Impulse noise parameter estimation method based on hybrid neural network

Abstract

The invention relates to the technical field of underwater acoustic communication and signals, and discloses an impulse noise parameter estimation method based on a hybrid neural network, which comprises the following steps of: 1, constructing a nonlinear mapping model of non-Gaussian parameter estimation; 2, establishing a noise sample database; step 3, establishing a noise parameter estimator based on the hybrid neural network; 4, network training and online deployment: respectively carrying out network training through the noise training sample set, storing network parameters after network training, and carrying out online deployment application; step 5, receiving an actually measured noise sample, performing parameter off-line estimation to obtain a noise parameter, and feeding back the noise parameter to a communication receiver; sixthly, the communication receiver designs a signal detector through the received noise parameters, and the signal detection effect under the impulse noise is optimized; according to the scheme, small-sample, rapid and accurate parameter estimation is realized by using the hybrid neural network, and the method is suitable for effective deployment of a noise parameter estimation module of a communication system under impulse noise.

Description

Technical Field

[0001] The present invention relates to the field of underwater acoustic communication and signal technology, and in particular to a method for estimating pulse noise parameters based on a hybrid neural network. Background Art

[0002] Non-Gaussian noise is widely present in the marine environment, but Gaussian noise conditions are generally assumed in the design of marine communication modules. The actual noise statistics deviate significantly from the Gaussian distribution, which contradicts the traditional design concept of communication systems based on Gaussian noise. The key to communication systems under impulse noise is how to effectively eliminate the influence of impulse noise.

[0003] Nonlinear memoryless methods are typical impulse noise suppression methods, such as clipping, denoising, or a combination of clipping / denoising. Non-memory algorithms utilize the large amplitude characteristics of the impulse component, but cannot adaptively select the truncation threshold, and the residual noise component generated by it is prone to error accumulation. Sparsity-based impulse noise suppression utilizes the sparse structural characteristics of impulse noise to effectively restore the impulse component and completely remove the influence of impulse noise from the contaminated signal. Sparsity-based suppression methods are effective for impulse noise with a small number of impulse components. The above methods cannot restore the impulse component or even excessively remove the useful signal component for non-sparse or cluster-sparse impulse noise. Statistical parameter-based impulse suppression methods have less structural constraints on noise, and the theoretical performance boundary can be derived or approximated by introducing statistical parameters into the system design. However, the statistical parameters of impulse noise are usually assumed to be known, which is not realistic in practical applications. The current maximum likelihood (ML) optimal estimation has high complexity and requires a large number of noise samples. Considering low complexity and suboptimal estimation implementation, a general, low-cost, and highly generalizable estimation method is currently needed, which is suitable for non-stationary ocean noise environments. Summary of the invention

[0004] The purpose of the present invention is to solve the above problems. A pulse noise parameter estimation method based on a hybrid neural network is designed. A noise database is generated using a typical pulse noise model, a nonlinear mapping model of noise parameters is established, a hybrid neural network parameter estimator is designed, and the estimated parameters are fed back to the communication receiver to improve the parameter estimation accuracy and communication performance. This method is suitable for scenarios with different pulse noises, and better estimation effects can be obtained using a small amount of noise sample data.

[0005] The present invention provides a method for estimating impulse noise parameters based on a hybrid neural network. The method for estimating impulse noise parameters based on a hybrid neural network mainly comprises the following steps:

[0006] Step 1: Construct a nonlinear mapping model for non-Gaussian parameter estimation: Obtain noise samples X through noise models or measured noise, construct a nonlinear mapping function f(·), take the noise samples as input and the target parameters as output, and the parameter estimation problem is expressed as:

[0007]

[0008] In the formula, are the estimated parameters;

[0009] Step 2: Establish a noise sample database: determine the non-Gaussianity of the noise through probability distribution fitting, combine the typical ocean non-Gaussian noise model α steady-state noise model and Gaussian mixed noise model to generate samples X and labels Θ;

[0010] Step 3: Establish a noise parameter estimator based on a hybrid neural network: Establish the characteristic representation of the noise sample through the hybrid neural network. According to the temporal characteristics of the noise, the noise sample at time t is represented as x t , taking noise samples at multiple moments as input and target parameters as output, the parameter estimation problem is expressed as

[0011]

[0012] In the formula, is the estimated parameter, X={x t ,x t-1 ,…,x t-T+1}, F(·) represents the estimation function of the multi-layer network nesting, and T is the time window length of the historical data;

[0013] Step 4: Network training and online deployment: Perform network training using noise training sample sets, save network parameters after network training, and deploy applications online;

[0014] Step 5: Receive the measured noise sample to perform offline parameter estimation, obtain the noise parameter, and feed it back to the communication receiver;

[0015] Step 6: The communication receiver uses the received noise parameters to design a signal detector to optimize the signal detection effect under impulse noise.

