A deep learning-based channel equalization method, system, device, and medium

Abstract

A channel equalization method, system, device and medium based on deep learning, the method is, a data set is constructed and divided into a training set, a validation set and a test set for training, validating and testing a deep learning equalizer; a deep learning equalization network is constructed to realize channel equalization; the deep learning equalization network is trained, validated and tested, the performance of the final deep learning equalization network is evaluated, and the channel equalization based on deep learning is completed; the system, device and medium based on the channel equalization method based on deep learning realize channel equalization; the application can significantly improve the recovery precision of signals and the communication quality of the system, and is particularly suitable for modern communication systems in high dynamic and complex environments.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of channel equalization, and particularly relates to a channel equalization method, system, device and medium based on deep learning. BACKGROUND

[0002] With the rapid development of modern communication technology and the continuous improvement of information warfare demand, channel equalization technology has become one of the key technologies to improve the performance of communication systems. Channel equalization is mainly used to correct the distortion caused by factors such as multipath effect, attenuation, interference and other factors encountered by signals in the transmission process, the purpose is to restore the original state of the signal and ensure the accuracy and reliability of data transmission. In modern communication systems, signals often experience complex propagation environments before reaching the receiver, resulting in significant differences between the received signal and the transmitted signal, which requires effective channel equalization technology to adjust the received signal to make it as close to the original signal as possible.

[0003] Among numerous channel equalization techniques, the least mean square (LMS) algorithm and the recursive least square (RLS) algorithm are two widely used adaptive filtering techniques. The working principle of the LMS algorithm is based on the gradient descent idea, which pursues the minimization of the sum of squared output errors by continuously adjusting the filter weights. The operation of the LMS algorithm (document [Widrow B, Hoff ME. Jr. Adaptive switching circuits. IRE WESCON Conv. Rec., Pt. 4, 1960. 96-104.]) is iterative, adjusting the weights step by step with each received signal to adapt to changes in channel characteristics. This gradual adjustment process makes the LMS algorithm very effective in steady-state, but its performance can be limited in cases of rapid changes in channel characteristics, mainly because it relies on gradual iterations to gradually approach the optimal solution. In contrast to the LMS algorithm, the RLS algorithm provides a more complex but superior performance solution. The RLS algorithm (document [McWhirter J G. Recursive least-squares minimization using systolic arrays. Proc. SPIE, Realtime Signal Processing VI, 1983, 431.]) optimizes the weight adjustment process by adapting more quickly to changes in signals and channels. It uses all past data to update the weights at each time, which allows the RLS algorithm to converge more quickly to the optimal weight configuration, especially in environments where channel conditions change rapidly. This comprehensive consideration of historical data allows the RLS algorithm to quickly respond to any changes in the channel, resulting in more accurate channel estimation and faster adjustment. Although the RLS algorithm has a higher computational complexity and requires more processing power and resources, its advantages in handling high dynamic range signals and complex changing channels are unmatched.

[0004] Existing channel equalization techniques, such as the least mean square (LMS) algorithm and the recursive least square (RLS) algorithm, are widely used in communication systems, but they face many challenges in dynamic channel environments and high-speed data transmission conditions. Specifically, the LMS algorithm performs poorly in high-speed or variable channel conditions due to its slow convergence speed and low stability; while the RLS algorithm, although fast in convergence, has high computational complexity and requires a large amount of computing resources, limiting its application on resource-constrained devices. In addition, these traditional equalization methods usually require manual adjustment and rely on accurate prior channel information, making it difficult to adapt to unknown or rapidly changing channel characteristics. SUMMARY

[0005] In order to overcome the above-mentioned deficiencies of the prior art, the purpose of the present application is to provide a channel equalization method, system, device and medium based on deep learning, which introduces a deep learning-based channel equalization framework, uses the strong learning ability of deep neural networks to automatically extract channel features from a large amount of data without manual adjustment and complex parameter configuration; by using a deep learning model, it can quickly adapt to changes in signals in complex channel environments and effectively process signal recovery under non-linear and non-stationary channel conditions; through practical application, it can significantly improve the recovery accuracy of signals and the communication quality of the system, and is particularly suitable for modern communication systems in high dynamic and complex environments.

[0006] In order to achieve the above-mentioned purpose, the present application adopts the following technical solutions:

[0007] A channel equalization method based on deep learning, comprising the following steps:

[0008] Step 1, data set construction; first generate a bit stream, generate a transmission signal after modulation, shape filter the generated transmission signal and modulate it into a form that meets the communication requirements, perform analog transmission, and finally, the output signal after channel transmission is constructed into a data set, and is divided into a training set, a validation set and a test set;

[0009] Step 2, build a deep learning equalization network that meets the data set format of step 1, based on a convolutional autoencoder (CAE), integrate a residual network structure (ResNet), a self-attention mechanism (Self-Attention), a layer normalization (LayerNorm) and a Swish activation function to realize channel equalization;

[0010] Step 3, train, validate and test the deep learning equalization network; first, input the training set into the deep learning equalization network to learn the equalization features of the signal, and use the validation set to monitor the training process to prevent overfitting; during the training process, use mean square error (MSE) as the loss function and use Adam optimizer to update the parameters to improve the training efficiency and speed up the convergence speed; in the validation phase, use the mean absolute error (MAE) to evaluate the performance of the deep learning equalization network on the validation set to help adjust the training strategy; after training, use the test set to evaluate the final deep learning equalization network to test its generalization ability and equalization performance on the test set, and complete the channel equalization based on deep learning.

