Convolutional cyclic network channel equalization system based on data enhancement
Through the convolutional cyclic network channel equalization system based on data enhancement, the problems of robustness and low spectrum utilization of channel equalization in complex communication environments are solved, and effective correction of distorted signals and improved bit error rate are achieved.
Patent Information
- Application Number
- CN202510890194.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-30
- Publication Date
- 2025-09-05
AI Technical Summary
The existing adaptive equalization technology algorithm has low spectrum utilization in complex communication environments and cannot effectively suppress intersymbol interference and noise in the channel, resulting in a decline in communication quality.
The convolutional recurrent network channel equalization system based on data augmentation is adopted, including a sample generation module, a convolutional recurrent neural network feature extraction module, a data augmentation training module and a discrimination module. By generating IQ timing signals, feature extraction and data augmentation training, the channel equalization performance is improved.
Effectively correct distorted signals in complex communication channel environments, improve the robustness and balance ability of channel equalization, and improve the bit error rate performance of the communication system.
Smart Images
Figure CN120602276A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of digital communication technology, and in particular to a convolutional cyclic network channel equalization system based on data enhancement. Background Art
[0002] With the rapid advancement of technology, the demand for digital communications is growing, requiring high-speed data transmission over channels with limited bandwidth. Factors such as multipath and additive noise in wireless communications can cause inter-symbol interference (ISI) at the receiving end, severely impacting communication quality. To improve communication quality and mitigate the negative impact of ISI, channel equalization techniques are often employed at the receiving end to enhance wireless channel transmission performance. Traditional adaptive equalization algorithms, primarily the least mean square (LMS) algorithm and the recursive least squares (RLS) algorithm, require sufficient training sequences to achieve effective equalization. This results in low spectrum efficiency, hindering widespread adoption and application in increasingly complex and volatile communication environments.
[0003] Deep learning, a highly complex machine learning technology, possesses powerful computational capabilities, enabling computers to achieve human-like perception and cognitive processes. Deep learning models can capture the temporal and spatial characteristics of the channel, effectively suppressing intersymbol interference and noise within the channel, thereby improving the bit error rate performance of communication systems. Furthermore, deep learning offers end-to-end processing capabilities, integrating equalization with tasks such as demodulation and detection. This simplifies the system architecture and improves the efficiency and robustness of equalization. It demonstrates high efficiency in handling both linear and nonlinear transmission channels, and its rapid convergence and excellent interference suppression capabilities are noteworthy. Deep neural networks utilize a distributed architecture, so if local structures are damaged, the system's operation will only be moderately slowed, but not catastrophically.
[0004] In summary, the channel equalization method based on deep neural networks has strong feasibility and practicality. In order to further improve the equalization ability of deep neural networks for distorted signals, a channel equalization method based on data-enhanced convolutional recurrent neural networks is proposed on the basis of drawing on the existing deep neural network model. This method effectively equalizes the distorted signal output by the channel, improves the equalization reliability of channel equalization in real environments, and thus meets the requirements for channel equalization systems in complex channel environments. Summary of the Invention
[0005] The technical problem to be solved by this invention is to address the shortcomings of the above-mentioned existing technologies and provide a data-augmented convolutional recurrent network channel equalization system. This system solves the problem of equalizing distorted signals at the receiving end using deep neural networks in complex communication channel environments. By designing a sample generation module, a convolutional recurrent neural network feature extraction module, a data-augmented training module, and a discrimination module, the system improves equalization performance in complex communication environments.
[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is:
[0007] A convolutional recurrent network channel equalization system based on data enhancement, including a sample generation module, a convolutional recurrent neural network feature extraction module, a data enhancement training module and a discrimination module;
[0008] The sample generation module is used to generate an IQ timing signal; generate a distorted signal sample using a transmitted signal X and its sample label y; X is a segment of original transmitted data for network training, with a dimension of 1×64, expressed as: The dimension of the sample label y is 64×4. The nth symbol is represented as a feature matrix with nM to n+M symbols and its real and imaginary components are extracted, where M is the number of adjacent symbols. The generated samples are divided into a training set and a test set in a ratio of 7:3.
[0009] The convolutional recurrent neural network feature extraction module is used to capture key information in the training signal and perform dimensionality reduction and denoising operations simultaneously. The convolutional recurrent neural network includes a convolution layer, a pooling layer, an LSTM layer, a BN layer, and an activation function. The time series related features of the distorted signal are extracted through a network consisting of three convolution layers in series, an LSTM layer, a BN layer, and an activation layer. The BN layer prevents overfitting and gradient vanishing, thereby improving network balancing performance. The LSTM layer integrates the features extracted from the first two convolution layers through its memory capacity and associates contextual information at different time steps.
