Training method of OFDM system receiver model

Abstract

The invention relates to the technical field of wireless communication, in particular to a training method of an OFDM system receiver model, which comprises the following steps: acquiring a parameter-initialized signal detection model and a channel equalization model, and acquiring training set data and test set data; the signal detection model and the channel equalization model are both constructed based on a complex valued neural network; the signal detection model is trained based on the training set data until a loss function of the signal detection model converges; cascading the trained signal detection model to a channel equalization model to obtain an OFDM system receiver model; the OFDM system receiver model is trained based on the training set data, parameters of the channel equalization model are updated in the training process until the loss function of the OFDM system receiver model is converged, and training of the OFDM system receiver model is completed. The model trained by the method of the specification combines signal detection and signal estimation equalization, the data recovery capability is stronger, and the prediction result is more accurate.

Description

Technical Field

[0001] The embodiments of the present specification relate to the field of wireless communication technology, and in particular to a training method for an OFDM system receiver model. Background Art

[0002] Wireless communication is a communication method that uses the characteristics of electromagnetic wave signals propagating in free space to exchange information. In wireless communication systems, transmitters and receivers are crucial components. The transmitter converts the baseband signal into a radio frequency signal suitable for transmission in the wireless channel. The conversion process specifically involves signal modulation, up-conversion, power amplification and filtering. The receiver receives the radio frequency signal in the wireless channel and restores the signal sent by the transmitter through signal reception, demodulation, down-conversion, filtering, etc. In actual wireless communication scenarios, the performance of wireless communication systems may be affected by a variety of non-ideal factors such as multipath channel fading, Doppler frequency shift, radio frequency damage and noise. Therefore, the receiver needs to perform complex signal processing and analysis to remove interference factors as much as possible and restore the signal data sent by the transmitter. Traditional physical layer receivers usually use serial signal processing methods. The residual in the signal processing module may affect the optimization of subsequent modules, resulting in error accumulation and the inability to guarantee the performance of the entire system. In addition, the traditional channel model based on strict mathematical modeling cannot adaptively learn the changes in the channel to reflect the actual communication environment, and the computational complexity and amount are large. Summary of the invention

[0003] The embodiments of the present specification provide a method for training an OFDM system receiver model, aiming to solve one or more of the above problems and other potential problems.

[0004] In order to achieve the above objectives, the following technical solutions are provided: This specification provides a training method for an OFDM system receiver model, the method comprising: Obtain parameter-initialized signal detection model and channel equalization model, and obtain training set data and test set data; wherein the training set data and the test set data include a transmission signal and a reception signal of an OFDM communication system; the signal detection model is used to restore a distorted signal with noise, the channel equalization model is used for channel estimation and equalization, and both the signal detection model and the channel equalization model are constructed based on a complex-valued neural network; The signal detection model is trained based on the training set data until the loss function of the signal detection model converges and the training of the signal detection model is completed; the trained signal detection model is cascaded to the channel equalization model to obtain an OFDM system receiver model, and the input of the signal detection model is the output of the channel equalization model; The OFDM system receiver model is trained based on the training set data, and the parameters of the channel equalization model are updated during the training process until the loss function of the OFDM system receiver model converges, and the training of the OFDM system receiver model is completed; The input of the OFDM system receiver model is a received signal of the OFDM communication system, and the output is a predicted transmitted signal of the OFDM system.

[0005] The training method of the OFDM system receiver model in the embodiment of this specification, the channel equalization model is used to exploit the time-frequency characteristics of the channel to achieve accurate channel estimation and equalization, and reduce the impact of the fading channel; the signal detection model is used to restore the noisy distorted signal output by the channel equalization model to improve the adaptability of the entire model. The signal detection model is trained separately and then combined with the channel detection model for training to ensure the stability and reliability of each function in the signal processing process. At the same time, the two models are constructed based on complex-valued neural networks. Compared with real-valued neural networks, they can directly process complex-valued data, have a small network scale, strong computing power and high stability, and can learn channel changes more accurately to ensure the accuracy of signal recovery.