[0011] The specific method of step 1 is:

[0012] Step 1.1. Select the bit rate of the transmitted signal and the modulation mode of the transmitted signal; according to the selected modulation mode, first randomly generate the corresponding original data sequence, and modulate the generated original data sequence into the corresponding symbol to generate the transmission signal; save these original data sequences as labels for use in subsequent training process;

[0013] Step 1.2. Perform shaping filtering on the transmission signal generated in step 1.1, modulate the shaped filtered signal into a form that meets the communication requirements, and then select a communication channel to simulate the transmission of the signal that meets the communication requirements; finally, the output signal after channel transmission constitutes a data set, which is divided into a training set, a validation set and a test set for training, validation and testing of the deep learning equalizer.

[0014] Step 1.1. Select the bit rate of the transmitted signal and the modulation mode of the transmitted signal, specifically: Let the binary original data sequence be a bit stream of length K:

[0015] b=[b1,b2,...,b K ],b i ∈{0,1}

[0016] Select the modulation mode, map every log2(M) bits to a complex constellation symbol:

[0017]

[0018] Where M is the modulation order, N = K / log2(M) is the total number of symbols; the complex symbol can be decomposed into in-phase (I) and quadrature (Q) components:

[0019]

[0020] Step 1.2. Perform shaping filtering on the transmission signal generated in step 1.1, modulate the shaped filtered signal into a form that meets the communication requirements, and then select a communication channel to simulate the transmission of the signal that meets the communication requirements, specifically:

[0021] Let the expected output symbol length of the deep learning equalization network be L, then reshape the transmission signal S IQ generated in step 1.1 into Output ∈ R B×L×2 , where is the batch size:

[0022]

[0023] Then interpolate each symbol by sps times, generate a baseband signal using a shaping filter and discretize it to get a symbol sequence, convolve the symbol sequence with the filter, and the output shaped signal is:

[0024]

[0025] Next, the shaped signal is passed through a wireless channel and added with noise:

[0026]

[0027] After separating the real and imaginary components of the received signal passed through a wireless channel and added with noise, the received signal that meets the communication requirements is:

[0028]

[0029] The received signal y that meets the communication requirements is: IQ Reshape to Input ∈R according to sps x L B×(sps·L)×2 :

[0030]

[0031] The final data set is obtained as an input-output pair:

[0032]

[0033] The specific method of step 2 is:

[0034] Step 2.1 constructs a balanced network of deep learning, including residual network structure (ResNet), self-attention mechanism (Self-Attention), layer normalization (LayerNorm), and Swish activation function.

[0035] Step 2.1.1 in feature extraction, residual network structure (ResNet) is introduced in both encoder and decoder, residual network structure (ResNet) is composed of consecutive Conv1D layers, each Conv1D layer is followed by layer normalization and activation function to improve training stability; at the end of residual network structure (ResNet), skip connection (SkipConnection) is used to directly add input to output to alleviate the problem of gradient disappearance and ensure that information can be effectively transmitted in the balanced network of deep learning.

[0036] Wherein, the forward propagation formula of residual network structure (ResNet) is:

[0037] h l+1 = F(h l ,W l ) + h l

[0038] F(·) is a residual function, including convolution, layer normalization and activation function; when backpropagation, gradient calculation is decomposed into two paths:

[0039]

[0040] Even if the gradient of the residual function F(·) approaches zero, the skip connection can still retain at least the same order of magnitude of the input gradient through the unit matrix I

[0041] Step 2.1.2 Residual Network Structure (ResNet) reconstructs the complete mapping H(y) = x from the received signal y to the transmitted signal x into learning residual function F(y) = H(y) - y; the deep learning equalization network output should satisfy Where Δ is the channel distortion compensation amount, then the deep learning equalization network only needs to fit the channel distortion compensation amount Δ:

[0042]

[0043] By reconstructing the complete mapping into the residual mapping, the deep learning equalization network is converted from learning the absolute mapping to learning the relative correction amount; the channel distortion can be decomposed into the linear part H linear (y) and the nonlinear interference ò(y), the residual network structure (ResNet) hierarchically approximates the linear part H linear (y) and the nonlinear interference ò(y):

[0044] H linear (y) = W2·σ(W1y + b1) + b2

[0045]

[0046] Where σ is the activation function, and φ(·) is the high-order nonlinear function;

[0047] Step 2.2 introduces a self-attention mechanism (Self-Attention) after the residual network structure in step 2.1 to enhance the global feature modeling capability of the deep learning equalization network;

[0048] Multi-head self-attention (Multi-Head Attention, MHA) structure is adopted to enhance the information of low-dimensional features;

[0049] Assuming that the output signal of the residual network structure is Y = [y1, y2,..., y N ] ∈ R N×d , the multi-head self-attention (Multi-Head Attention, MHA) structure first generates a query matrix Q, a key matrix K, and a value matrix V: ​

[0050] Q = YW Q K = YW K V = YW V

[0051] where W Q W K W V ∈R d×d is a learnable parameter; the attention weight matrix A is calculated by scaling dot product:

[0052]

[0053] A ij represents the attention degree of position i to position j; the final output is a weighted value matrix:

[0054] Z = AV∈R N×d

[0055] Step 2.3 introduces the layer normalization (LN) technology in the balanced network of deep learning, and all feature dimensions of a single sample are independently normalized by layer normalization (LN);

[0056] Step 2.3.1 sets the feature tensor of the output of the l-th layer as h∈R C×H×W , where C is the number of channels, HxW is the spatial dimension, and the normalization process of layer normalization (LN) is:

[0057]

[0058] where γ c and β c are learnable scaling and offset parameters, and the feature distribution of each channel is independently normalized without waiting for batch data accumulation;

[0059] Step 2.3.2, through layer normalization (LN), the feature distribution of each channel is independent of each other, which can suppress the abnormal propagation of gradient, and the gradient calculation of the l-th layer is calculated by back propagation:

[0060]

[0061] The numerical range of the gradient term is constrained, avoiding the exponential decay or growth of the gradient amplitude of the balanced network of deep learning;

[0062] Step 2.4 selects Swish as the activation function to improve the stability and balance ability of training, and the mathematical expression of Swish is as follows:

[0063]

[0064] Allow negative values to pass when x < 0;

[0065] The first derivative of Swish is calculated as follows:

[0066]

[0067] Where x is the function input, representing the variable input to Swish.

