deep learning-based adaptive line spectrum enhancement method.

[0088] Figure 2 This is a diagram of the structure of a fully connected neural network model.

[0089] Figure 3 This is a flowchart of the adaptive line spectrum enhancement method based on deep learning.

[0090] Figure 4 Time-frequency diagram analysis for a simulated signal with SNR = 7dB;

[0091] Figure 4 In the figure, Figure (a) is the time-frequency diagram of the signal FFT (Fast Fourier Transform) method when SNR = 7dB, Figure (b) is the time-frequency diagram of the ALE (Adaptive Line Spectrum Enhancement) method, Figure (c) is the time-frequency diagram of the DLE (Deep Learning-based Line Spectrum Enhancement) method proposed in this invention, and Figure (d) is the time-frequency diagram of the NSG-DLE (Deep Learning-based Line Spectrum Enhancement via Noise Suppression Gate) method proposed in this invention.

[0092] Figure 5 Time-frequency diagram analysis for a simulated signal with SNR = 3dB;

[0093] Figure 5 In the figure, Figure (a) is the time-frequency diagram of the signal FFT (Fast Fourier Transform) method when SNR = 3dB, Figure (b) is the time-frequency diagram of the ALE (Adaptive Line Spectrum Enhancement) method, Figure (c) is the time-frequency diagram of the DLE (Deep Learning-based Line Spectrum Enhancement) method proposed in this invention, and Figure (d) is the time-frequency diagram of the NSG-DLE (Deep Learning-based Line Spectrum Enhancement via Noise Suppression Gate) method proposed in this invention.

[0094] Figure 6 The gain calculation results are for input signals with different signal-to-noise ratios.

[0095] Figure 6 In the diagram, (a) represents the output signal-to-noise ratio of the fully connected neural network model; (b) represents the gain of the fully connected neural network model.

[0096] Figure 7 The time-frequency plot shows the results of the experimental data processing.

[0097] Figure 7 In the figure, Figure (a) is the time-frequency diagram of the measured signal using the FFT (Fast Fourier Transform) method, Figure (b) is the time-frequency diagram of the ALE (Adaptive Line Spectrum Enhancement) method, Figure (c) is the time-frequency diagram of the DLE (Deep Learning-based Line Spectrum Enhancement) method proposed in this invention, and Figure (d) is the time-frequency diagram of the NSG-DLE (Deep Learning-based Line Spectrum Enhancement via Noise Suppression Gate) method proposed in this invention. Detailed Implementation

[0098] Mathematical software includes all software similar to MATLAB that can generate simulated signals.

[0099] To compensate for the inadequacy of simulation data, the field of underwater acoustic signal detection tends to use real-world datasets. This involves setting up experimental scenarios to simulate the emission of line spectrum signals and using acoustic sensors to collect data. The simulation environment is conducted in laboratory settings or real waters to ensure that the data covers a variety of underwater sound targets, including fish, whales, submarines, and ships. Different noise backgrounds include marine environmental noise and biological noise.

[0100] This invention discloses a deep learning-based adaptive line spectrum enhancement method, belonging to the field of signal processing. It replaces the iterative process of traditional adaptive line spectrum enhancement with a deep learning approach, achieving weight iteration under low signal-to-noise ratio (SNR) conditions and ultimately improving the SNR of the line spectrum signal. Simultaneously, preprocessing techniques are added before the fully connected neural network to enhance the line spectrum enhancement effect. Even under low SNR conditions, this invention can accurately identify and enhance target signals, improving the processing performance of traditional adaptive line spectrum enhancement (ALE) methods at low SNR.

[0101] like Figure 1 As shown, an adaptive line spectrum enhancement method based on deep learning is described, which specifically includes the following steps:

[0102] Step 1: Generate target acoustic signal sample data;

[0103] Step 2: Normalization processing of target acoustic signal sample data;

[0104] Step 3, construct the dataset; the dataset includes a training dataset and a test dataset;

[0105] Step 4: Construct a fully connected neural network model;

[0106] Step 5: Train the fully connected neural network model;

[0107] Step 6: Verify the effectiveness using a fully connected neural network model in actual signal processing.

[0108] In step 1, target acoustic signal sample data is generated using simulation or sonar methods. The target acoustic signals collected using simulation or sonar methods can ensure the diversity and authenticity of the target acoustic signals.