[0006] In some embodiments, the loss function of the signal detection model and the loss function of the OFDM system receiver model are both a weighted sum of a binary cross entropy loss and a regularization loss.

[0007] In some embodiments, the Adam optimization algorithm is used to update model parameters during training of the signal detection model and the OFDM system receiver model.

[0008] In some embodiments, the signal detection model includes a preprocessing module, a feature extraction module and a classification module; the preprocessing module converts the output of the channel equalization model into a three-dimensional complex-valued tensor, and inputs the three-dimensional complex-valued tensor into the feature extraction module; the feature extraction module includes a one-dimensional complex-valued convolutional layer, a first activation function layer and a complex-valued fully connected layer to extract sample features from the three-dimensional complex-valued tensor; the classification module converts the extracted sample features into a bit stream signal as a predicted transmission signal of the OFDM system.

[0009] In some embodiments, the feature extraction module also includes a two-dimensional convolution layer and a second activation function layer, which extracts sample features from the three-dimensional complex-valued tensor, including: inputting the three-dimensional complex-valued tensor into the one-dimensional complex-valued convolution layer, the complex-valued fully connected layer and the first activation function layer in sequence, and inputting the output of the first activation function layer into the complex-valued fully connected layer and the first activation function layer in sequence again to extract complex-valued local features; inputting the complex-valued local features into the two-dimensional convolution layer and the second activation function layer to obtain real-valued local features; performing deep feature fusion on the complex-valued local features and the real-valued local features to obtain a feature map, and performing feature extraction on the feature map to obtain the sample features.

[0010] In some embodiments, the first activation function layer and the second activation function layer use a Leaky ReLU function as an activation function.

[0011] In some embodiments, the channel equalization model includes a time-frequency transformation module, a channel equalization module and a signal recovery module; the time-frequency transformation module is used to convert the received signal of the OFDM communication system into a frequency domain signal; the channel equalization module includes several layers of complex-valued fully connected layers, and the frequency domain signal tensor is reduced in dimension and input into several layers of complex-valued fully connected layers, and the pilot signal channel estimation, channel interpolation and channel estimation of the entire coherent time slot are performed on the frequency signal in turn, and then the estimated channel tensor is stretched in four dimensions and input into a two-dimensional complex-valued convolution layer for processing; the signal recovery module is used to perform signal recovery based on the output signal of the signal equalization module.

[0012] In some embodiments, the time-frequency transformation module includes a Flatten layer, a complex-valued fully connected layer, and a one-dimensional complex-valued convolutional layer. The received signal of the OFDM system is reduced in dimension through the Flatten layer and then input into the complex-valued fully connected layer. The output of the complex-valued fully connected layer is stretched in four dimensions and then input into the one-dimensional complex-valued convolutional layer to output the frequency domain signal.

[0013] In some embodiments, the signal recovery module includes a residual network, two one-dimensional complex-valued convolutional layers and a complex-valued fully connected layer. The residual module is used to obtain the output signal of the signal equalization module, and the output signal of the residual network and its autocorrelation function are respectively input into the two one-dimensional complex-valued convolutional layers, and the outputs of the two one-dimensional complex-valued convolutional layers are deeply fused to obtain a feature map; the feature map tensor is input into the complex-valued fully connected layer after dimensionality reduction, and the output of the complex-valued fully connected layer is used as the output of the channel equalization model.

[0014] In some embodiments, the test set data is used to perform a BER test on the trained OFDM system receiver model, and when the BER is higher than a BER threshold, the signal detection model and the OFDM system receiver model are retrained until the BER is lower than the BER threshold or the number of training iterations reaches a threshold. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] By reading the following detailed description with reference to the accompanying drawings, the above and other objects, features and advantages of the embodiments of the present specification will become readily understood. In the accompanying drawings, several embodiments of the present specification are shown by way of example and not limitation.

