A Channel Equalization Method, Computer Device, and Readable Storage Medium Based on Convolutional Recurrent Neural Network

Abstract

A channel equalization method, computer device, and readable storage medium based on a convolutional recurrent neural network belong to the field of communication technologies, and solve the problems of difficult control of convergence error and long convergence process. The method of the present invention includes: collecting and preprocessing a data set, which is divided into a training set, a validation set, and a test set; establishing a channel equalizer based on a convolutional recurrent neural network; determining the parameters of the channel equalizer based on the convolutional recurrent neural network; and equalizing data using the trained channel equalizer based on the convolutional recurrent neural network. The present invention is mainly used in the field of channel equalization for frequency-selective fading channels, especially for channel equalization of Rayleigh multipath channels.

Description

Technical Field

[0001] This application relates to the field of communication technologies, and in particular, to a channel equalization method, a computer device, and a readable storage medium based on a convolutional recurrent neural network. Background Art

[0002] In a wireless communication system, the channel is usually time-varying. Since a preset equalizer with constant tap coefficients cannot track a time-varying channel, it performs poorly in a wireless scenario. In order to be able to respond to channel changes in real time, an adaptive equalizer that can update parameters needs to be used.

[0003] Deep learning has excellent learning ability and data recognition ability. In some fields, as long as partial prior information is given, the neural network can obtain a good fitting effect. Currently, in the field of communication, many scholars have used deep learning theory to conduct related research on the physical layer and achieved certain results. Schemes for using deep learning for channel equalization in a receiver also emerge in an endless stream.

[0004] Scholars have early studied using neural networks for channel equalization. As early as 1995, some scholars proposed a neural network-based decision feedback equalizer to process complex phase shift keying signals on a complex nonlinear satellite channel, maintaining almost the same computational complexity as a real-valued training algorithm.

[0005] In subsequent developments, more and more neural networks have been used in the field of wireless channel equalization, including single-layer Chebyshev neural networks, multi-layer perceptrons, functional link artificial neural networks, radial basis functions, etc. However, the above-mentioned schemes all have certain deficiencies, and the main common defects include the following two points:

[0006] First, it is difficult to control the convergence error. The selection of network parameters, such as the learning rate, activation function, number of neurons, and the connection method of the network, all have a direct impact on the final training effect, and the training result is directly related to subsequent applications. This requires researchers to have relatively proficient usage experience to quickly select reasonable parameters and structures.

[0007] Second, the neural network structure is relatively complex and the training has a longer convergence process. If the selection of parameters or optimization algorithms is inappropriate, the convergence process may be further lengthened. In a real-time communication system, any operation that may introduce excessive delay is intolerable. Summary of the Invention

[0008] The purpose of the present invention is to solve the problems of difficult control of existing convergence error and long convergence process, and provide a channel equalization method, a computer device, and a readable storage medium based on a convolutional recurrent neural network.

[0009] The present invention is implemented through the following technical solutions. On the one hand, the present invention provides a channel equalization method based on a convolutional recurrent neural network, and the method includes:

[0010] Step 1, generate and process the original data sequence, obtain the data set, and divide the data set into a training set, a validation set, and a test set;

[0011] Step 2, establish a channel equalizer based on a convolutional recurrent neural network, specifically including:

[0012] The channel equalizer of the convolutional recurrent neural network includes an input module, a two-dimensional feature extraction module, a temporal feature extraction module, and an output layer, wherein,

[0013] The input module includes an input layer and a fully connected layer;

[0014] The two-dimensional feature extraction module includes a convolutional layer, a pooling layer, and batch normalization;

[0015] The temporal feature extraction module is constructed by using a long short-term memory artificial neural network;

[0016] The output layer includes two fully connected layers and Dropout;

[0017] Step 3, determine the parameters of the channel equalizer based on the convolutional recurrent neural network through simulation, and the parameters include the training set length, the number of training epochs, and the learning rate;

[0018] Step 4, use the data set to train the channel equalizer based on the convolutional recurrent neural network with the determined parameters;

[0019] Step 5, use the trained channel equalizer based on the convolutional recurrent neural network to equalize the data.