[0068] The specific method of step 3 is:

[0069] Step 3.1: Input the training set into the deep learning equalization network. The deep learning equalization network learns the signal equalization features in the training set data and gradually optimizes the model parameters. During the training process, the mean square error (MSE) is used as the loss function, defined as:

[0070]

[0071] Where n1 is the number of training set samples, is the true value of the i-th sample in the training set, is the predicted value of the i-th sample in the training set. The loss function is used to measure the difference between the predicted value and the true value. By minimizing the loss function, the deep learning equalization network can better restore the original characteristics of the transmission signal generated in step 1.1.

[0072] Use the Adam optimizer to update the parameters, improve the training efficiency and speed up the convergence speed. The update rule of the Adam optimizer is as follows:

[0073]

[0074] Where θ t is the current deep learning equalization network parameter, θ t+1 is the deep learning equalization network parameter to be updated at the next moment, m t and v t are the first and second moment estimates of the gradient, η is the learning rate, and e is a constant to prevent division by zero errors.

[0075] Step 3.2: In the validation phase, use the validation set to monitor the training effect and prevent overfitting. For the validation set, use the mean absolute error (MAE) as the evaluation indicator:

[0076]

[0077] Use the early stopping strategy. When the validation set is in the continuous k rounds, the current moment L MAE(t) is less than a set threshold μ, i.e. MAE (t-1) are less than a set threshold μ, i.e.

[0078]

[0079] At this time, it is considered that the deep learning equalization network has converged. In order to avoid overfitting, the verification set L MAE The deep learning equalization network parameter with the minimum value is selected as the final deep learning equalization network parameter.

[0080] Step 3.3 inputs the received signal in the test set after being transmitted through the wireless channel and adding noise in step 1.2 into the deep learning equalization network, and the deep learning equalization network outputs the corresponding predicted signal. According to the mean square error (MSE) calculation formula in step 3.1 and the mean error (MAE) calculation formula in step 3.2, the mean square error (MSE) and the mean error (MAE) are calculated by difference calculation. If the error between the calculated mean square error (MSE) and the optimal mean square error (MSE) in the training stage in step 3.1 is less than a set value μ1, and the error between the calculated mean error (MAE) and the optimal mean error (MAE) in the verification stage in step 3.2 is less than a set value μ2, then the model has good generalization performance.

[0081] A channel equalization system based on deep learning, comprising:

[0082] A data set construction module for step 1, which constructs a complete data stream by bit stream generation, modulation, shaping filtering, analog transmission and channel transmission, realizes data set construction, and divides the data set into a training set, a verification set and a test set.

[0083] An equalization network construction module for step 2, which realizes channel equalization by constructing a deep learning equalization network based on a convolutional autoencoder (CAE), integrating a residual network structure (ResNet), a self-attention mechanism (Self-Attention), a layer normalization (LayerNorm) and a Swish activation function.

[0084] An equalization network training module for step 3, which realizes network training by inputting the training set into the deep learning equalization network, learning equalization features, using mean square error (MSE) as a loss function, and using an Adam optimizer to update parameters, improve training efficiency and speed up convergence.

[0085] The balanced network verification module is used for step 3, and the training effect of the deep learning balanced network is monitored through a verification set to prevent overfitting, the performance of the deep learning balanced network on the verification set is evaluated by using mean absolute error (MAE), and dynamic adjustment and optimization of the training strategy are realized.

[0086] The balanced network test module is used for step 3, and the deep learning balanced network is evaluated through a test set, the error of the deep learning balanced network on the test set is calculated, including mean square error (MSE) and mean absolute error (MAE), the generalization ability and balanced performance of the deep learning balanced network are tested, and comprehensive testing and evaluation of the performance of the deep learning balanced network are realized.

[0087] A deep learning-based channel equalization device comprises:

[0088] A memory is used for storing a computer program.

[0089] A processor is used for realizing the deep learning-based channel equalization method in steps 1 to 3 when the computer program is executed.

[0090] A computer readable storage medium stores a computer program, and the computer program can realize deep learning-based channel equalization by the method in steps 1 to 3 when the computer program is executed by a processor.

[0091] Compared with the prior art, the present application has the following advantages:

[0092] 1. The present application proposes a deep learning channel equalization network training data set construction method for digital modulation signals, which is especially suitable for communication scenarios with shaping filtering and multipath interference. The method solves the key problems of dimension mismatch, insufficient complex processing and single channel modeling in traditional equalization data set construction by systematically designing the input-output dimension mapping relationship and channel distortion simulation process.

[0093] 2. The present application proposes a deep learning channel equalization network architecture that integrates residual network, self-attention mechanism, layer normalization and Swish activation function, which is specially designed for high-precision equalization of digital modulation signals in shaping filtering and multipath interference scenarios. The architecture solves the bottleneck problems of gradient disappearance, limited local perception and insufficient channel time-varying adaptability of traditional equalization networks through multi-level feature reuse, global dependence modeling, dynamic distribution adaptation and non-linear enhancement mechanism.

[0094] In summary, in a complex channel environment, the deep learning equalization method of the present application has significant advantages compared to the traditional RLS (recursive least squares) algorithm, especially in dealing with factors such as multipath effect, Doppler effect and signal noise, and performs better. Based on MSE (mean square error), it can be clearly seen that the deep learning equalization method has advantages in complex channel environment. In the above simulation experiment, based on QPSK signal, the MSE of the present application is reduced by about 3dB compared with the MSE of RLS, and based on 8PSK signal, the MSE of the present application is reduced by about 2.6dB compared with the MSE of RLS.