[0016] Figure 1 A flow chart of a training method for an OFDM system receiver model in an embodiment of this specification is shown; Figure 2 A BER baseline comparison diagram of an OFDM system receiver model according to an embodiment of this specification in a FLAT Rayleigh fading channel scenario with different algorithms is shown; Figure 3 A BER baseline comparison diagram of an OFDM system receiver model according to an embodiment of this specification in an EPA Rayleigh fading channel scenario with different algorithms is shown; Figure 4 A BER baseline comparison diagram of an OFDM system receiver model in an embodiment of this specification in an EVA Rayleigh fading channel scenario with different algorithms is shown; Figure 5 A BER baseline comparison diagram of an OFDM system receiver model according to an embodiment of this specification in an ETU Rayleigh fading channel scenario with different algorithms is shown. DETAILED DESCRIPTION

[0017] The preferred embodiments of the present specification will be described in more detail below with reference to the accompanying drawings. Although the preferred embodiments of the present specification are shown in the accompanying drawings, it should be understood that the present specification can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to make the present specification more thorough and complete, and to fully convey the scope of the present specification to those skilled in the art.

[0018] The term "including" and its variations used in this document represent open inclusion, that is, "including but not limited to". Unless otherwise stated, the term "or" means "and / or". The term "based on" means "based at least in part on". The terms "an example embodiment" and "an embodiment" mean "at least one example embodiment". The term "another embodiment" means "at least one other embodiment". Terms such as "upper", "lower", "front", and "rear" indicating placement or positional relationships are based on the orientation or positional relationships shown in the drawings, and are only for the convenience of describing the principles of this specification, and do not indicate or imply that the referred elements must have a specific orientation, be constructed or operate in a specific orientation, and therefore should not be understood as limitations on this specification.

[0019] The following is a detailed description of a training method for an OFDM system receiver model according to an embodiment of the present specification in conjunction with the accompanying drawings. The training method for an OFDM system receiver model according to the present specification includes: S1. Obtain parameter-initialized signal detection model and channel equalization model, and obtain training set data and test set data; wherein the training set data and the test set data include transmission signals and reception signals of an OFDM communication system; the signal detection model is used to restore distorted signals with noise, and the channel equalization model is used for channel estimation and equalization, and both the signal detection model and the channel equalization model are constructed based on a complex-valued neural network; S3, the signal detection model is trained based on the training set data until the loss function of the signal detection model converges, and the training of the signal detection model is completed; S5. cascading the trained signal detection model to the channel equalization model to obtain an OFDM system receiver model; wherein the input of the signal detection model is the output of the channel equalization model, the input of the OFDM system receiver model is the received signal of the OFDM communication system, and the output is the predicted transmitted signal of the OFDM system; S7. The OFDM system receiver model is trained based on the training set data. During the training process, the parameters of the channel equalization model are updated until the loss function of the OFDM system receiver model converges and the training of the OFDM system receiver model is completed.