[0020] Further, in step 1, the generating and processing the original data sequence specifically includes:

[0021] Generate and process the original data sequence according to the bit rate of the transmitted signal and the signal modulation method.

[0022] Further, in step 1, the generating and processing the original data sequence, obtaining the data set specifically includes:

[0023] Step 1.1, generate the original data sequence according to the modulation method, modulate it into corresponding symbols, and save the original data sequence as a label;

[0024] Step 1.2, perform channel modeling according to the channel parameters, calculate the number of symbols that can be transmitted within a coherence time according to the coherence time and the signal rate, and provide guidance for the selection of the training set length;

[0025] Step 1.3: Estimate the number of taps of the equalizer according to the multipath delay information;

[0026] Step 1.4: Pass the data through the channel to obtain the fading signal. According to the number of taps N, perform N cyclic shifts on the data sequence, separate the real and imaginary parts of the signal, and then splice them to obtain the data set.

[0027] Further, the input layer in Step 2 is specifically:

[0028] The input layer is responsible for feeding the data set into the neural network and performing data shaping to adjust the positional relationship between the real and imaginary parts of the data. The data is reshaped from 1×2N to 2×N.

[0029] Further, the two-dimensional feature extraction module in Step 2 is specifically:

[0030] The first convolutional layer can perform convolution on the output of the input layer to generate a feature map and extract shallow information;

[0031] Use average pooling to retain background information and prevent some key features from being discarded;

[0032] Batch normalization normalizes the output to make the input of the next layer have a more stable distribution, which is beneficial to the training of the network;

[0033] The second convolutional layer is used to extract deep information;

[0034] Use max pooling to retain boundary information and reduce redundant information.

[0035] Further, the long short-term memory artificial neural network in Step 2 is specifically:

[0036] The LSTM layer of the long short-term memory artificial neural network models time, learns the time-varying characteristics of the channel, and memorizes the state of the previous moment as part of the input at this time.

[0037] Further, the Dropout in Step 2 is specifically: Dropout is a regularization method. Its role is to discard part of the input of the upper layer and add Dropout between layers to reduce overfitting.

[0038] Further, Step 3 specifically includes:

[0039] Step 3.1: Select the length of the training set according to the bit error rate and high code rate;

[0040] Step 3.2: Select the number of training epochs according to the number of error symbols and training time;

[0041] Step 3.3: Select the learning rate according to the convergence error and the momentum factor.

[0042] In a second aspect, the present invention provides a computer device, including a memory and a processor. A computer program is stored in the memory. When the processor runs the computer program stored in the memory, the steps of a channel equalization method based on a convolutional recurrent neural network as described above are executed.

[0043] In a third aspect, the present invention provides a computer-readable storage medium. Multiple computer instructions are stored in the computer-readable storage medium. The multiple computer instructions are used to cause a computer to execute a channel equalization method based on a convolutional recurrent neural network as described above.

[0044] Advantages of the present invention:

[0045] First of all, referring to the system model of the multi-layer perceptron equalizer, the present invention improves the network structure and proposes a system model of a convolutional recurrent neural network. Combining the input method of the traditional channel equalizer and the characteristics of the convolutional neural network, the data set is shaped to facilitate feature extraction by the network. In terms of the neural network, the present invention proposes a brand-new convolutional recurrent network structure including four modules. The convolutional layer is used for two-dimensional feature extraction, reducing the data dimension while extracting deep information. The long short-term memory artificial neural network (LSTM) is used for temporal feature extraction to discover the correlation between symbols.

[0046] In addition, the present invention selects the key parameters of the channel equalizer based on the convolutional recurrent neural network through simulation, selects appropriate training set length, number of training epochs and learning rate, speeds up the convergence speed and reduces the convergence error.

[0047] The present invention is mainly used in the field of channel equalization for frequency-selective fading channels, especially for channel equalization in Rayleigh multipath channels. Description of the Drawings

[0048] In order to more clearly illustrate the technical solutions of the present application, the drawings required for use in the embodiments will be briefly introduced below. Obviously, for those of ordinary skill in the art, other drawings can also be obtained based on these drawings without creative efforts.