[0095] This shows that the deep learning equalization method of the present application has stronger adaptability and robustness, and can learn the complex mapping relationship of the signal through the deep neural network model. This method no longer relies on traditional linear models, but automatically extracts features from data and performs nonlinear mapping, thereby adapting to nonlinear distortion and time-varying changes in the channel. In a complex channel environment, the deep learning equalization method can effectively capture key features in the signal, especially in terms of multipath and fading, and has more advantages than the RLS algorithm. Therefore, in a complex environment, the deep learning equalization method can significantly reduce the MSE value and provide more accurate signal recovery than the RLS algorithm. BRIEF DESCRIPTION OF DRAWINGS

[0096] Figure 1 is a deep learning-based equalization system diagram of the present application.

[0097] Figure 2 is a deep learning-based equalization network architecture diagram of the present application.

[0098] Figure 3 is a residual network structure diagram of the present application.

[0099] Figure 4 is a multi-head self-attention structure diagram of the present application.

[0100] Figure 5 is a QPSK equalization performance diagram under different algorithms of the present application.

[0101] Figure 6 is an 8PSK equalization performance diagram under different algorithms of the present application.

[0102] Figure 7 is a QPSK equalization constellation diagram under LMS algorithm of the present application.

[0103] Figure 8 is an 8PSK equalization constellation diagram under LMS algorithm of the present application.

[0104] Figure 9 is a QPSK equalization constellation diagram under RLS algorithm of the present application.

[0105] Figure 10 is the 8PSK constellation diagram under the RLS algorithm.

[0106] Figure 11 is the QPSK constellation diagram under the deep learning equalization network.

[0107] Figure 12 is the 8PSK constellation diagram under the deep learning equalization network. DETAILED DESCRIPTION

[0108] The application will be further described in detail below in combination with the drawings and simulation experiments.

[0109] As shown in the figure, a channel equalization method based on deep learning comprises the following steps: Figure 1

[0110] Step 1, data set construction; first generate a bit stream, generate a transmission signal after modulation, shape filter the generated signal and modulate it into a form conforming to the communication requirements, perform analog transmission, finally, the output signal after channel transmission constitutes a data set, and is divided into a training set, a validation set and a test set; the specific method is:

[0111] Step 1.1, select the bit rate of the transmitted signal and the modulation mode of the signal; according to the selected modulation mode, first randomly generate the corresponding original data sequence, and modulate the generated original data sequence into the corresponding symbol to generate the transmission signal; save these original data sequences as labels for use in the subsequent training process;

[0112] Step 1.2, shape filter the transmission signal generated in step 1.1, modulate the shaped filtered signal into a form conforming to the communication requirements, then select a communication channel to perform analog transmission on the signal conforming to the communication requirements; finally, the output signal after channel transmission constitutes a data set, and the data set is divided into a training set, a validation set and a test set for training, validation and testing of the deep learning equalizer.

[0113] The bit rate of the transmitted signal and the modulation mode of the signal selected in step 1.1 are as follows: let the binary original data sequence be a bit stream of length K:

[0114] b = [b1, b2,..., b K ], b i ∈ {0, 1}

[0115] Select a modulation mode (such as QPSK), map every log2(M) bits to a complex constellation symbol:

[0116] ​

[0117] where M is the modulation order (M = 4 for QPSK, 1 symbol is mapped for every 2 bits), N = K / log2(M) is the total number of symbols; the complex symbols can be decomposed into in-phase (I) and quadrature (Q) components:

[0118]

[0119] Step 1.2: The transmission signal generated in step 1.1 is shaped filtered, and the shaped filtered signal is modulated into a form that meets the communication requirements, and then a communication channel is selected to transmit the signal that meets the communication requirements analogically, specifically:

[0120] Let the length of the equalization network of deep learning be L, and then reshape the transmission signal S IQ generated in step 1.1 into Output ∈ R B×L×2 , where is the batch size:

[0121]

[0122] Then, each symbol is multiplied by sps times, a shaping filter (such as a raised cosine filter) is used to generate a baseband signal and discretize it to obtain a symbol sequence, and the symbol sequence is convolved with the filter. The output shaped signal is:

[0123]

[0124] Next, the shaped signal is transmitted through a wireless channel (such as a Doppler shift, multipath fading channel model) and noise is added:

[0125]

[0126] After separating the real and imaginary parts of the received signal that has been transmitted through a wireless channel (such as a Doppler shift, multipath fading channel model) and added noise, the received signal that meets the communication requirements is:

[0127]

[0128] The received signal y IQ that meets the communication requirements is reshaped according to sps x L into Input ∈ R B×(sps·L)×2 :

[0129]

[0130] The final data set is obtained as an input-output pair:

[0131]

[0132] Step 2: Construct a deep learning equalization network conforming to the dataset format of Step 1. Based on a convolutional autoencoder (CAE), integrate a residual network structure (ResNet), a self-attention mechanism, layer normalization (LayerNorm), and the Swish activation function to achieve channel equalization. The equalization network architecture is as follows: Figure 2 As shown; the specific method is as follows:

[0133] Step 2.1 Construct a balanced deep learning network, including a residual network structure (ResNet), a self-attention mechanism, layer normalization (LayerNorm), and the Swish activation function;

[0134] Step 2.1.1 In terms of feature extraction, a residual network structure (ResNet) is introduced into both the encoder and decoder. The residual network structure is as follows: Figure 3 As shown, the ResNet residual network structure consists of consecutive Conv1D layers. Each Conv1D layer is followed by a normalization and activation function to improve training stability. At the end of the ResNet residual network structure, a SkipConnection is used to directly add the input to the output to alleviate the gradient vanishing problem and ensure that information can be effectively transferred in the balanced network of deep learning. This allows the deep network to learn the structural information of the signal more effectively and improve the signal reconstruction quality.