[0010] The data augmentation training module sends individual samples as a batch to the neural network training and sets a data loader for each sample independently;
[0011] The discrimination module uses the test set to adjust the learning weights, and the convolutional recurrent neural network converts the output feature information into corresponding bit information, compares it with the decision threshold value, and obtains the equalized bit error rate (BER).
[0012] Furthermore, the specific operation process of the sample generation module is as follows:
[0013] Step 1.1: Generate B × Num_s bits, with each B bit group as a Gray coding mapping group; B represents the number of bits corresponding to the modulation symbol, and Num_s is the number of symbols in a single signal;
[0014] Step 1.2: Store the bits from group M+1 to group Num_s-M as the label y.
[0015] Step 1.3: Convert each group of bits into a decimal index, perform mapping encoding, and generate Num_s modulation symbols;
[0016] Step 1.4: After the modulation symbols pass through the channel, add Gaussian white noise;
[0017] Step 1.5: Oversample a single symbol with 4 sampling points.
[0018] Step 1.6: Put the nth symbol It is encapsulated with its adjacent M symbols to construct Characteristic matrix of symbols :
[0019] ;
[0020] in, The dimension is 4×(2M+1);
[0021] Step 1.7: Convert the feature matrix Convert it into an IQ timing signal, then perform a Cat operation to obtain X. After passing through the channel environment, use Dataset to concatenate it with the sample label y to form a distorted signal sample. The IQ timing-related features of the transmitted signal are expressed as:
[0022] ;
[0023] in, The dimension is 2×(2M+1), is the in-phase component, is the orthogonal component;
[0024] Step 1.8: Divide the generated distorted signal samples into a training set and a test set in a ratio of 7:3.
[0025] Furthermore, the steps of network training in the data enhancement training module are as follows:
[0026] Step 2.1: Load each sample in the training set into Num data loaders, where Num is the number of received signals;
[0027] Step 2.2: Connect the Num data loaders that have loaded the data;
[0028] Step 2.3: Initialize network parameters, including network weights With bias ;
[0029] Step 2.4: Extract the training data from the concatenated data loader and feed it into the convolutional recurrent network for training using the gradient descent algorithm and backpropagation mechanism;
[0030] The kth feature map output after the sample passes through the first convolutional layer , whose dimension is 4×7, is expressed as:
[0031] ;
[0032] in, Expressed as an activation function, Z is the received sample after the channel output, represents the batch normalization operation, represents the kth convolution kernel of the first convolutional layer, Is the bias introduced into the convolution result;
[0033] The kth feature map output after the sample passes through the second convolutional layer , whose dimension is 1×64, is expressed as:
[0034] ;
[0035] in, is the kth convolution kernel in the second convolution layer, For the corresponding offset, the feature map output by the first convolutional layer is encapsulated as ;
[0036] Step 2.5: Compare the predicted results with the true labels and calculate the loss function Loss:
[0037] ;
[0038] in, is the reconstructed signal after the neural network; y is the sample label of the received signal; H is the number of samples in the current batch;
[0039] Step 2.6: Repeat steps 2.3 to 2.5 to minimize the loss function Loss until the maximum number of iterations MAX is reached. Use the trained convolutional recurrent network model to infer the soft bit vector of the network output. Dataset;
[0040] Soft bit vector output by the network Expressed as:
[0041] ;
[0042] in, and They represent the kth convolution kernel and the corresponding bias in the third convolution layer, respectively, and encapsulate the feature map output by the second convolution layer as .
[0043] Furthermore, the specific operation process of the discrimination module is as follows:
[0044] Step 3.1: Input the test set into the trained network model to obtain the output soft bit vector Dataset;
[0045] Step 3.2: Output soft bit vector Dataset and threshold Compare, the soft bit vector greater than the threshold value The judgment is 1, and the judgment is 0 if it is less than the threshold value;
[0046] Step 3.3: Repeat steps 3.1 to 3.2, and record the number of discrepancies between all judged data and labels, denoted as c.
[0047] Step 3.4: After all test samples are compared, the equalized bit error rate (BER) is calculated, which is expressed as:
[0048] ;
[0049] Where c is the number of elements whose judgment output is different from the sample, and m is the total number of elements in the test sample.