[0016] Optionally, in the first implementation of the present invention, the noise model α in step 2 is a characteristic function of steady-state noise:

[0017]

[0018] Wherein, ω(t,α) and sign(t) are both piecewise functions. When α≠1, ω(t,α)=tan(πα / 2); when α=1, ω(t,α)=π / 2log|t|; when t>0, sign(t)=1; when t=0, sign(t)=0; when t<0, sign(t)=-1;

[0019] From the steady-state distribution of random variables X, denoted as:

[0020] X~S(α,β,γ,δ)

[0021] Wherein, α is the characteristic index, which describes the degree of heavy tail of the distribution. When α=2, the α steady-state distribution degenerates into Gaussian distribution. When α=1, the α steady-state distribution degenerates into Cauchy distribution. The smaller α is, the stronger the impulse is and the higher the degree of heavy tail is. The larger α is, the stronger the Gaussian is and the weaker the degree of heavy tail is. β∈[-1,1] is the slope parameter, which determines the degree of asymmetry of the distribution. δ∈(-∞,+∞) is the location parameter, which indicates the location of the center of the distribution. When 0<α<1, δ corresponds to the median of the distribution. When 1<α≤2, δ corresponds to the mean of the distribution. γ>0 is the scale parameter, which determines the degree of diffusion of the distribution around the location parameter δ.

[0022] When β=0 and δ=0, the α steady-state distribution is symmetric about , and the α steady-state distribution is called the symmetric α steady-state distribution.

[0023] Optionally, in a second implementation of the present invention, the noise model α in step 2 is a probability density function of Gaussian mixture noise:

[0024]

[0025] Where M is the number of Gaussian components, f m (x) is the probability density function of the mth Gaussian component, 0≤p m ≤1 is the weighting coefficient, The total variance of the Gaussian mixture noise e(n) is

[0026] Optionally, in a third implementation of the present invention, the noise model α in step 2 is a probability density function of MCA noise:

[0027]

[0028] Where A is the pulse index. The larger A is, the closer the probability density is to the Gaussian distribution. The smaller A is, the stronger the pulse is. is the power ratio of background noise to impulse noise,

[0029] Optionally, in a fourth implementation of the present invention, the timing characteristics of the noise signal in step three include at least the characteristics of large amplitude, the probability of pulses, the interval of pulse components, and the power ratio of pulses and Gaussian components.

[0030] Optionally, in a fifth implementation of the present invention, in step 3, the convolutional neural layer and the LSTM module are used to fully explore the temporal characteristics of the impulse noise. The characteristics of the pulse signal are refined and extracted by the convolutional layer, and the high-level features are input into the stacked LSTM layer for learning the temporal features and outputting the estimation results, specifically including:

[0031] Assume that the noise sample length is L, divide the sample into T sample segments, each with a length of To round down, the convolution kernel size is 1, and the output of the convolution layer is:

[0032]

[0033] Where W c and B c Represent the weight matrix and bias of the convolutional layer, X in is the input of the convolutional layer, X output is the output feature, represents the convolution operation, and Activation(·) represents the nonlinear activation function;

[0034] Select two convolution layers to convolve with the input data. The convolution kernel sizes are 1×c1 and 1×c2 respectively. The 'same' padding mode is used. The data passes through the two convolution layers, and the internal timing information is integrated and merged. The output of the convolution layer is:

[0035] X c =F c (X)

[0036] In the formula, F c (·) A mapping function representing input and output features;

[0037] Noise timing characteristics X c ={x c,t-T+1 ,…,x c,t} Input multi-layer LSTM for the second stage of feature learning and parameter estimation, LSTM designs the forget gate f g,t , input gate i g,t and output gate o g,t , forget gate f g,t Determines which information from the previous time step should be discarded, input gate i g,t Determines the current input information, expressed as:

[0038] f g,t =σ(Wf,h h t-1 +W f,x x t +W f,c c t-1 +b f )

[0039] i g,t =σ(W i,h h t-1 +W i,x x t +W i,c c t-1 +b i )

[0040] Where W h , W x and W c Indicates the previous output state h t-1 , input x t and the previous state c t-1 , b represents the bias vector, weights and biases are distinguished by subscripts f, i, etc., and σ represents the sigmoid activation function;

[0041] After the data passes through LSTM, the output cell state is updated to:

[0042] c t =f g,t ⊙c t-1 +i g,t ⊙c g,t

[0043] In the formula, c g,t =tanh(W c,h h t-1 +W c,x x t +b c ) represents the candidate state vector generated by the input through the tanh function, ⊙ represents the Hadamard product;

[0044] The LSTM output process is expressed as:

[0045] o g,t =σ(W o,h h t-1 +W o,x x t +W o,c c t-1 +b o ),

[0046] h t =o g,t ⊙tanh(c t )

[0047] In stacked LSTM, the output of the LSTM module is the input of the next LSTM. Each LSTM block considers T time steps of data. By stacking LSTM layers, the noise feature X c The high-level temporal features in , take the last hidden state of the stacked LSTM as the output, and return the aggregate information. The output of the stacked LSTM is input into a fully connected layer, which converts the output into a parameter set Θ.

[0048] Optionally, in a sixth implementation of the present invention, in step 4, during the training process, the mean square error of parameter estimation is used as a loss function, which is expressed as:

[0049]

[0050] Where φ is the network parameter, and Θ represent the estimated parameters and true parameters respectively. Z is the number of data batches. The network parameters are updated using the mini-batch gradient descent method. The Adam optimizer is used to update the network parameters. After network training, the network parameters are determined. When a new noise sample is input, the model outputs the estimated noise parameters.

[0051] Optionally, in a seventh implementation of the present invention, the bit error rate of the receiver in step 6 is

[0052]

[0053] Where ρ is the signal-to-noise ratio, Γ(·) is the gamma function.

[0054] The beneficial effects of the present invention are as follows: the method utilizes a hybrid deep neural network to mine the temporal information of pulse noise and establishes a unified estimation framework. In order to solve the problem of unexplored noise features, the convolution layer is used as a shallow layer to learn the noise feature space, and then the output of the convolution layer is sent to a stacked LSTM module to deeply explore the temporal information of the pulse noise. The model realizes the feature extraction and mapping of small sample noise statistical parameters, breaks through the limitation of large sample number in traditional machine learning estimation methods, realizes low-complexity and real-time parameter estimation, and introduces the parameters of pulse noise into the design of communication receivers, thereby improving the performance of communication receivers under pulse noise. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the following detailed description of the preferred embodiment.The drawings are only for the purpose of illustrating the preferred embodiments and are not to be construed as limiting the invention.

[0056] Figure 1 A schematic diagram of a hybrid neural network structure provided by an embodiment of the present invention;

[0057] Figure 2 A schematic diagram of the MSE effect of noise parameter estimation for noise samples with different time window lengths provided by an embodiment of the present invention;

[0058] Figure 3 A schematic diagram of MSE results of parameters of different noise models provided in an embodiment of the present invention;

[0059] Figure 4 A schematic diagram of comparing bit error rates of communication receivers under different estimation methods provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0060] The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, device, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0061] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 , the method specifically comprises the following steps:

[0062] Step 1: Construct a nonlinear mapping model for non-Gaussian parameter estimation: Obtain noise samples X through noise models or measured noise, construct a nonlinear mapping function f(·), take the noise samples as input and the target parameters as output, and the parameter estimation problem is expressed as:

[0063]

[0064] In the formula, are the estimated parameters;

[0065] Step 2: Establish a noise sample database: determine the non-Gaussianity of the noise through probability distribution fitting, combine the typical ocean non-Gaussian noise model α steady-state noise model and Gaussian mixed noise model to generate samples X and labels Θ;