[0135] The forward propagation formula for the ResNet structure is as follows:

[0136] h l+1 =F(h) l W l )+h l

[0137] F(·) is the residual function, which includes convolution, layer normalization, and activation functions; during backpropagation, gradient calculation is decomposed into two paths:

[0138]

[0139] Even the gradient of the residual function F(·) As the distance approaches zero, skip connections can still retain gradients of at least the same order of magnitude as the input through the identity matrix I. This characteristic enables the network to be trained stably even when it has more than 100 layers, making it particularly suitable for equalization tasks that require modeling complex channel responses.

[0140] Step 2.1.2 Residual network structure (ResNet) reconstructs the complete mapping H(y) = x from the received signal y to the transmitted signal x into a learned residual function F(y) = H(y) - y; the deep learning equalization network output should satisfy where Δ is the channel distortion compensation amount, then the deep learning equalization network only needs to fit the channel distortion compensation amount Δ:

[0141]

[0142] By reconstructing the complete mapping into a residual mapping, the deep learning equalization network is converted from learning an absolute mapping to learning a relative correction amount; the channel distortion can be decomposed into a linear part H linear (y) and a nonlinear interference ò(y), the residual network structure (ResNet) hierarchically approximates the linear part H linear (y) and the nonlinear interference ò(y):

[0143] H linear (y) = W2·σ(W1y + b1) + b2

[0144]

[0145] where σ is an activation function and φ(·) is a high-order nonlinear function;

[0146] Step 2.2 Introduce a self-attention mechanism (Self-Attention) after the residual network structure in step 2.1 to enhance the global feature modeling capability of the deep learning equalization network; traditional CNNs only rely on local convolution operations, and the receptive field is limited, making it difficult to capture global features, while the self-attention mechanism can effectively learn the long-range dependencies between data;

[0147] A multi-head self-attention (Multi-HeadAttention, MHA) structure is used to enhance the information of low-dimensional features, as shown in Figure 4

[0148] Assuming that the output signal of the residual network structure is Y = [y1, y2,..., y N ] ∈ R N×d , the multi-head self-attention (Multi-HeadAttention, MHA) structure first generates a query matrix Q, a key matrix K, and a value matrix V:

[0149] Q = YW Q , K = YW K , V = YW V

[0150] where W Q , W K , W V ​∈ R d×d is the learnable parameter; the attention weight matrix A is calculated by scaling dot product:

[0151]

[0152] A ij represents the attention degree of position i to position j; the final output is a weighted value matrix:

[0153] Z = AV∈ R N×d

[0154] Step 2.3 introduces the layer normalization (LN) technology in the balanced network of deep learning, and all feature dimensions of a single sample are independently normalized by layer normalization (LN); the core advantage of layer normalization (LN) technology is to standardize the feature distribution within a single sample, overcoming the dependence on batch statistics of traditional batch normalization (BN), so as to better adapt to the dynamic time-varying characteristics and sequence processing requirements in the channel equalization task;

[0155] Step 2.3.1 sets the feature tensor of the output of the l-th layer as h∈ R C×H×W , where C is the number of channels, HxW is the spatial dimension, and the normalization process of layer normalization (LN) is:

[0156]

[0157] , where γ c and β c are learnable scaling and offset parameters, and the feature distribution of each channel is independently normalized without waiting for batch data accumulation;

[0158] Step 2.3.2, through layer normalization (LN), the feature distribution of each channel is independent of each other, which can suppress the abnormal propagation of gradient, and the gradient calculation of the l-th layer is calculated by back propagation:

[0159]

[0160] The numerical range of the gradient term is constrained, avoiding the exponential decay or growth of the gradient amplitude of the deep learning balanced network;

[0161] Step 2.4 selects Swish as the activation function to improve the stability and balancing ability of training, and the mathematical expression of Swish is as follows:

[0162]

[0163] In x Allow negative values to pass through when <0;

[0164] The first derivative of Swish is calculated as follows:

[0165]

[0166] Where x is the function input, representing the variable input to Swish.

[0167] It can be seen that even in the case of x < 0, the gradient of Swish will not become completely 0, but will maintain a small gradient flow, making Swish able to alleviate the vanishing gradient problem and improve the training efficiency of deep learning balanced networks; Finally, the deep learning balanced network is constructed.

[0168] Step 3, train, validate and test the deep learning balanced network; First, input the training set into the balanced network, learn the balanced features of the signal, and use the validation set to monitor the training process to prevent overfitting; During the training process, mean square error (MSE) is used as the loss function, and Adam optimizer is used for parameter update to improve the training efficiency and speed up the convergence speed; In the validation phase, the mean absolute error (MAE) is used to evaluate the performance of the model on the validation set, which helps to adjust the training strategy. After training, the test set is used to evaluate the final model, to test its generalization ability and balanced performance on the test set;

[0169] The specific method is:

[0170] Step 3.1 input the training set into the deep learning balanced network, the deep learning balanced network learns the signal balanced features in the training set data, and gradually optimizes the model parameters; During the training process, mean square error (MSE) is used as the loss function, defined as:

[0171]

[0172] Where n1 is the number of training set samples, is the true value of the i-th sample in the training set, is the predicted value of the i-th sample in the training set, the loss function is used to measure the difference between the predicted value and the true value, by minimizing the loss function, the deep learning balanced network can better restore the original characteristics of the transmission signal generated in step 1.1.