[0020] OFDM is a multi-frequency carrier modulation technology. When transmitting signals, a high-speed data stream is decomposed into multiple parallel low-speed data streams, which are transmitted simultaneously on multiple subcarriers. The signal on each subcarrier is modulated using a modulation method and then sent simultaneously. Since the subcarriers are mutually orthogonal, they can overlap in the spectrum, thereby improving spectrum utilization. In order to adaptively learn channel changes in an OFDM system, in step S1, the training set data and the test set data used by the receiver model of the embodiment of this specification include a transmission signal and a reception signal of an OFDM communication system. The transmission signal of the OFDM communication system is a signal constructed from an information bit stream, and the reception signal of the OFDM communication system is also in the form of an information bit stream. Specifically, the traditional OFDM communication system model is first obtained through Matlab. The sampling rate of the system model is set to 0.96MHz, and the subcarrier bandwidth is 15KHz; the number of OFDM symbols contained in a coherent time slot is 7, the number of subcarriers is 64, the number of DC protection subcarriers is 2, the total number of edge protection subcarriers is 14, and the two edges have 7 edge protection subcarriers. The number of time-frequency resource blocks allocated to pilots and data in a coherent time slot is 16 and 320 respectively. The pilot insertion method uses a constant signal discrete pilot, and the value is set to. The transmission signal of the OFDM communication system is randomly generated, transmitted into the channel through the OFDM system transmitter, and then the received signal of the OFDM communication system is captured at the receiving end of the channel. The transmission signal of the OFDM communication system is modulated by BPSK, QPSK, 8QAM, 16QAM and other modulation methods, so that the trained model can adapt to the channel estimation and denoising process of signals modulated by various modulation methods. The maximum amplitude of the constellation diagram is 1, and the signal peak-to-average power ratio is limited to within 9dB. In addition, different 3GPP multipath fading channel simulation models are used to simulate the multipath fading of the received signal of the OFDM communication system. The multipath fading channel simulation models are flat fading channel, extended pedestrian channel, extended vehicle channel, extended typical urban channel, etc. The maximum Doppler frequency shifts corresponding to the four simulation models are 0Hz, 5Hz, 70Hz, and 300Hz, respectively. When the DFT length is 64, the channel filter lengths of the four simulation models are 1, 9, 11, and 13, respectively.

[0021] The channel equalization model described in step S1 is used for channel estimation and equalization, specifically to explore the characteristics of impulse response, noise and interference in the channel, restore the received signal of the OFDM system based on the channel state characteristics, and reduce the impact of the fading channel; the signal detection model is used to further restore the noisy distorted signal output by the channel equalization model, improve the signal recovery effect and adaptability of the entire receiver model; the channel equalization model is trained separately to ensure the stability and reliability of each function in the signal process. At the same time, both models are constructed based on complex-valued neural networks. Complex-valued neural networks extend the neural network in the real number domain to the complex number domain, introduce new operations such as complex addition and complex multiplication, and the neural network has stronger expression ability and more flexible models; compared with the model constructed by real-valued neural networks, the model in the embodiment of this specification can directly process complex-valued data, with a small network scale, strong computing power and higher stability.

[0022] Specifically, the signal detection model includes a preprocessing module, a feature extraction module and a classification module. The preprocessing module converts the output deformation of the channel equalization model into a frame structure format, and then decomposes it into a three-dimensional complex valued tensor so that the feature extraction module can extract features from the complex input signal.

[0023] The feature extraction module includes a one-dimensional complex-valued convolution layer, a first activation function layer and a complex-valued fully connected layer, and also includes a two-dimensional convolution layer and a second activation function layer. The three-dimensional complex-valued tensor output by the preprocessing module is stretched in four dimensions and then input into the one-dimensional complex-valued convolution layer to achieve functions similar to CP removal and discrete Fourier transform, and the implicit information in the CP is used to obtain performance gain. At the same time, the time domain symbols are converted into frequency domain symbols and input into the first activation function layer. The first activation function layer is activated and then input into the complex-valued fully connected layer through the Flatten operation. The output of the complex-valued fully connected layer is input into the second activation function layer, and the output of the second activation function layer is input into the complex-valued fully connected layer and the first activation function layer in turn to extract complex-valued local features; the complex-valued local features are input into the two-dimensional convolution layer and the second activation function layer to obtain real-valued local features; the complex-valued local features and the real-valued local features are deeply fused to obtain feature maps, and the feature maps are input into a fully connected layer again for feature extraction to obtain sample features. Wherein, the first activation function layer and the second activation function layer both use the Leaky ReLU function as the activation function.

[0024] Finally, the classification module converts the extracted sample features into a bit stream signal as the predicted transmission signal of the OFDM system. Specifically, the sample features are output along the last dimension through the softmax layer to restore the soft information bit stream, and a simple judgment is made on the output soft bit stream to obtain the predicted bit stream.