[0066] In this step, a typical ocean impulse noise model is used to construct a noise database. Three typical non-Gaussian ocean noise models are selected, including α noise, GM noise and MCA noise. The parameters of α steady-state noise are Θ = {α, β, γ, δ}, and the parameters of GM noise are The noise parameters of Middleton Class A noise are Θ = {A, Γ, σ};

[0067] The characteristic function of α steady-state noise is:

[0068]

[0069] Wherein, ω(t,α) and sign(t) are both piecewise functions. When α≠1, ω(t,α)=tan(πα / 2), when α=1, ω(t,α)=π / 2log|t|ω(t,α)=π / 2log|t|, when t>0, sign(t)=1; when t=0, sign(t)=0; when t<0, sign(t)=-1. The random variable X that obeys the steady-state distribution is denoted as X~S(α,β,γ,δ), and the relevant parameters are:

[0070] (1) α is the characteristic index, which describes the degree of heavy tail of the distribution. When α = 2, the α steady-state distribution degenerates into a Gaussian distribution. When α = 1, the α steady-state distribution degenerates into a Cauchy distribution. The smaller α is, the stronger the impulse is and the higher the degree of heavy tail is. The larger α is, the stronger the Gaussian is and the weaker the degree of heavy tail is.

[0071] (2) β∈[-1,1] is the slope parameter, which determines the degree of asymmetry of the distribution;

[0072] (3)δ∈(-∞,+∞) is the location parameter, which indicates the location of the center of the distribution. When 0<α<1, δ corresponds to the median of the distribution; when 1<α≤2, δ corresponds to the mean of the distribution;

[0073] (4) γ>0 is the scale parameter, which determines the degree of spread of the distribution around the location parameter δ, similar to the variance of the Gaussian distribution;

[0074] When β=0 and δ=0, the α-stable distribution is symmetric about δ. At this time, the α-stable distribution is called the symmetric α-stable (SαS) distribution.

[0075] The probability density function of Gaussian mixture (GM) noise is:

[0076]

[0077] Where M is the number of Gaussian components, f m (x) is the probability density function of the mth Gaussian component, 0≤pm ≤1 is the weighting coefficient,

[0078] The probability density function of the mth Gaussian component is expressed as:

[0079]

[0080] Where μ m and They represent the mean and variance of the mth Gaussian component respectively. When M = 2, it is a commonly used binary Gaussian mixture distribution;

[0081] In a bivariate Gaussian mixture distribution, the Gaussian component can be considered as having a variance of The Gaussian mixture model can describe the scene where Gaussian thermal noise and impulse noise exist at the same time. The total variance of Gaussian mixture noise e(n) is

[0082] The probability density function of Middleton Class A (MCA) noise is:

[0083]

[0084] Where A is the pulse index. The larger A is, the closer the probability density is to the Gaussian distribution; the smaller A is, the stronger the pulse is. is the power ratio of background noise to impulse noise,

[0085] Step 3: Establish a noise parameter estimator based on a hybrid neural network: Establish the characteristic representation of the noise sample through the hybrid neural network. According to the temporal characteristics of the noise, the noise sample at time t is represented as x t , taking noise samples at multiple moments as input and target parameters as output, the parameter estimation problem is expressed as

[0086]

[0087] In the formula, is the estimated parameter, X={x t ,x t-1 ,…,x t-T+1}, F(·) represents the estimation function of the multi-layer network nesting, and T is the time window length of the historical data;

[0088] In this step, the core of the model is to use convolutional neural layers and LSTM modules to fully explore the temporal characteristics of impulse noise. As a representative time series, the characteristics of the impulse signal are refined and extracted through the convolutional layer, and the high-level features are input into the stacked LSTM layer to learn the temporal features and output the estimation results.

[0089] 1) Assuming the noise sample length is L, divide the sample into T sample segments, each with a length of To round down, the convolution layer has an excellent ability to extract features through the convolution kernel. The convolution kernel increases the number of channels and reduces the input width. In order to avoid information loss in the time dimension, the convolution kernel size is selected as 1 and the pooling layer is abandoned because the pooling layer extracts important information of adjacent elements and reduces the data dimension. The output of the convolution layer is:

[0090]

[0091] Where W c and B c Represent the weight matrix and bias of the convolutional layer, X in is the input of the convolutional layer, X output is the output feature, represents the convolution operation, and Activation(·) represents the nonlinear activation function.