[0173] Adam optimizer is used for parameter update to improve training efficiency and speed up convergence, the update rule of Adam optimizer is as follows:

[0174]

[0175] wherein, θ t is the deep learning equalization network parameter at the current time, θ t+1 is the deep learning equalization network parameter to be updated at the next time, m t and v t are the first and second moment estimates of the gradient respectively, η is the learning rate, and e is a constant to prevent division by zero error.

[0176] Step 3.2 In the verification phase, the training effect is monitored using the verification set to prevent overfitting. For the verification set, the mean absolute error (MAE) is used as the evaluation index:

[0177]

[0178] The early stopping strategy is used. When the verification set L MAE (t) is smaller than the previous time L MAE (t-1) for continuous k rounds, and both are smaller than the set threshold μ, that is,

[0179]

[0180] At this time, it is considered that the deep learning equalization network has converged. To avoid overfitting, the deep learning equalization network parameter with the smallest verification set L MAE value is selected as the final deep learning equalization network parameter.

[0181] Step 3.3 The received signal in the test set after passing through the wireless channel (such as the Doppler shift, multipath fading channel model) and adding noise in step 1.2 is input into the deep learning equalization network. The deep learning equalization network outputs the corresponding predicted signal. For the transmission signal in the test set without channel transmission and the predicted signal of the deep learning equalization network, the mean square error (MSE) and the mean error (MAE) are calculated according to the calculation formula in 3.1 and the calculation formula in 3.2, respectively. If the error between the calculated mean square error (MSE) and the optimal mean square error (MSE) in the training phase of step 3.1 is less than the set value μ1, and the error between the calculated mean error (MAE) and the optimal mean error (MAE) in the verification phase of step 3.2 is less than the set value μ2, then the model has good generalization performance.

[0182] Simulation experiment:

[0183] Under complex channel conditions containing multipath and Doppler effect, the performance of the deep learning equalization network proposed in the application is compared with that of the traditional adaptive equalization algorithm minimum mean square (LMS) and recursive least squares (RLS) method to evaluate the optimization ability of the application under complex channel conditions.

[0184] The deep learning equalization network is constructed on a Tensorflow 2.4.0 framework, the GPU is NVIDIA 3090, the python version is 3.8, the CUDA version is 11.0.3, and the CUDNN version is 8.0.5.39.

[0185] Table 1 simulation parameter setting

[0186] Parameter Value Modulation QPSK, 8PSK Sampling frequency F s ]]> 8MHz Symbol rate F d ]]> 1MHz Sampling time T 10ms Gaussian white noise signal-to-noise ratio SNR -5dB: 1: 10dB Number of multipath channel paths n 3 Path delays 0, 0.4, 0.9μs Path gains 0, -5, -10dB Doppler shift 0.5Hz

[0187] As shown in Table 1, the modulation mode is selected as QPSK and 8PSK, the sampling frequency is set to 8MHz, the symbol rate is set to 1MHz, the sampling time is 10ms, the signal-to-noise ratio of Gaussian white noise changes in the range of -5dB to 10dB with an increment of 1dB, the number of paths of the multipath channel is 3, the time delay of each path is 0, 0.4 and 0.9us respectively, the gain of each path is 0, -5 and -10dB respectively, and the Doppler frequency offset is set to 0.5Hz.

[0188] The experimental data are respectively input into the LMS algorithm, the RLS algorithm and the method of the present application, and the simulation results obtained are as shown in the following table Figures 5-12

[0189] Figure 5 Figure 6 ​​QPSK and 8PSK equalization performance figures under different algorithms, the horizontal axis is SNR, and the vertical axis is MSE. With the increase of signal-to-noise ratio, the MSE of the three methods shows a downward trend, which indicates that the quality of the signal increases with the increase of signal-to-noise ratio. Specifically, in the low signal-to-noise ratio environment (-5dB to 0dB), the method of the application effectively suppresses noise interference by virtue of the residual network feature extraction capability, and for QPSK signal, the MSE value decreases from 1.5e-1 to 6e-2, which is about 36.4% and about 32.3% higher than the performance of LMS and RLS respectively, and for 8PSK signal, the MSE value decreases from 1.45e-1 to 6e-2, which is about 37.9% and about 33.6% higher than the performance of LMS and RLS respectively; in the medium signal-to-noise ratio environment (1dB to 5dB), the method of the application can converge to the optimal solution faster and reduce the steady-state error, thanks to the dynamic modeling of the multi-head attention to the channel, and for QPSK signal, the MSE value decreases from 5e-2 to 2e-2, which is about 63.2% and about 61.3% higher than the performance of LMS and RLS respectively, and for 8PSK signal, the MSE value decreases from 5e-2 to 2.3e-2, which is about 64.5% and about 61.9% higher than the performance of LMS and RLS respectively; in the high signal-to-noise ratio environment (6dB to 10dB), the multi-hop connection structure of the residual network in the method of the application shows unique advantages, effectively eliminating the residual error in the high signal-to-noise ratio region, and further improving the accuracy and stability of the algorithm, and for QPSK signal, the MSE value is further compressed from 1.5e-2 to 8e-3, which is about 80.8% and about 79.8% higher than the performance of LMS and RLS respectively, and for 8PSK signal, the MSE value is further compressed from 1.8e-2 to 1.1e-3, which is about 81.3% and about 79.5% higher than the performance of LMS and RLS respectively.

[0190] Figure 7 、 Figure 8 is the constellation diagram comparison effect before and after QPSK and 8PSK equalization under LMS algorithm when SNR is 5dB, Figure 9 、 Figure 10 is the constellation diagram comparison effect before and after QPSK and 8PSK equalization under RLS algorithm when SNR is 5dB, Figure 11 、 Figure 12To compare the constellation diagram effects of QPSK and 8PSK before and after equalization under the condition that the SNR is 5dB. Specifically, although LMS and RLS algorithms make the QPSK signal and 8PSK signal after equalization gather around the ideal position, the residual phase offset and the compensation limitation of high-order modulation non-linear distortion are still exposed. The method of the application improves the signal recovery quality in a complex noise environment through the residual architecture and the self-attention mechanism, and reduces the phase offset and amplitude jitter standard deviation of the QPSK signal and 8PSK signal to a lower level on the basis of maintaining a lower MSE.