[0049] Figure 1 is the system model of the convolutional recurrent neural network equalizer;

[0050] Figure 2 is the pipelined convolutional recurrent neural network structure;

[0051] Figure 3Schematic diagram of the influence of the training set length on the bit error rate;

[0052] Figure 4 Schematic diagram of the influence of the number of training rounds on the number of error symbols;

[0053] Figure 5 Four groups of mean square error simulation curves of the training set;

[0054] Figure 6 Four groups of mean square error simulation curves of the validation set;

[0055] Figure 7 QPSK signal with severe distortion of the constellation diagram;

[0056] Figure 8 Constellation diagram of the QPSK signal after channel equalization using the present invention;

[0057] Figure 9 Bit error rate performance of the present invention. Specific embodiments

[0058] The following details the embodiments of the present invention. The examples of the embodiments are shown in the drawings, where the same or similar reference numerals represent the same or similar elements or elements with the same or similar functions throughout. The embodiments described below with reference to the drawings are exemplary and are intended to explain the embodiments and should not be construed as limiting the embodiments.

[0059] Embodiment 1. A channel equalization method based on a convolutional recurrent neural network, the method comprising:

[0060] Step 1, generate and process the original data sequence, obtain a data set, and divide the data set into a training set, a validation set, and a test set;

[0061] It should be noted that this Step 1 is for the generation and processing of the data set. In this Step 1, the training set is the training sequence, and the network parameters can be trained and updated using the training set; the validation set is responsible for testing the generalization and robustness of the system during the training process to determine whether overfitting occurs; the test set can simulate the information sequence and is used to test the trained network to illustrate the final performance of the equalizer.

[0062] Step 2, establish a channel equalizer based on a convolutional recurrent neural network, specifically including:

[0063] Establish a convolutional recurrent neural network, the convolutional recurrent neural network including an input module, a two-dimensional feature extraction module, a temporal feature extraction module, and an output layer, where,

[0064] The input module includes an input layer and a fully connected layer;

[0065] The two-dimensional feature extraction module includes a convolutional layer, a pooling layer, and batch normalization;

[0066] The temporal feature extraction module is constructed using a long short-term memory artificial neural network;

[0067] The output layer includes two fully connected layers and Dropout;

[0068] The final network output is the real and imaginary parts of the equalized signal.

[0069] It should be noted that in this step 2, it is the network structure and module design. Both convolutional neural networks and recurrent neural networks have better non-linear characteristics. And recurrent neural networks have very good performance when processing time series signals. However, the defect of recurrent neural networks is that the algorithm complexity is relatively high and the training time is long. Due to the existence of parameter sharing and pooling layers in convolutional neural networks, they can reduce the data dimension and shorten the training time. Connecting the two networks in series can not only pay attention to the correlation between adjacent symbols in the sequence, effectively improve the network's ability to handle non-linear problems, but also reduce the training time to a certain extent and improve the training efficiency. Therefore, in this embodiment, a pipeline convolutional recurrent neural network is designed, and the network structure is as Figure 2 shown.

[0070] Step 3, determine the parameters of the channel equalizer based on the convolutional recurrent neural network through simulation, and the parameters include the training set length, the number of training epochs, and the learning rate;

[0071] It should be noted that in this step 3, it is the equalizer parameter selection. In the iterative process of the neural network, the problem is fitted by adjusting the network weights. However, in addition to the weights, many other parameters will also affect the iterative effect of the network. Specifically in this embodiment, the main parameters affecting the training results include the selection of the training set length l, the number of training epochs epoch, and the learning rate α. With other parameters fixed, simulate the above three parameters and select the values with the best equalization performance.

[0072] Step 4, use the data set to train the channel equalizer based on the convolutional recurrent neural network with the determined parameters;

[0073] Step 5, use the trained channel equalizer based on the convolutional recurrent neural network to equalize the data.

[0074] In this embodiment, a new network model is used to propose a channel equalization scheme based on a convolutional recurrent neural network. The scheme mainly includes the generation and processing of the data set and the optimization of the network structure and parameters. Figure 1 Shows the system model of the convolutional recurrent neural network equalizer.