[0109] The simulation method uses MATLAB as the data tool, and the specific simulation settings for MATLAB are as follows:

[0110] First, the frequency, duration, and power of the target sound signal are preset; then, a white noise background is preset to simulate the real environment; next, the bandwidth and power of the noise signal are set.

[0111] The signal frequency is 300Hz, the sampling frequency is 2kHz, the broadband noise range is [100, 500], and the signal duration is 5s. The designed signal-to-noise ratios are 7dB and 3dB, respectively, and experimental sampling data is used.

[0112] The target acoustic signal is denoted as x(n), which includes the line spectrum signal S(n) and the noise signal N. noise (n);

[0113]

[0114] Where A is the amplitude of the line spectrum signal; f0 is the frequency of the line spectrum signal; T is the total duration of the line spectrum signal; f s The sampling frequency is denoted by n; n represents the nth sample. Let U be the random phase of the line spectrum signal, and let U be the random phase that satisfies the distribution U(0,2π).

[0115] In step 2, the normalization processing method for the target acoustic signal sample data is as follows:

[0116] To ensure a consistent convergence speed during training of the sample data, a minimum-maximum normalization method is used to normalize the target acoustic signal sample data. The normalized target acoustic signal... as follows:

[0117]

[0118] Where, min x and max x These represent the minimum and maximum values ​​of the target acoustic signal sequence, respectively.

[0119] In step 3, the dataset is constructed as follows:

[0120] The dataset will be the normalized target acoustic signal sample data, which includes a training set and a test set; the top T values ​​of the normalized target acoustic signal will be taken. s Time, i.e., T s *f s The signal of the number of sampling points is used as the training set data x T (n), the signal of the remaining sampling points is the test set data x. E (n).

[0121] like Figure 2 As shown, in step 4, the fully connected neural network model includes an input layer, a hidden layer, and an output layer; in this example, the number of neurons in each layer is 2000, 800, and 1, respectively.

[0122] A fully connected neural network model is used to replace the adaptive weight iteration process of ALE; target acoustic signal x T(n) serves as the input to the fully connected neural network model;

[0123] The input vector of the input layer:

[0124] x T (nk)=[x T (nk),x T (nk-1),…,x T (nk-(N-1)] T ,

[0125] Where, x T (nk) represents the latency of the training set data, where k represents the pre-latency points, and the pre-latency is set to 1 second, i.e., 1*f s The number of digital sampling points, [·] T is the transpose of the matrix, and N is the order of the input layer;

[0126] x training set data T (n) and the delay vector x of the training set data T (nk) serves as the input to the fully connected neural network model; the data at the corresponding sampling points in the training set are the output signals of the fully connected neural network model.

[0127] like Figure 3 As shown, in step 5, the training process of the fully connected neural network model includes:

[0128] Step 5.1: Perform initial training on the fully connected neural network model:

[0129] x training set data T If (n) is directly input into the fully connected neural network model, then the output signal y(n) of the fully connected neural network model is:

[0130] y(n)=f θ (x T (n))

[0131] Among them, f θ (·) represents the function mapping relationship of a fully connected neural network model;

[0132] Obtain the parameters of the trained fully connected neural network model

[0133]

[0134] Where N represents the order of the input layer; i is the sampling sequence number; x(i) is the input acoustic signal sampled at the i-th time; y(i) is the output acoustic signal sampled at the i-th time; L is the loss function; θ is the parameters of the fully connected neural network model;

[0135] The loss function L is:

[0136] L(x,y)=||xy|| 2

[0137] Where |||·|| represents the squared error operator; x and y are the sampled input acoustic signal and the sampled output acoustic signal, respectively;

[0138] Step 5.2: Use the Adam algorithm to process the fully connected neural network model parameters θ obtained in Step 5.1. * Optimize:

[0139]

[0140] In the Adam algorithm, the gradient descent method is used to solve for the learning rate α, which is set to 0.001. β′ and β″ are the first and second moments of the initial estimated gradient, respectively, with the first and second moments set to 0.9 and 0.999, respectively. The Adam algorithm can adaptively adjust the learning rate without manual adjustment, effectively solve the gradient noise problem, and converge quickly, achieving good results.