[0050] The beneficial effects of adopting the above technical solution are: the data-enhanced convolutional recurrent network channel equalization system provided by the present invention simulates the real communication environment by constructing a small-scale fading channel model, and then uses deep neural network technology to extract the timing-related features of the signal. By preprocessing the received data to fully utilize the characteristic information of the received signal, the bit prediction is modeled as a binary classification problem by performing a convolution operation on the input array obtained by preprocessing to restore the transmitted data, thereby achieving the purpose of improving the robustness and equalization capability of the method. The system of the present invention can meet the requirements of achieving effective correction and equalization of distorted signals in complex communication channel environments, thereby improving the channel equalization capability of deep neural networks. BRIEF DESCRIPTION OF THE DRAWINGS
[0051] Figure 1 Schematic diagram of a convolutional recurrent network channel equalization system based on data enhancement provided by an embodiment of the present invention;
[0052] Figure 2 A diagram showing the structure of a convolutional cyclic equalization network provided by an embodiment of the present invention;
[0053] Figure 3 This is a block diagram of the data enhancement module training system provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0054] The following embodiments of the present invention are described in further detail with reference to the accompanying drawings and examples. The following examples are used to illustrate the present invention but are not intended to limit the scope of the present invention.
[0055] like Figure 1 As shown, the system of this embodiment involves four parts, namely, a sample generation module, a convolutional recurrent neural network-based feature extraction module, a data enhancement training module, and a discrimination module.
[0056] The sample generation module generates IQ time-series signals, enabling comprehensive analysis of signal characteristics. During signal transmission, the original transmitted sequence undergoes QPSK modulation. The modulated signal is a non-real signal. Given the characteristics of complex signals in actual communications, neural networks using IQ components can directly process complex data, simplifying data preparation and conversion. For tasks involving signal phase and amplitude (e.g., modulation and demodulation, channel estimation, etc.), IQ components provide a more natural representation. Using this representation in neural networks helps the model learn the phase and amplitude characteristics of the signal, better capturing signal properties that traditional real-space models may not be able to capture. Representing the nth symbol as a feature matrix consisting of nM to n+M symbols and feeding it into the neural network, incorporating the symbol's context (or window information) into the neural network's input, allows for better understanding of the symbol's context and identification of relationships between symbols, thereby improving the neural network's ability to learn complex patterns. This is particularly true in communication systems, where signals are often affected by noise and interference. By leveraging the contextual information of multiple symbols, the neural network can better extract valid signals from noise, thereby enhancing system robustness.
[0057] Based on the convolutional recurrent neural network feature extraction module, it is used to capture key information in the training signal and perform dimensionality reduction and denoising operations to reduce the dimension of the data and remove unnecessary noise, thereby improving the computational efficiency and balanced performance of the model. Figure 2As shown in the figure, the convolutional recurrent neural network feature extraction module consists of a convolutional layer, a pooling layer, an LSTM layer, a batch normalization layer, and an activation function. The time series features of the distorted signal are extracted through a network consisting of three convolutional layers connected in series, an LSTM layer, a batch normalization layer, and an activation layer. The batch normalization layer effectively prevents overfitting and gradient vanishing, thereby improving network balancing performance. The LSTM layer uses its memory capacity to integrate features extracted from the first two convolutional layers and associate contextual information at different time steps. This integration enables the network to better understand long-term dependencies between features, improving its ability to recognize and process complex time series data.
[0058] The data augmentation training module sends a single sample as a batch to the neural network training, and sets a data loader for each sample independently, such as Figure 3 As shown in the figure, this method acts as data augmentation. By designing independent data loading and processing logic for each sample, the network can dynamically simulate a variety of channel conditions and signal characteristics during training. For example, this can include random noise, frequency offset, amplitude imbalance, or multipath fading. This ensures that each training input data is slightly different, thus preventing the model from overfitting to a fixed signal pattern. This dynamic augmentation not only expands the effective size of the dataset but also greatly increases the diversity and complexity of the input signals, enabling the model to learn to adapt to a wider range of practical communication scenarios. Furthermore, because each signal sample is trained in an independent environment, the model can focus more on learning the characteristics of the current sample without being distracted by the differences in other samples, further improving the model's generalization and robustness. Ultimately, this approach enables neural networks to perform equalization tasks more stably and efficiently when faced with unknown channel conditions.
[0059] The discriminant module uses the test set to adjust the learning weights and converts the output feature information into corresponding bit information to achieve the final balanced mapping function.
[0060] In summary, combined with the design of various parts of the data-enhanced convolutional recurrent network, it is possible to achieve effective correction and equalization of distorted signals in complex communication channel environments, thereby improving the channel equalization capabilities of deep neural networks.