[0092] Properly increasing the network depth can improve performance, but simply deepening the network will cause gradient degradation or gradient explosion. In order to avoid overfitting, two layers of convolutional layers are selected to convolve with the input data. The convolution kernel sizes are 1×c1 and 1×c2 respectively. In order to keep enough information, the 'same' padding mode is adopted. The data passes through the two layers of convolutional layers, and the internal time series information is integrated and merged. When the time series features are not destroyed, the feature space, the output of the convolutional layer is X c =F c (X), F c (·) represents the mapping function between input and output features, and the output of CNN Input LSTM for the next stage of processing;

[0093] 2) The timing information hidden in the impulse noise helps improve the noise parameter estimation effect. The noise timing characteristics X c ={x c,t-T+1 ,…,x c,t} Input multi-layer LSTM for the second stage of feature learning and parameter estimation. LSTM belongs to recurrent neural network (RNN), which establishes a connection path from the previous input to the current output. Compared with traditional RNN, LSTM is good at learning long-term sequence data, long-term memory and time information hidden in the data, solving problems such as long-term dependence, gradient disappearance and gradient explosion in data flow and training process. LSTM designs forget gate f g,t , input gate i g,t and output gate o g,t , forget gate f g,t Determines which information from the previous time step should be discarded, input gate i g,tDetermines the current input information, expressed as:

[0094] f g,t =σ(W f,h h t-1 +W f,x x t +W f,c c t-1 +b f )

[0095] i g,t =σ(W i,h h t-1 +W i,x x t +W i,c c t-1 +b i )

[0096] Where W h , W x and W c Indicates the previous output state h t-1 , input x t and the previous state c t-1 , b represents the bias vector, weights and biases are distinguished by subscripts f, i, etc., and σ represents the sigmoid activation function;

[0097] LSTM keeps the cell state c partially flowing on the network. After the data passes through LSTM, the output cell state is updated to:

[0098] c t =f g,t ⊙c t-1 +i g,t ⊙c g,t

[0099] In the formula, c g,t =tanh(W c,h h t-1 +W c,x x t +b c ) represents the candidate state vector generated by the input through the tanh function. ⊙ represents the Hadamard product. The LSTM output process is expressed as:

[0100] o g,t =σ(W o,h h t-1 +W o,x x t +W o,c c t-1 +b o ),

[0101] ht =o g,t ⊙tanh(c t )

[0102] Stacked LSTM is an improved version of LSTM. It consists of multiple LSTM modules, each of which contains multiple LSTM units. In stacked LSTM, the output of the LSTM module is the input of the next LSTM. Each LSTM block considers T time steps of data. The stacking structure deepens the network and improves the estimation performance. By stacking LSTM layers, the noise feature X is used to c The high-level temporal features in the stacked LSTM are used as output, and the aggregated information is returned;

[0103] The output of the stacked LSTM is fed into a fully connected layer that transforms the output into a parameter set Θ. There are n d neurons, n d Corresponding to the number of parameters, the features pass through the fully connected layer and the network outputs the parameters Θ.

[0104] Step 4: Network training and online deployment: Perform network training using noise training sample sets, save network parameters after network training, and deploy applications online;

[0105] In this step, the network parameters are learned through model training. The data samples are divided into training data sets and test data sets. Considering the generation of data sets and labels and the diversity of data, the data sets are uniformly generated within the allowable range of each noise model, rather than selecting a limited number of parameters and then inputting them into the model to generate samples. For the test data set, the parameters are regenerated to maintain independence from the training data set and improve versatility;

[0106] During the training process, the mean square error (MSE) of parameter estimation is used as the loss function, which is expressed as:

[0107]

[0108] Where φ is the network parameter, and Θ represent estimated parameters and true parameters respectively, Z is the number of data batches, and the network parameters are updated using the mini-batch gradient descent method. The Adam optimizer is used to update the network parameters. After network training, the network parameters are determined. In the online testing phase, the parameters are regenerated to maintain independence from the training data set and improve versatility. When a new noise sample is input, the model outputs the estimated noise parameters.

[0109] Step 5: Receive the measured noise sample to perform offline parameter estimation, obtain the noise parameter, and feed it back to the communication receiver;

[0110] In this step, the noise parameters are output and fed back, and a noise parameter feedback link is established. The noise parameter estimation module can estimate the parameters in real time, thereby realizing fast processing of the receiver.