[0075] The system model of this embodiment is similar to a multi-layer perceptron equalizer and has a training mode and a direct decision mode. In the training mode, a known sequence is sent and used as the desired output at the receiving end. The network parameters are iteratively updated by the least mean square error algorithm. After the network training is completed, it switches to the direct decision mode, where the network parameters remain unchanged. The transmitter sends the information sequence, and the received signal at the receiving end is directly fed into the trained neural network to obtain the network output, and then a decision is made. (Overview of this embodiment)

[0076] Compared with the multi-layer perceptron with a simple structure, due to the two characteristics of local perception and parameter sharing of the convolutional neural network, the complexity of the network model is greatly reduced, the number of weights is also significantly reduced, it has better recognition ability for similar features in the data, has better non-linear characteristics and generalization ability, and is very suitable for the modern digital communication mode of high-speed information transmission in wireless non-linear channels. For the received signal with inter-symbol interference, due to the multi-path effect, the symbol being decided at a certain moment will be interfered by some symbols before and after. Therefore, there is a certain connection between several adjacent symbols in the received sequence. The recurrent neural network has memory, can pay attention to the connection between symbols, and has a strong learning ability for the non-linear characteristics of the sequence. Therefore, it is also very suitable for channel equalization in frequency-selective fading channels.

[0077] Embodiment 2 further limits a channel equalization method based on a convolutional recurrent neural network described in Embodiment 1. In this embodiment, the generation and processing of the original data sequence in step 1 are further limited, specifically including:

[0078] Generate and process the original data sequence according to the bit rate and signal modulation method of the transmitted signal.

[0079] Before generating the data set in this embodiment, it is necessary to clarify the bit rate and signal modulation method of the transmitted signal, which gives the basis for the generation and processing of the original data sequence.

[0080] This embodiment combines the input method of the traditional channel equalizer and the characteristics of the convolutional neural network to shape the data set, which is convenient for the network to extract features.

[0081] Embodiment 3 further limits a channel equalization method based on a convolutional recurrent neural network described in Embodiment 1. In this embodiment, the generation and processing of the original data sequence and the acquisition of the data set in step 1 are further limited, specifically including:

[0082] Step 1.1: Generate the original data sequence according to the modulation method, modulate it into corresponding symbols, and save the original data sequence as a label.

[0083] Step 1.2: Perform channel modeling according to the channel parameters. Calculate the number of symbols that can be transmitted within one coherence time based on the coherence time and signal rate, which provides guidance for the selection of the training set length.

[0084] Step 1.3: Estimate the number of taps of the equalizer according to the multipath delay information.

[0085] Step 1.4: Pass the data through the channel to obtain the fading signal. According to the number of taps N, perform N cyclic shifts on the data sequence, separate the real and imaginary parts of the signal, and then splice them to obtain the data set.

[0086] In this embodiment, if the selected training set length is L, after splitting the real and imaginary parts and shifting and splicing, a data matrix of L×2N can be obtained. Then, by performing "trimming" processing to remove the unusable data, the final data set matrix can be obtained.

[0087] This embodiment combines the input method of the traditional channel equalizer and the characteristics of the convolutional neural network to shape the data set, which is convenient for the network to extract features.

[0088] Embodiment 4: This embodiment further limits the channel equalization method based on the convolutional recurrent neural network described in Embodiment 1. In this embodiment, the input layer in Step 2 in Step 1 is further limited as follows:

[0089] The input layer is mainly responsible for feeding the data set into the neural network and performing data shaping to adjust the positional relationship between the real part data and the imaginary part data. The data is reshaped from 1×2N to 2×N.

[0090] The input layer of this embodiment can facilitate subsequent input to the convolutional layer.

[0091] Embodiment 5: This embodiment further limits the two-dimensional feature extraction module in Step 2 in the method for channel equalization based on the convolutional recurrent neural network described in Embodiment 1. Specifically:

[0092] The first convolutional layer can perform convolution on the output of the input layer to generate a feature map and extract shallow information.

[0093] Use average pooling to retain background information and prevent some key features from being discarded.

[0094] Batch normalization normalizes the output to make the input of the next layer have a more stable distribution, which is beneficial to the training of the network;

[0095] The second convolutional layer is used to extract deep - level information;

[0096] Use max - pooling to retain boundary information and reduce redundant information.