[0141] The Adam algorithm applies the parameters θ of a fully connected neural network model. * The optimization process is as follows:

[0142] During the optimization process, letter subscripts all represent the current iteration number;

[0143] First, given initial parameters: given learning rate α; given parameters β′ and β″; parameters β′ and β″ are used to estimate the first and second moments of the gradient; define the optimization objective function f(θ) and the initial values ​​of the optimization parameters θ0;

[0144] Secondly, set the initial values ​​t = 0, m0 = 0, v0 = 0, where t is the number of iterations, m is the first moment estimate of the gradient, and v is the second moment estimate;

[0145] Then, update the time from t+1 to t, and calculate the gradient g of the objective function. t Gradient of objective function g t Represented as: Among them, the first moment estimate m of the updated gradient t As follows: m t =β′m t-1 +(1-β′)g t Calculate the first-order gradient moment estimate after bias correction. Update the second moment estimate of the gradient and calculate the bias-corrected second moment estimate of the gradient.

[0146] v t =β″v t-1 +(1-β″)gt +g t

[0147]

[0148] Update parameters:

[0149]

[0150] Where ε = 10 -8 It is an infinitesimal quantity to ensure that the denominator is not zero;

[0151] Finally, repeat the above iterative process to obtain the optimal θ. *

[0152] Step 5.3: During the training of the fully connected neural network model, activation functions are added between the input layer, hidden layers, and output layer to ensure the nonlinear transformation of the fully connected neural network model.

[0153] The modified linear activation function ReLU(z) is used as the activation function for the fully connected neural network model.

[0154]

[0155] Where z is the calculation result of the fully connected neural network model during the forward propagation process;

[0156] When the input of the modified linear activation function is less than 0, the output of the fully connected neural network model is 0; the ReLU function obtains a sparse representation of the data, enabling the fully connected neural network model to accurately extract the relevant line spectrum frequencies in the acoustic signal; the ReLU function is more conducive to the fast and accurate descent of the gradient.

[0157] In step 6, the verification process of the fully connected neural network model in actual signal processing tasks is as follows:

[0158] Step 6.1: Test the trained fully connected neural network model:

[0159] The test set data is input into a fully connected neural network model for adaptive line spectrum enhancement. The method for adaptive line spectrum enhancement using a fully connected neural network model is called DLE (Deep Learning Line Enhancement). The enhancement result of DLE is the output result x of the fully connected neural network model. DLE (n) is:

[0160]

[0161] Step 6.2: Preprocess the tested fully connected neural network model to improve the signal-to-noise ratio:

[0162] The noise suppression gate (NSG) method is used as a preprocessing technique before inputting into a fully connected neural network model. The noise suppression gate method is as follows:

[0163] Test set data signal x E The autocorrelation function T of (n) x (n) is:

[0164] T x (n)=x E (-n)*x E (n)

[0165] The noise suppression gate I(n) is:

[0166]

[0167] Where, α T The zeroing time range is defined as n, where n is the nth sample, and N is the nth time interval. s This represents the total number of sampling points;

[0168] The amplitude equalization window E(n) is:

[0169]

[0170] Where A is the amplitude of the test set data signal x(n);

[0171] Preprocessed test set data x NSG (n) is:

[0172] x NSG (n)=T x (n)·I(n)·E(n)

[0173] Step 6.3: Apply the preprocessed fully connected neural network model to actual signal processing.

[0174] By inputting the original data and the preprocessed data (passed through a noise suppression gate) into the trained model, the following results can be obtained:

[0175] The processing gain of each method was calculated using FFT (Fast Fourier Transform), ALE (Adaptive Line Spectrum Enhancement), and the proposed DLE (Deep Learning Line Spectrum Enhancement) and NSG-DLE (Line Spectrum Enhancement via a noise suppression gate followed by deep learning) respectively. With an input signal-to-noise ratio of 7dB, the output results are as follows: Figure 4As shown in (a)-(c), the output signal-to-noise ratio and gain calculation results of the conventional adaptive line spectrum enhancement method and the deep learning-based adaptive line spectrum enhancement method are shown in Table 1, where the gain is defined as the output signal-to-noise ratio of different methods minus the output signal-to-noise ratio of the FFT result;

[0176] Table 1. Gain calculation results of various line spectrum enhancement methods when SNR = 7dB.