[0111] Step 6: The communication receiver uses the received noise parameters to design a signal detector to optimize the signal detection effect under impulse noise.

[0112] In this step, the communication receiver is optimized. Taking α noise as an example, under known noise parameters, the bit error rate of the receiver is:

[0113]

[0114] Where ρ is the signal-to-noise ratio, Γ(·) is the gamma function.

[0115] The performance advantages of the method of the present invention are verified by experimental analysis of simulated noise and measured ocean noise. The network structure and parameters used in the experiment are shown in Table 1:

[0116] Table 1 Network parameters

[0117]

[0118] The comparison algorithms include DNN, CNN, LSTM and other DL estimation algorithms. The structures and parameters of the comparison algorithms are adjusted according to the requirements of noise parameter estimation. The DNN model has 6 layers of networks, each containing {256, 512, 1024, 512, 256, 4} neurons. The CNN contains 2 convolutional layers with output channels of {32, 64} and 2 fully connected layers. The LSTM network consists of 2 layers of LSTM modules stacked, each containing 32 LSTM units. The batch size of the data sample set is 64, and the maximum number of epochs is 300. The time window length T = 200, and the learning rate is set to 0.001. All networks are implemented by Tensorflow2.0, and the running computer is configured with NVDIA GeForce GTX 1060GPU. After the network is trained, the network parameters are fixed and configured online, and the results are analyzed by Matlab platform.

[0119] Consider α steady-state noise, GM noise and MCA noise as simulation noise sample generation models. Each noise model generates 2000 training samples and 200 test samples. The parameters during sample generation fluctuate randomly to improve the generalization of network training. The range of α steady-state noise parameters is set to 1<α≤2, -1≤β≤1, 0≤δ≤2 and 0≤γ≤3. The binary GM noise parameters are set to 0≤μ1≤0.1, 0≤μ2≤0.1, k∈[10,15] and c∈[0.1,0.2], where c is the probability of occurrence of the pulse component. For MCA noise, L is set to 3, A∈[0.01,1], σ 2 ∈[1,10],Γ∈[10 -3 ,1].

[0120] Figure 2 The MSE effects of estimation of α noise and GM noise under different networks and different time window lengths are compared. As shown in the figure, the method of the present invention has good estimation performance, but DNN and CNN cannot capture statistical parameters in all cases, even when the time window length T=300. The LSTM estimation network introduces the time series information inside the data, and the estimation shows higher accuracy as the time window length increases. For α steady-state noise, the average MSE of the comparison algorithm is [1.627, 1.1576, 0.7356, 0.4578], indicating that the model of the present invention not only performs better under different sample lengths, but also has a better average MSE. For GM noise, the average MSEs are [0.2460, 0.9095, 0.1854, 0.1685] respectively. For MCA noise, the average MSEs are [1.8393, 1.4242, 0.79, 0.596] respectively.

[0121] Figure 3 The MSE performance of different impulse noise model parameters is studied. For α steady-state noise, the estimation performance based on LSTM model is better than that of DNN and CNN models. The estimation performance of the proposed algorithm is better than that of CNN and DNN, with a value of 3.93×10 -3 The MSE performance of the proposed algorithm is slightly worse than that of LSTM. For parameters β, γ and δ, the proposed algorithm performs better than other algorithms. For GM noise, for parameters c, μ1 and μ2, the estimation errors are all within [10 -5 ,10 -4 ] of the order of magnitude. 2 The MSE of the algorithm of the present invention is 0.03179, 3.52×10 -3 , 5.71×10 -3Outperforms DNN, CNN and LSTM. For parameter k, the proposed model has the smallest MSE error, and the MSE of CNN exceeds 1, which is caused by the large fluctuation range of k value. For MCA noise, for parameter Γ, the proposed method is estimated to be slightly worse than LSTM, with a performance gap of about 2.8×10 -5 For parameter A, CNN has the best performance, but the CNN estimation performance is unstable because the present invention has a large error rate for parameter σ. 2 The estimation performance is an order of magnitude higher.