[0097] In this embodiment, the two - dimensional feature extraction module consists of a convolutional layer, a pooling layer, and batch normalization. The first convolutional layer can perform convolution on the output of the input layer to generate a feature map, extracting shallow - level information such as amplitude and phase. Subsequently, average - pooling is used to retain more background information to prevent some key features from being discarded. Batch normalization normalizes the output to make the input of the next layer have a more stable distribution, which is beneficial to the training of the network. The second convolutional layer is similar to the first convolutional layer, but can extract deeper - level information such as Doppler frequency shift. Subsequently, max - pooling is used to retain more boundary information and reduce redundant information. Through two convolutions, a set of filters that can overcome characteristics such as signal rotation, time - shift, and scaling can be learned, which are used for the restoration of wireless signals with inter - symbol interference.

[0098] Embodiment 6: This embodiment further limits the long short - term memory artificial neural network in step 2 of the method for channel equalization based on a convolutional recurrent neural network described in Embodiment 1. Specifically:

[0099] The LSTM layer of the long short - term memory artificial neural network models time, learns the time - varying characteristics of the channel, and remembers the state of the previous moment and uses it as part of the input at this time.

[0100] In this embodiment, the time - series feature extraction module is mainly completed by using a long short - term memory artificial neural network. The LSTM layer can model time and learn the time - varying characteristics of the channel. The LSTM layer can remember the state of the previous moment and use it as part of the input at this time, so as to find the time - correlation information of the sequence and help signal restoration.

[0101] Embodiment 7: This embodiment further limits the Dropout in step 2 of the method for channel equalization based on a convolutional recurrent neural network described in Embodiment 1. Specifically:

[0102] Dropout is a regularization method, and its function is to discard part of the input of the upper layer. Adding Dropout between layers reduces the over - fitting phenomenon.

[0103] In this embodiment, the final network output is the real part and the imaginary part of the equalized signal.

[0104] Embodiment 8. This embodiment further defines a channel equalization method based on a convolutional recurrent neural network described in Embodiment 1. In this embodiment, Step 3 is further defined, specifically including:

[0105] Step 3.1: Select the training set length according to the bit error rate and the high code rate;

[0106] Step 3.2: Select the number of training rounds according to the number of error symbols and the training time;

[0107] Step 3.3: Select the learning rate according to the convergence error and the momentum factor.

[0108] In this embodiment, in a digital communication system adopting channel equalization technology, in order to meet the requirements of real-time communication, the time to complete one iteration should be less than the symbol interval. An overly long training sequence will lead to increased redundancy, reduced code rate, and lower spectrum utilization, while an overly short training sequence will result in insufficient training and underfitting. Therefore, the selection of the training sequence length in this embodiment is to obtain better equalization performance as much as possible on the premise of ensuring spectrum utilization. The design idea of the training sequence is as follows:

[0109] a) Calculate the number of symbols that can be transmitted within the coherence time according to the channel coherence time and the signal transmission rate;

[0110] b) Each training sequence is configured with at least five information sequences to ensure a high code rate and spectrum utilization, and calculate the maximum training sequence length;

[0111] c) On the premise of meeting the maximum training sequence length, select the sequence length that can obtain the best equalization performance.

[0112] Figure 3 Shows the influence of the training set length on the bit error rate. When the transmitted signal-to-noise ratios are 10 dB and 14 dB respectively, the bit error rate of the equalized signal decreases as the training set length increases. When the training set length is greater than 100,000, the curve decline trend becomes slow and gradually converges. Selecting the training set length of 500,000 can not only ensure a high code rate, improve the effectiveness of the communication system, but also obtain good equalization performance, ensuring good reliability of the entire communication system.

[0113] Figure 4It shows the influence of the number of training epochs on the number of error symbols. When epoch = 30, the number of error symbols is the least. However, when epoch = 20, the increase in the number of error symbols compared to epoch = 30 is relatively small, only increasing by 10.6%, which has little impact on the system bit error rate performance, but can save 33.3% of the training time. After comprehensively considering the balance between performance and training time, the best balance effect can be obtained when the number of training epochs epoch = 20.