Claims

1. A channel equalization method based on deep learning, characterized in that, Specifically, the following steps are included: Step 1, Dataset Construction: First, a bit stream is generated, which is then modulated to generate a transmission signal. The generated transmission signal is then shaped, filtered, and modulated into a form that meets communication requirements for simulated transmission. Finally, the output signal after transmission through the channel constitutes the dataset and is divided into a training set, a validation set, and a test set. Step 2: Construct a deep learning equalization network conforming to the dataset format of Step 1. Based on a convolutional autoencoder (CAE), integrate a residual network structure (ResNet), a self-attention mechanism, layer normalization (LayerNorm), and the Swish activation function to achieve channel equalization; specifically: Step 2.1 Construct a balanced deep learning network, including a residual network structure (ResNet), a self-attention mechanism, layer normalization (LayerNorm), and the Swish activation function; Step 2.1.1 In terms of feature extraction, a residual network structure (ResNet) is introduced in both the encoder and decoder. The residual network structure (ResNet) consists of consecutive Conv1D layers. After each Conv1D layer, a normalization layer and activation function are added to improve training stability. At the end of the ResNet structure, a Skip Connection is used to directly add the input to the output to alleviate the gradient vanishing problem and ensure that information can be effectively transferred in the balanced network of deep learning. The forward propagation formula for the ResNet structure is as follows: h l+1 =F(h l ,W l )+h l F(·) is the residual function, which includes convolution, layer normalization, and activation functions; during backpropagation, gradient calculation is decomposed into two paths: Even the gradient of the residual function F(·) As the distance approaches zero, skip connections can still retain gradients of at least the same order of magnitude as the input through the identity matrix I. Step 2.1.2 The Residual Network (ResNet) structure reconstructs the complete mapping H(y) = x from the received signal y to the transmitted signal x into a learned residual function F(y) = H(y) - y; the output of the deep learning equalization network should satisfy... Where Δ is the channel distortion compensation amount, the deep learning equalization network only needs to fit the channel distortion compensation amount Δ: By reconstructing the complete mapping into a residual mapping, the deep learning equalization network transforms its learning from absolute mapping to learning relative corrections; channel distortion can be decomposed into a linear part H. linear (y) and nonlinear disturbances ∈ (y), the residual network structure (ResNet) hierarchically approximates the linear part H. linear (y) and nonlinear disturbance ∈ (y): H linear (y)=W2·σ(W1y+b1)+b2 Where σ is the activation function and φ(·) is a higher-order nonlinear function; Step 2.2 Introduce a self-attention mechanism after the residual network structure in Step 2.1 to enhance the global feature modeling capability of the deep learning balanced network; A multi-head attention (MHA) structure is used to enhance the information of low-dimensional features; Assume the output signal of the residual network structure is The Multi-Head Attention (MHA) structure first generates the query matrix Q, the key matrix K, and the value matrix V: Q=YES Q ,K=YW K ,V=YW V in, These are learnable parameters; the attention weight matrix A is calculated using a scaled dot product: A ij This represents the attention given to position i by position j; the final output is a weighted matrix: Step 2.3 Introduce Layer Normalization (LN) technology into the balanced network of deep learning, and perform independent normalization on all feature dimensions of a single sample through Layer Normalization (LN); Step 2.3.1 Let the feature tensor output by the l-th layer be... Where C is the number of channels, H×W is the spatial dimension, and the normalization process of Layer Normalization (LN) is as follows: Where, γ c and β c For learnable scaling and offset parameters, the feature distribution of each channel is independently normalized without waiting for batch data accumulation; Step 2.3.2: Through Layer Normalization (LN), the feature distribution of each channel is independent, which can suppress gradient anomaly propagation. Backpropagation is then performed to calculate the gradient of the l-th layer. gradient term The numerical range is constrained, thus avoiding the exponential decay or growth of gradient magnitude in deep learning equilibration networks. Step 2.4 Selects Swish as the activation function to improve training stability and balance. The mathematical expression for Swish is as follows: Negative values ​​are allowed when x < 0; The first derivative of Swish is calculated as follows: Where x is the function input, representing the variable input to Swish; Step 3 involves training, validating, and testing the deep learning equalization network. First, the training set is input into the deep learning equalization network to learn the equalization features of the signal. The validation set is used to monitor the training process and prevent overfitting. During training, mean squared error (MSE) is used as the loss function, and the Adam optimizer is used to update parameters to improve training efficiency and accelerate convergence. In the validation phase, mean absolute error (MAE) is used to evaluate the performance of the deep learning equalization network on the validation set, helping to adjust the training strategy. After training, the final deep learning equalization network is evaluated using the test set to verify its generalization ability and equalization performance on the test set, thus completing the deep learning-based channel equalization.

2. The channel equalization method based on deep learning according to claim 1, characterized in that, The specific method for step 1 is as follows: Step 1.

1. Select the bit rate and modulation scheme of the transmitted signal; based on the selected modulation scheme, first randomly generate the corresponding raw data sequence, and then modulate the generated raw data sequence into the corresponding symbols to generate the transmission signal; save these raw data sequences as tags for use in subsequent training processes; Step 1.

2. Perform shaping filtering on the transmission signal generated in Step 1.1, modulate the filtered signal into a form that meets communication requirements, and then select a communication channel to simulate the transmission of the signal that meets the communication requirements. Finally, the output signal after transmission through the channel constitutes a dataset, which is divided into a training set, a validation set, and a test set for training, validating, and testing the deep learning equalizer.