[0122] Figure 4 The receiver optimization effect after noise parameter feedback is studied. Considering the underwater multi-receiver system, the transmitted signal is modulated by BPSK, the receiving end adopts MRC reception, and the amplitude of the underwater acoustic channel is Rayleigh fading. It can be seen from the figure that the parameters obtained by the present invention can assist the system to improve the bit error rate effect. When SNR = 20dB, the BER performance gap is about 3.04×10 -4 When BER=0.01 is examined, the present invention achieves performance gains of 1dB, 7dB, and 9dB respectively compared with LSTM, CNN, and DNN.

[0123] It can be seen from the above experimental results that the present invention is superior to the comparative algorithm and has better robustness and effectiveness for different pulse noises. Among them, the hybrid neural network can better extract the implicit characteristics of the noise and improve the estimation performance. In summary, the present invention provides an ocean pulse noise parameter estimation method and a receiver design scheme, which belongs to the field of underwater acoustic signal processing. The advantage of the present invention is that it fully exploits the characterization characteristics and time correlation of noise data, improves the accuracy of parameter estimation, and at the same time utilizes network optimization design to adapt the noise parameter estimation target, and feeds back the estimated parameters to the receiver to improve the decoding performance of the system.

[0124] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. The above embodiments and descriptions are only preferred examples of the present invention and are not intended to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A method for estimating impulse noise parameters based on a hybrid neural network, characterized in that: The impulse noise parameter estimation method based on hybrid neural network mainly includes the following steps: Step 1: Construct a nonlinear mapping model for non-Gaussian parameter estimation: Obtain noise samples X through noise models or measured noise, construct a nonlinear mapping function f(·), take the noise samples as input and the target parameters as output, and the parameter estimation problem is expressed as: In the formula, are the estimated parameters; Step 2: Establish a noise sample database: determine the non-Gaussianity of the noise through probability distribution fitting, combine the typical ocean non-Gaussian noise model α steady-state noise model and Gaussian mixed noise model to generate samples X and labels Θ; Step 3: Establish a noise parameter estimator based on a hybrid neural network: Establish the characteristic representation of the noise sample through the hybrid neural network. According to the temporal characteristics of the noise, the noise sample at time t is represented as x t , taking noise samples at multiple moments as input and target parameters as output, the parameter estimation problem is expressed as In the formula, is the estimated parameter, X={x t ,x t-1 ,…,x t-T+1 }, F(·) represents the estimation function of the multi-layer network nesting, and T is the time window length of the historical data; Step 4: Network training and online deployment: Perform network training using noise training sample sets, save network parameters after network training, and deploy applications online; Step 5: Receive the measured noise sample to perform offline parameter estimation, obtain the noise parameter, and feed it back to the communication receiver; Step 6: The communication receiver uses the received noise parameters to design a signal detector to optimize the signal detection effect under impulse noise.

2. The method for estimating impulse noise parameters based on a hybrid neural network as claimed in claim 1, characterized in that: The noise model α in step 2 is the characteristic function of steady-state noise: Wherein, ω(t,α) and sign(t) are both piecewise functions. When α≠1, ω(t,α)=tan(πα / 2); when α=1, ω(t,α)=π / 2log|t|; when t>0, sign(t)=1; when t=0, sign(t)=0; when t<0, sign(t)=-1; From the steady-state distribution of random variables X, denoted as: X~S(α,β,γ,δ) Wherein, α is the characteristic index, describing the degree of heavy tail of the distribution. When α=2, the α steady-state distribution degenerates into Gaussian distribution. When α=1, the α steady-state distribution degenerates into Cauchy distribution. The smaller α is, the stronger the impulse is and the higher the degree of heavy tail is. The larger α is, the stronger the Gaussian is and the weaker the degree of heavy tail is. β∈[-1,1] is the slope parameter, which determines the degree of asymmetry of the distribution. δ∈(-∞,+∞) is the location parameter, which indicates the location of the center of the distribution. When 0<α<1, δ corresponds to the median of the distribution. When 1<α≤2, δ corresponds to the mean of the distribution. γ>0 is the scale parameter, which determines the degree of diffusion of the distribution around the location parameter δ. When β=0 and δ=0, the α steady-state distribution is symmetric about , and the α steady-state distribution is called the symmetric α steady-state distribution.