[0114] Figure 5 It shows the four mean square error simulation curves of the training set. From the simulation results, it can be seen that as the number of training epochs increases, the four mean square errors of the training set all show a stable downward trend and gradually tend to converge. When α = 0.001, the convergence speed of the training is the slowest, and there is the largest mean square error after convergence. The possible reason is that the training does not converge to the minimum value point but converges to a local minimum value. Due to the too small learning rate, it is unable to jump out of the minimum value, resulting in a relatively large mean square error in the end. When α = 0.01, the MSE curve is in a stepped pattern. In the early stage of training, the same situation as when α = 0.001 occurs, and the curve has a convergence trend, probably also falling into a local minimum. However, due to the relatively large learning rate, it successfully jumps out of this minimum value point in the subsequent training and conducts a second convergence, and finally converges to the minimum value point. When α = 0.1, the convergence speed is very fast, and the curve is very smooth, indicating that the minimum value is avoided during the convergence process and it directly converges to the vicinity of the minimum value. After introducing the momentum factor, that is, when α = 0.01, β = 0.9, the MSE curve converges the fastest and is always below the other three curves, with the smallest convergence error, and the training effect is the best at this time.

[0115] Figure 6 It shows the four mean square error simulation curves of the validation set. For the validation set, since the data is different from the training set data, the four curves all show different degrees of oscillation. However, as the number of training epochs increases, the oscillation amplitude of the other three curves except the mean square error curve of α = 0.1 weakens. The reason for the severe up and down fluctuations of the curve when α = 0.1 is that the training step size is too large, oscillating back and forth near the maximum and minimum points and unable to converge stably. By simulating the mean square error of the validation set, it can be seen that the network structure and training parameters of this embodiment are relatively appropriate, the regularization is sufficient, and there is no overfitting phenomenon. Moreover, after introducing the momentum term, the convergence effect of the validation set is also very good, indicating that the network has strong generalization and robustness. In addition, when choosing α = 0.01, β = 0.9, it only takes 3 - 5 rounds of training to converge to the vicinity of the minimum value.

[0116] Finally, based on the channel equalization method based on convolutional recurrent neural network described above, the reference dataset processing method, network structure and parameter settings, after selecting the optimal value, the performance of this embodiment is simulated and verified. The simulation includes two parts. The first part is the constellation diagram simulation, through which the correction effect of the convolutional recurrent neural network equalizer on the constellation diagram of the fading signal can be intuitively observed. The second part is the bit error rate simulation. For fading signals with different signal-to-noise ratios, the convolutional recurrent neural network equalizer designed in this embodiment is used for channel equalization. The output of the equalizer is judged and mapped to the corresponding ideal constellation points. After demodulation, it is compared with the original transmitted bit sequence, the bit error rate is calculated, and then plotted into a bit error rate curve. By comparing with the traditional decision feedback equalizer, it is proved that the convolutional recurrent neural network equalizer has better performance when processing wireless signals with frequency selective fading.

[0117] Figure 7 The QPSK signal with severely distorted constellation diagram is shown. The reason for the distortion is that the signal passes through a frequency selective fading channel with multipath effect and Doppler frequency shift without channel equalization. At this time, the distribution of constellation points is irregular and very scattered.

[0118] Figure 8 This is the constellation diagram of the QPSK signal after channel equalization using this embodiment. It can be seen that after parameter optimization, the convolutional recurrent neural network equalizer can very effectively restore the constellation diagram and has strong correction ability for the distorted constellation diagram.

[0119] Figure 9 The bit error rate performance of this embodiment is shown. Among them, the CRNN equalizer is the convolutional recurrent neural network equalizer. Compared with the multi-layer perceptron (MLP) equalizer and the decision feedback equalizer (DFE), this embodiment has better bit error rate performance. When the signal-to-noise ratio is 10-18 dB, the convolutional recurrent neural network equalizer has a performance improvement of 1-2 dB compared with the decision feedback equalizer when reaching the same bit error rate.