[0025] The channel equalization model includes a time-frequency conversion module, a channel equalization module and a signal recovery module. The time-frequency conversion module is used to convert the received signal of the OFDM communication system into a frequency domain signal. Specifically, the time-frequency conversion module includes a Flatten layer, a complex-valued fully connected layer, and a one-dimensional complex-valued convolutional layer. The received signal of the OFDM system is input into the complex-valued fully connected layer after dimension reduction through the Flatten layer. The output of the complex-valued fully connected layer is four-dimensionally stretched and then input into the one-dimensional complex-valued convolutional layer to output the frequency domain signal. First, the received signal of the OFDM system is converted into an OFDM frame structure format, and decomposed into real and imaginary parts in the last dimension, and then normalized layer by layer and tensor dimension reduction is performed through the Flatten layer, and then input into the complex-valued fully connected layer to extract a CP-removed OFDM frame. After four-dimensional stretching, it is sent to the one-dimensional complex-valued convolutional layer to convert the time domain symbols into frequency domain symbols.

[0026] The channel equalization module includes several layers of complex-valued fully connected layers, which reduce the dimension of the frequency domain signal tensor and input it into several layers of complex-valued fully connected layers, perform pilot signal channel estimation, channel interpolation and channel estimation of the entire coherent time slot on the frequency signal in turn, and then perform four-dimensional stretching on the estimated channel tensor and input it into a two-dimensional complex-valued convolutional layer for processing.

[0027] The signal recovery module is used to perform signal recovery based on the output signal of the signal equalization module. Specifically, it includes a residual network, two one-dimensional complex-valued convolutional layers and a complex-valued fully connected layer. The residual module is used to obtain the output signal of the signal equalization module, and the output signal of the residual network and its autocorrelation function are respectively input into the two one-dimensional complex-valued convolutional layers, and the outputs of the two one-dimensional complex-valued convolutional layers are deeply fused to obtain a feature map; the feature map tensor is reduced in dimension and input into the complex-valued fully connected layer to recover the transmission data with CP in a coherent time slot, and the recovered data is processed and converted into the same form as the input signal as the output of the channel equalization model.

[0028] Specifically, the network hyperparameters of the signal detection model and the channel equalization model are set as follows: in the signal detection model, the number of convolution kernels of the one-dimensional complex-valued convolution layer is 64; the number of neurons of the complex-valued fully connected layer is 320, the number of convolution kernels of the two-dimensional convolution layer is 2, the convolution kernel size is 1, and the number of neurons of the fully connected layer for extracting sample features is M. In the channel equalization model, the number of convolution kernels of the one-dimensional complex-valued convolution layer of the time-frequency transform module is 64; the number of neurons of the first complex-valued fully connected layer in the channel equalization module is P, the number of neurons of the second complex-valued fully connected layer is 4, and the number of neurons of the third and fourth complex-valued fully connected layers is 448; the convolution kernel in the two-dimensional complex-valued convolution layer in the channel equalization module is 1, and the convolution kernel size is (7, 64). The number of convolution kernels of the one-dimensional complex-valued convolution layer in the signal recovery module is 64, and the convolution kernel size is (1, 64). The training data is input into the model in small batches, each small batch of data contains 72 frames of coherent time slots, and the number of iterations is used to describe the outermost loop in the training. One training iteration contains 200 small batches of data. The optimization process is performed by the Adam optimizer, and the initial learning rate is set to 0.001. The learning rate is exponentially decayed for fine-tuning after every 500 small batches.