[0177]

[0178] Similarly, when the input signal-to-noise ratio is changed to 3dB, such as Figure 5 As shown in (a)-(c), the calculated output signal-to-noise ratio and gain are also shown in Table 2:

[0179] Table 2. Gain calculation results of various line spectrum enhancement methods when SNR = 3dB.

[0180]

[0181] Figure 6 The curves showing the output signal-to-noise ratio (SNR) and gain as a function of the input SNR are presented for various methods. Gain calculation is based on the FFT method. The input SNR increases from -15dB to 15dB in 2dB increments. The Monte Carlo simulation is performed 100 times. The output SNR and gain curves are shown below. Figure 6 As shown.

[0182] The proposed deep learning-based adaptive line spectrum enhancement method was validated using experimental data.

[0183] The transmitted signal frequency is 170Hz, the system sampling frequency is 8192Hz, the signal duration is 30s, the number of taps in the adaptive line spectrum enhancer is 1000, and the step size is 2×10⁻⁶. -6 The pre-delayed digital sampling points are 20, the number of neurons in the three layers are 1000, 300, and 1 respectively, the input vector dimension of the neural network model is 1000, the length of each frame of the time-frequency graph is 2 seconds, and the overlap rate between frames is 50%. Figure 7 As shown in (a)-(c), the table below gives the output signal-to-noise ratio and gain obtained by processing experimental data using various line spectrum enhancement methods.

[0184] Table 3. Gain calculation results of various line spectrum enhancement methods in experimental data processing.

[0185]

[0186]

[0187] The ALE processing gain is 0.31dB, which is poor at low signal-to-noise ratios. In contrast, the DLE method and NSG-DLE proposed in this invention have processing gains of 9.38dB and 20.85dB, respectively, which are higher. This verifies the correctness of the simulation analysis and demonstrates a good high-gain processing effect.

Claims

1. A deep learning-based adaptive line spectrum enhancement method, characterized in that, The adaptive line spectrum enhancement method specifically includes the following steps: Step 1: Generate target acoustic signal sample data; Step 2: Normalization processing of target acoustic signal sample data; Step 3, construct the dataset; the dataset includes a training dataset and a test dataset; Step 4: Construct a fully connected neural network model; Step 5: Train the fully connected neural network model; Step 6: Verify the effectiveness of the fully connected neural network model in actual signal processing. The effect verification process is as follows: Step 6.1: Test the trained fully connected neural network model: The test set data is input into a fully connected neural network model for adaptive line spectrum enhancement; the enhancement result of the adaptive line spectrum enhancement is the output of the fully connected neural network model. for: ； Step 6.2: Preprocess the tested fully connected neural network model to improve the signal-to-noise ratio: Noise suppression gates are used as a preprocessing technique before inputting into a fully connected neural network model. The noise suppression gate method is as follows: Test set data signals autocorrelation function for: ； The noise suppression gate is for: ； in, The time range for setting to zero is defined, where n is the nth sample. This represents the total number of sampling points; Amplitude Equalization Window for: ； Where A is the test set data signal. The amplitude; Preprocessed test set data for: ； Step 6.3: Apply the preprocessed fully connected neural network model to actual signal processing.

2. The adaptive line spectrum enhancement method based on deep learning according to claim 1, characterized in that, In step 1, target acoustic signal sample data is generated using simulation or sonar. The simulation tool used is MATLAB, and the specific simulation settings for MATLAB are as follows: First, the frequency, duration, and power of the target sound signal are preset; then, a white noise background is preset to simulate the real environment; next, the bandwidth and power of the noise signal are set. The target acoustic signal is denoted as Target acoustic signal Including line spectrum signals and noise signals ; ； Where A is the amplitude of the line spectrum signal; The frequency of the line spectrum signal is T; the total duration of the line spectrum signal is T. The sampling frequency is denoted by n; n represents the nth sample. Let the random phase of the line spectrum signal satisfy the distribution. .

3. The adaptive line spectrum enhancement method based on deep learning according to claim 1, characterized in that, In step 2, the normalization processing method for the target acoustic signal sample data is as follows: The target acoustic signal sample data is normalized using the min-max normalization method. The normalized target acoustic signal is then obtained. as follows: ； Where, min x and max x These represent the minimum and maximum values ​​of the target acoustic signal sequence, respectively.