Claims

1. A channel equalization method based on a convolutional recurrent neural network, characterized in that, The method includes: Step 1: Generate and process the original data sequence, obtain the data set, and divide the data set into a training set, a validation set, and a test set; Step 2: Establish a channel equalizer based on a convolutional recurrent neural network, specifically including: The channel equalizer of the convolutional recurrent neural network includes an input module, a two-dimensional feature extraction module, a temporal feature extraction module, and an output layer, where The input module includes an input layer and a fully connected layer; The two-dimensional feature extraction module includes a convolutional layer, a pooling layer, and batch normalization; The temporal feature extraction module is constructed using a long short-term memory artificial neural network; The output layer includes two fully connected layers and Dropout; Step 3: Determine the parameters of the channel equalizer based on the convolutional recurrent neural network through simulation, where the parameters include the training set length, the number of training epochs, and the learning rate; Step 4: Use the data set to train the channel equalizer based on the convolutional recurrent neural network with the determined parameters; Step 5: Use the trained channel equalizer based on the convolutional recurrent neural network to equalize the data; The two-dimensional feature extraction module includes a convolutional layer, a pooling layer, and batch normalization. Specifically: The two-dimensional feature extraction module sequentially includes a convolutional layer, an average pooling layer, BN, a convolutional layer, a max pooling layer, and BN. Specifically: The first convolutional layer convolves the output of the input layer to generate a feature map and extract shallow information, where the shallow information includes amplitude and phase; Subsequently, average pooling is used to retain background information to prevent some key features from being discarded, and batch normalization normalizes the output; The second convolutional layer extracts deeper information, where the deeper information includes Doppler frequency shift. Subsequently, max pooling is used to retain boundary information and reduce redundant information. Through two convolutions, it is used for the restoration of wireless signals with inter-symbol interference; The temporal feature extraction module is constructed using a long short-term memory artificial neural network. Specifically: The temporal feature extraction module sequentially includes an LSTM layer and BN; The output layer includes two fully connected layers and Dropout. Specifically: The output layer sequentially includes a fully connected layer, Dropout, and a fully connected layer; The role of Dropout is to discard some of the inputs of the upper layer. Adding Dropout between layers reduces the overfitting phenomenon. The final network output is the real part and the imaginary part of the equalized signal; In Step 1, specifically including: Step 1.1: Generate the original data sequence according to the modulation method, modulate it into corresponding symbols, and save the original data sequence as a label; Step 1.2: Perform channel modeling according to the channel parameters, calculate the number of symbols that can be transmitted within a coherence time according to the coherence time and the signal rate, and provide guidance for the selection of the training set length; Step 1.3: Estimate the number of taps of the equalizer according to the multipath delay information; Step 1.4: Pass the data through the channel to obtain the fading signal. According to the number of taps , perform times of cyclic shift on the data sequence, separate the real part and the imaginary part of the signal, and then splice them together.

2. The channel equalization method based on a convolutional recurrent neural network according to claim 1, wherein In Step 1, the generation and processing of the original data sequence specifically include: Generate and process the original data sequence according to the bit rate and signal modulation method of the transmitted signal.

3. The channel equalization method based on a convolutional recurrent neural network according to claim 1, wherein The input layer in Step 2 is specifically: The input layer is responsible for feeding the dataset into the neural network and performing data shaping to adjust the positional relationship between the real part data and the imaginary part data. The data is reshaped from to .

4. A channel equalization method based on a convolutional recurrent neural network according to claim 1, characterized in that The long short-term memory artificial neural network in Step 2 is specifically as follows: The LSTM layer of the long short-term memory artificial neural network models time, learns the time-varying characteristics of the channel, and memorizes the state of the previous moment and uses it as part of the input at this time.

5. A channel equalization method based on a convolutional recurrent neural network according to claim 1, characterized in that Step 3 specifically includes: Step 3.1: Select the training set length according to the bit error rate and the high code rate; Step 3.2: Select the number of training epochs according to the number of error symbols and the training time; Step 3.3: Select the learning rate according to the convergence error and the momentum factor.

6. A computer device, comprising a memory and a processor, wherein a computer program is stored in the memory, characterized in that, When the processor runs the computer program stored in the memory, it executes the steps of the method according to any one of claims 1 to 5.

7. A computer-readable storage medium, characterized in that, Multiple computer instructions are stored in the computer-readable storage medium, and the multiple computer instructions are used to cause the computer to execute the method according to any one of claims 1 to 5.

Images