[0061] The specific implementation steps are as follows:
[0062] Step 1: Use the antenna to send signal X and tag y to generate distorted signal samples;
[0063] Step 1.1: Generate B × Num_s bits, with each B bit group as a Gray coding mapping group; B represents the number of bits corresponding to the modulation symbol, and Num_s is the number of symbols in a single signal;
[0064] Step 1.2: Store the bits from group M to group Num_s-M as the label y.
[0065] Step 1.3: Convert each group of bits into a decimal index, perform mapping encoding, and generate Num_s modulation symbols.
[0066] Step 1.4: After the modulated symbols pass through the channel, add Gaussian white noise.
[0067] Step 1.5: Oversample a single symbol by a sampling factor of 4.
[0068] Step 1.6: Put the nth symbol It is encapsulated with its adjacent M symbols to construct Characteristic matrix of symbols :
[0069] ;
[0070] in, The dimension is 4×7, where 4 represents the number of sampling points of the symbol.
[0071] Step 1.7: Convert the feature matrix Convert it into an IQ timing signal, then perform a Cat operation to obtain X. After passing through the channel environment, use Dataset to concatenate it with the sample label y to form a distorted signal sample. The IQ timing-related features of the transmitted signal are expressed as:
[0072] ;
[0073] in, The dimensions are 2×7, is the in-phase component, are orthogonal components.
[0074] Step 1.8: Divide the generated distorted signal samples into a training set and a test set in a ratio of 7:3.
[0075] Step 2: Train the convolutional recurrent network based on the training samples.
[0076] Step 2.1: Load each sample in the training set into Num data loaders, where Num is the number of received signals.
[0077] Step 2.2: Connect the Num data loaders that have loaded the data.
[0078] Step 2.3: Initialize network parameters, including network weights With bias .
[0079] Step 2.4: Extract the training data from the concatenated data loader and feed it into the convolutional recurrent network for training using the gradient descent algorithm and backpropagation mechanism.
[0080] The kth feature map output after the sample passes through the first convolutional layer , whose dimension is 4×7, is expressed as:
[0081] ;
[0082] in, Expressed as an activation function, Z is the received sample after the channel output, represents the batch normalization operation, represents the kth convolution kernel of the first convolutional layer, It is the bias introduced into the convolution result.
[0083] The kth feature map output after the sample passes through the second convolutional layer , whose dimension is 1×64, is expressed as:
[0084] ;
[0085] in, is the kth convolution kernel in the second convolution layer, For the corresponding offset, the feature map output by the first convolutional layer is encapsulated as .
[0086] Step 2.5: Compare the predicted results with the true labels and calculate the loss function Loss:
[0087] ;
[0088] in, is the reconstructed signal after the neural network; y is the sample label of the received signal; H is the number of samples in the current batch.
[0089] Step 2.6: Repeat steps 2.3 to 2.5 to minimize the loss function Loss until the maximum number of iterations MAX is reached. Use the trained convolutional recurrent network model to infer the soft bit vector of the network output. Dataset.
[0090] Soft bit vector output by the network Expressed as:
[0091] ;
[0092] in, and They represent the kth convolution kernel and the corresponding bias in the third convolution layer, respectively, and encapsulate the feature map output by the second convolution layer as .
[0093] Step 3: Compare the soft bit output of the trained network with the decision threshold to obtain the equalized bit error rate (BER).
[0094] Step 3.1: Input the test set into the trained network model to obtain the output soft bit vector Dataset.
[0095] Step 3.2: Output soft bit vector Dataset and threshold Compare, the soft bit vector greater than the threshold value The judgment is 1, and the judgment is 0 if it is less than the threshold value; the judgment threshold value The value is 0.5.
[0096] Step 3.3: Repeat steps 3.1 to 3.2, and record the number of inconsistent data and labels, which is recorded as c.
[0097] Step 3.4: After all test samples are compared, the equalized bit error rate (BER) is calculated, which is expressed as:
[0098] ;
[0099] Where c is the number of elements whose judgment output is different from the sample, and m is the total number of elements in the test sample.
[0100] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the present invention.