4. The adaptive line spectrum enhancement method based on deep learning according to claim 1, characterized in that, In step 3, the dataset is constructed as follows: The dataset will be the normalized target acoustic signal sample data, which includes a training set and a test set; the first normalized target acoustic signal will be taken. Moment, that is * The signal of the number of sampling points is used as training set data. The signals from the remaining sampling points constitute the test set data. .

5. The adaptive line spectrum enhancement method based on deep learning according to claim 1, characterized in that, In step 4, the fully connected neural network model includes an input layer, a hidden layer, and an output layer; the target acoustic signal Used as input to a fully connected neural network model; The input layer is a delay vector of the training set data. : ， Where k represents the pre-delay point, and the pre-delay is set to 1 second, i.e., 1* The number of digital sampling points, is the transpose of the matrix, and N is the order of the input layer; training set data Delay vectors of training set data The data from the training set is used as input to the fully connected neural network model; the data from the corresponding sampling points are the output signals of the fully connected neural network model.

6. The adaptive line spectrum enhancement method based on deep learning according to claim 1, characterized in that, Step 5, the training process of the fully connected neural network model includes: Step 5.1: Perform initial training on the fully connected neural network model: training set data If the input is directly fed into a fully connected neural network model, the output signal of the fully connected neural network model will be... for: ； in, Represents the function mapping relationship of a fully connected neural network model; Obtain the parameters of the trained fully connected neural network model : ； Where N represents the order of the input layer; i is the sampling sequence number; The input acoustic signal is the i-th sample. Let L be the output acoustic signal of the i-th sample; L is the loss function. These are the parameters for a fully connected neural network model. The loss function L is: ； in, This represents the squared error operator; x and y are the sampled input acoustic signal and the sampled output acoustic signal, respectively; Step 5.2: Use the Adam algorithm to process the parameters of the fully connected neural network model obtained in Step 5.

1. Optimize: ； In the Adam algorithm, the learning rate is calculated using gradient descent. Take 0.001, and The first and second moments of the initial gradient estimation are 0.9 and 0.999, respectively. The Adam algorithm can adaptively adjust the learning rate without manual adjustment, effectively solve the gradient noise problem, and converge quickly, achieving good results. The Adam algorithm applies parameters to fully connected neural network models. The optimization process is as follows: Step 5.2.1, Given initial parameters: given learning rate Given parameters , ;parameter , Used to estimate the first and second moments of the gradient; to define the objective function. and the initial values ​​of the optimized parameters ; Step 5.2.2, Set initial values Where t is the number of iterations, m is the first moment estimate of the gradient, and v is the second moment estimate; Step 5.2.3, Update Time Calculate the gradient of the objective function Gradient of the objective function Represented as: Among them, the first moment estimate of the updated gradient as follows: Calculate the first-order gradient moment estimate after bias correction. Update the second moment estimate of the gradient, and calculate the bias-corrected second moment estimate of the gradient. : ； ； Update parameters: ； in, It is an infinitesimal quantity to ensure that the denominator is not zero; Step 5.2.4: Repeat steps 5.2.1 to 5.2.4 iteratively. When the Adam algorithm has completely traversed the database more than 40 times, the optimal result is finally obtained. ; Step 5.3: In the training of the fully connected neural network model, activation functions are added between the input layer, hidden layer, and output layer of the fully connected neural network model to ensure the non-linear transformation of the fully connected neural network model. The modified linear activation function is used as the activation function in the fully connected neural network model. for: ； in, This represents the computational results of the fully connected neural network model during the forward propagation process. When the input of the modified linear activation function is less than 0, the output of the fully connected neural network model is 0; the ReLU function obtains a sparse representation of the data, enabling the fully connected neural network model to accurately extract the relevant line spectrum frequencies in the acoustic signal; the ReLU function is more conducive to the fast and accurate descent of the gradient.

7. An electronic device, characterized in that, include: One or more processors; Memory; One or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs being configured to perform the method as described in any one of claims 1 to 6.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program code that can be invoked by a processor to execute the method as described in any one of claims 1 to 6.

Images