3. The channel equalization method based on deep learning according to claim 2, characterized in that, Step 1.1, which involves selecting the bit rate and modulation scheme of the transmitted signal, specifically involves: assuming the original binary data sequence is a bit stream of length K: b=[b1,b2,...,b K ],b i ∈{0,1} Choose a modulation scheme to map each log2(M) bits to a complex constellation symbol: Where M is the modulation order, and N = K / log2(M) is the total number of symbols; complex symbols can be decomposed into in-phase (I) and quadrature (Q) components:

4. The channel equalization method based on deep learning according to claim 2, characterized in that, Step 1.2 involves shaping and filtering the transmitted signal generated in step 1.1, modulating the filtered signal into a form that meets communication requirements, and then selecting a communication channel to simulate the transmission of this signal. Specifically: Let the expected output symbol length of the deep learning equalization network be L, then the transmission signal S generated in step 1.1... IQ Reshaped according to symbol length L in For batch quantity: Next, SPS interpolation is performed on each symbol, a shaping filter is used to generate the baseband signal, and the signal is discretized to obtain a symbol sequence. The symbol sequence is then convolved with the filter to output the shaped signal: Next, the forming signal is transmitted through a wireless channel with noise added: After separating the real and imaginary components of the received signal after passing through a wireless channel and adding noise, the received signal that meets the communication requirements is obtained as follows: Receive the signal y that meets the communication requirements IQ Remodeled according to SPS×L The final dataset consists of input-output pairs:

5. The channel equalization method based on deep learning according to claim 1, characterized in that, The specific method for step 3 is as follows: Step 3.1 Input the training set into the deep learning equalization network. The deep learning equalization network learns the signal equalization features in the training set data and gradually optimizes the model parameters. During training, the mean squared error (MSE) is used as the loss function, defined as: Where n1 is the number of samples in the training set. It is the true value of the i-th sample in the training set. It is the predicted value of the i-th sample in the training set. The loss function is used to measure the difference between the predicted value and the true value. By minimizing the loss function, the deep learning equalization network can better recover the original characteristics of the transmission signal generated in step 1.

1. The Adam optimizer is used for parameter updates to improve training efficiency and accelerate convergence. The update rules of the Adam optimizer are as follows: Where, θ t θ represents the current deep learning equilibrium network parameters. t+1 This refers to the deep learning balanced network parameters that will be updated in the next moment, m. t and v t These are the first and second moment estimates of the gradient, respectively, where η is the learning rate and e is a constant to prevent division by zero errors. Step 3.2 In the validation phase, the validation set is used to monitor the training effect and prevent overfitting. For the validation set, the mean absolute error (MAE) is used as the evaluation metric. Using an early stopping strategy, when the validation set has been in k consecutive rounds, the current time L MAE (t) compared to the previous time L MAE (t-1) are all less than the set threshold μ, that is At this point, the deep learning equilibrium network is considered to have converged. To avoid overfitting, a validation set L is selected. MAE The deep learning equilibrium network parameter with the smallest value is taken as the final deep learning equilibrium network parameter. Step 3.3 Input the received signals from the test set, after passing through the wireless channel and adding noise as in Step 1.2, into the deep learning equalization network. The deep learning equalization network outputs the corresponding prediction signal. For the transmitted signals from the test set that have not passed through the channel and the prediction signal from the deep learning equalization network, calculate the mean square error (MSE) and mean square error (MAE) respectively according to the mean square error (MSE) calculation formula in Step 3.1 and the mean square error (MAE) calculation formula in Step 3.

2. If the error between the calculated mean square error (MSE) and the optimal mean square error (MSE) in the training phase in Step 3.1 is less than the set value μ1, and the error between the calculated mean square error (MAE) and the optimal mean square error (MAE) in the validation phase in Step 3.2 is less than the set value μ2, then the model has good generalization performance.

6. The channel equalization system based on the deep learning-based channel equalization method according to any one of claims 1 to 5, characterized in that, include: The dataset construction module is used in step 1. It constructs a complete data stream through steps such as bit stream generation, modulation, shaping filtering, analog transmission and channel transmission, realizes dataset construction, and divides the dataset into training set, validation set and test set. The equalization network construction module is used in step 2. It constructs a deep learning equalization network based on a convolutional autoencoder (CAE) and integrates a residual network structure (ResNet), a self-attention mechanism, layer normalization (LayerNorm), and the Swish activation function to achieve channel equalization. The equalization network training module is used in step 3. By inputting the training set into the deep learning equalization network, it learns equalization features, uses mean squared error (MSE) as the loss function, and uses the Adam optimizer to update parameters, thereby improving training efficiency and accelerating convergence speed, and realizing network training. The equalization network validation module is used in step 3 to monitor the training effect of the deep learning equalization network through the validation set, prevent overfitting, and use the mean absolute error (MAE) to evaluate the performance of the deep learning equalization network on the validation set, so as to realize the dynamic adjustment and optimization of the training strategy. The equalization network testing module is used in step 3 to evaluate the deep learning equalization network using a test set, calculate the error of the deep learning equalization network on the test set, including mean squared error (MSE) and mean absolute error (MAE), and test the generalization ability and equalization performance of the deep learning equalization network, so as to achieve a comprehensive test and evaluation of the performance of the deep learning equalization network.

7. The channel equalization device according to any one of claims 1 to 5 based on the deep learning-based channel equalization method, characterized in that, include: Memory, used to store computer programs; A processor, used to implement deep learning-based channel equalization when executing the computer program.

8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it can implement deep learning-based channel equalization according to any one of claims 1 to 5.

Images