Claims
1. A convolutional recurrent network channel equalization system based on data enhancement, characterized by: The system includes a sample generation module, a feature extraction module based on a convolutional recurrent neural network, a data enhancement training module, and a discrimination module; The sample generation module is used to generate an IQ timing signal; generate a distorted signal sample using a transmitted signal X and its sample label y; X is a segment of original transmitted data for network training, with a dimension of 1×64, expressed as: The dimension of the sample label y is 64×4. The nth symbol is represented as a feature matrix with nM to n+M symbols and its real and imaginary components are extracted, where M is the number of adjacent symbols. The generated samples are divided into a training set and a test set in a ratio of 7:
3. The convolutional recurrent neural network feature extraction module is used to capture key information in the training signal and perform dimensionality reduction and denoising operations at the same time; the convolutional recurrent neural network includes a convolution layer, a pooling layer, an LSTM layer, a BN layer, and an activation function; the time series related features of the distorted signal are extracted through a network consisting of three convolution layers in series, an LSTM layer, a BN layer, and an activation layer; The role of the BN layer is to prevent overfitting and gradient disappearance, thereby improving the network's balanced performance; the LSTM layer integrates the features extracted from the first two convolution layers through its memory capacity and associates the contextual information of different time steps; The data augmentation training module sends individual samples as a batch to the neural network training and sets a data loader for each sample independently; The discrimination module uses the test set to adjust the learning weights, and the convolutional recurrent neural network converts the output feature information into corresponding bit information, compares it with the decision threshold value, and obtains the equalized bit error rate (BER).
2. The data-enhanced convolutional recurrent network channel equalization system according to claim 1, characterized in that: The specific operation process of the sample generation module is as follows: Step 1.1: Generate B × Num_s bits, with each B bit group as a Gray coding mapping group; B represents the number of bits corresponding to the modulation symbol, and Num_s is the number of symbols in a single signal; Step 1.2: Store the bits from group M+1 to group Num_s-M as the label y. Step 1.3: Convert each group of bits into a decimal index, perform mapping encoding, and generate Num_s modulation symbols; Step 1.4: After the modulation symbols pass through the channel, add Gaussian white noise; Step 1.5: Oversample a single symbol with 4 sampling points. Step 1.6: Put the nth symbol It is encapsulated with its adjacent M symbols to construct Characteristic matrix of symbols : ； in, The dimension is 4×(2M+1); Step 1.7: Convert the feature matrix Convert it into an IQ timing signal, then perform a Cat operation to obtain X. After passing through the channel environment, use Dataset to concatenate it with the sample label y to form a distorted signal sample. The IQ timing-related features of the transmitted signal are expressed as: ； in, The dimension is 2×(2M+1), is the in-phase component, is the orthogonal component; Step 1.8: Divide the generated distorted signal samples into a training set and a test set in a ratio of 7:
3.
3. The data-enhanced convolutional recurrent network channel equalization system according to claim 2, characterized in that: The steps of network training in the data enhancement training module are as follows: Step 2.1: Load each sample in the training set into Num data loaders, where Num is the number of received signals; Step 2.2: Connect the Num data loaders that have loaded the data; Step 2.3: Initialize network parameters, including network weights With bias ; Step 2.4: Extract the training data from the concatenated data loader and feed it into the convolutional recurrent network for training using the gradient descent algorithm and backpropagation mechanism; The kth feature map output after the sample passes through the first convolutional layer , whose dimension is 4×7, is expressed as: ； in, Expressed as an activation function, Z is the received sample after the channel output, represents the batch normalization operation, represents the kth convolution kernel of the first convolutional layer, Is the bias introduced into the convolution result; The kth feature map output after the sample passes through the second convolutional layer , whose dimension is 1×64, is expressed as: ； in, is the kth convolution kernel in the second convolution layer, For the corresponding bias, the feature map output by the first convolutional layer is encapsulated as ; Step 2.5: Compare the predicted results with the true labels and calculate the loss function Loss: ； in, is the reconstructed signal after the neural network; y is the sample label of the received signal; H is the number of samples in the current batch; Step 2.6: Repeat steps 2.3-2.5 to minimize the loss function Loss until the maximum number of iterations MAX is reached. Use the trained convolutional recurrent network model to infer the soft bit vector of the network output. Dataset; Soft bit vector output by the network Expressed as: ； in, and They represent the kth convolution kernel and the corresponding bias in the third convolution layer, respectively, and encapsulate the feature map output by the second convolution layer as .
4. The data-enhanced convolutional recurrent network channel equalization system according to claim 3, characterized in that: The specific operation process of the discrimination module is as follows: Step 3.1: Input the test set into the trained network model to obtain the output soft bit vector Dataset; Step 3.2: Output soft bit vector Dataset and threshold Compare, the soft bit vector greater than the threshold value The judgment is 1, and the judgment is 0 if it is less than the threshold value; Step 3.3: Repeat steps 3.1-3.2, and record the number of inconsistent data and labels, which is recorded as c. Step 3.4: After all test samples are compared, the equalized bit error rate (BER) is calculated, which is expressed as: ； Where c is the number of elements whose judgment output is different from the sample, and m is the total number of elements in the test sample.