3. The method for estimating impulse noise parameters based on a hybrid neural network according to claim 1, characterized in that: The noise model α in step 2 is the probability density function of Gaussian mixture noise: Where M is the number of Gaussian components, f m (x) is the probability density function of the mth Gaussian component, 0≤p m ≤1 is the weighting coefficient, The total variance of the Gaussian mixture noise e(n) is 4. The method for estimating impulse noise parameters based on a hybrid neural network as claimed in claim 1, characterized in that: The noise model α in step 2 is the probability density function of MCA noise: Where A is the pulse index. The larger A is, the closer the probability density is to the Gaussian distribution. The smaller A is, the stronger the pulse is. is the power ratio of background noise to impulse noise, 5. The method for estimating impulse noise parameters based on a hybrid neural network according to claim 1, characterized in that: The time series characteristics of the noise signal in step three at least include the characteristics of large amplitude, the probability of pulses, the interval of pulse components, and the power ratio of pulses and Gaussian components.

6. The method for estimating impulse noise parameters based on a hybrid neural network according to claim 1, characterized in that: In step 3, the convolutional neural layer and LSTM module are used to fully explore the temporal characteristics of impulse noise. The characteristics of the impulse signal are refined and extracted through the convolutional layer, and the high-level features are input into the stacked LSTM layer for temporal feature learning and output estimation results, including: Assume that the noise sample length is L, divide the sample into T sample segments, each with a length of To round down, the convolution kernel size is 1, and the output of the convolution layer is: Where W c and B c Represent the weight matrix and bias of the convolutional layer, X in is the input of the convolutional layer, X output is the output feature, represents the convolution operation, and Activation(·) represents the nonlinear activation function; Select two convolution layers to convolve with the input data. The convolution kernel sizes are 1×c1 and 1×c2 respectively. The 'same' padding mode is used. The data passes through the two convolution layers, and the internal timing information is integrated and merged. The output of the convolution layer is: X c =F c (X) In the formula, F c (·) A mapping function representing input and output features; Noise timing characteristics X c ={x c,t-T+1 ,…,x c,t } Input multi-layer LSTM for the second stage of feature learning and parameter estimation, LSTM designs the forget gate f g,t , input gate i g,t and output gate o g,t , forget gate f g,t Determines which information from the previous time step should be discarded, input gate i g,t Determines the current input information, expressed as: f g,t =σ(W f,h h t-1 +W f,x x t +W f,c c t-1 +b f ) i g,t =σ(W i,h h t-1 +W i,x x t +W i,c c t-1 +b i ) Where W h , W x and W c Indicates the previous output state h t-1 , input x t and the previous state c t-1 , b represents the bias vector, weights and biases are distinguished by subscripts f, i, etc., and σ represents the sigmoid activation function; After the data passes through LSTM, the output cell state is updated to: c t =f g,t ⊙c t-1 +i g,t ⊙c g,t In the formula, c g,t =tanh(W c,h h t-1 +W c,x x t +b c ) represents the candidate state vector generated by the input through the tanh function, ⊙ represents the Hadamard product; The LSTM output process is expressed as: o g,t =σ(W o,h h t-1 +W o,x x t +W o,c c t-1 +b o ), h t =o g,t ⊙tanh(c t ) In stacked LSTM, the output of the LSTM module is the input of the next LSTM. Each LSTM block considers T time steps of data. By stacking LSTM layers, the noise feature X c The high-level temporal features in , take the last hidden state of the stacked LSTM as the output, and return the aggregate information. The output of the stacked LSTM is input into a fully connected layer, which converts the output into a parameter set Θ.

7. The method for estimating impulse noise parameters based on a hybrid neural network as claimed in claim 1, characterized in that: In step 4, during the training process, the mean square error of parameter estimation is used as the loss function, which is expressed as: Where φ is the network parameter, and Θ represent estimated parameters and true parameters respectively, Z is the number of data batches, and the network parameters are updated using the mini-batch gradient descent method. The Adam optimizer is used to update the network parameters. After network training, the network parameters are determined. When a new noise sample is input, the model outputs the estimated noise parameters.

8. The method for estimating impulse noise parameters based on a hybrid neural network as claimed in claim 1, characterized in that: The bit error rate of the receiver in step 6 is: Where ρ is the signal-to-noise ratio, Γ(·) is the gamma function.

Images