[0029] In step S3 and step S7, the loss function of the signal detection model and the loss function of the OFDM system receiver model are both the weighted sum of the binary cross entropy loss and the regularization loss. The Adam optimization algorithm is used to update the model parameters when training the signal detection model and the OFDM system receiver model. When the Adam optimization algorithm updates the model parameters, it can adaptively adjust the learning rate according to the gradient characteristics of different parameters, combining the advantages of momentum and adaptive learning rate, accelerating the convergence process and improving training efficiency. During training, a hybrid Rayleigh fading model is used for training, and the Rayleigh fading is set to FLAT, EPA, EVA and ETU for alternating training to improve the generalization ability of the model.

[0030] After step S7, the test set data is used to perform a BER test on the trained OFDM system receiver model. When the BER is higher than the BER threshold, the signal detection model and the OFDM system receiver model are retrained until the BER is lower than the BER threshold or the number of training iterations reaches a threshold.

[0031] like Figure 2 , Figure 3 , Figure 4 , Figure 5The figure shows the comparison of the OFDM system receiver model CVCE-CVSD of the embodiment of this specification with the BER baselines of different algorithms in the FLAT, EPA, EVA, and ETU Rayleigh fading channel scenarios. The horizontal axis in the figure is the signal-to-noise ratio of the training set data and the test set data. The baselines include the linear minimum mean square error baseline ILMMSE; the traditional receiver baseline LS-Spline using the LS estimation algorithm, where the interpolation method of the traditional receiver baseline is the spline interpolation method; the traditional receiver baseline ALMMSE using the approximate linear minimum mean square error estimation algorithm; the signal recovery baseline Perfect with completely known channel information; and the deep learning receiver model baseline DCCN in the literature [Wen Kechao, Shi Tianwei, Jin Sheng. Application of deep learning in large-scale MIMO CSI feedback [J]. IEEE Wireless Communications Express, 2018, 7(5): 748-751].

[0032] Although several specific implementation details are included in the above discussion, these should not be interpreted as limiting the scope of this specification. Certain features described in the context of separate embodiments may also be implemented in a single implementation in combination. Conversely, various features described in the context of a single implementation may also be implemented in multiple implementations individually or in any suitable sub-combination.

[0033] Although the subject matter has been described in language specific to structural features and / or methodological logical actions, it should be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or actions described above. On the contrary, the specific features and actions described above are merely example forms of implementing the claims.

[0034] The embodiments of the present specification have been described above, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and changes will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The selection of terms used herein is intended to best explain the principles of the embodiments, practical applications, or technical improvements in the market, or to enable other persons of ordinary skill in the art to understand the embodiments disclosed herein.

Claims

1. A training method for an OFDM system receiver model, characterized in that: Methods include: Obtain parameter-initialized signal detection model and channel equalization model, and obtain training set data and test set data; wherein the training set data and the test set data include a transmission signal and a reception signal of an OFDM communication system; the signal detection model is used to restore a distorted signal with noise, the channel equalization model is used for channel estimation and equalization, and both the signal detection model and the channel equalization model are constructed based on a complex-valued neural network; The signal detection model is trained based on the training set data until the loss function of the signal detection model converges and the training of the signal detection model is completed; the trained signal detection model is cascaded to the channel equalization model to obtain an OFDM system receiver model, and the input of the signal detection model is the output of the channel equalization model; The OFDM system receiver model is trained based on the training set data, and the parameters of the channel equalization model are updated during the training process until the loss function of the OFDM system receiver model converges, and the training of the OFDM system receiver model is completed; The input of the OFDM system receiver model is a received signal of the OFDM communication system, and the output is a predicted transmitted signal of the OFDM system.

2. The method according to claim 1, characterized in that The loss function of the signal detection model and the loss function of the OFDM system receiver model are both weighted sums of binary cross entropy loss and regularization loss.

3. The method according to claim 2, characterized in that The Adam optimization algorithm is used to update model parameters during the training of the signal detection model and the OFDM system receiver model.

4. The method according to claim 3, characterized in that The signal detection model includes a preprocessing module, a feature extraction module and a classification module; the preprocessing module converts the output of the channel equalization model into a three-dimensional complex valued tensor, and inputs the three-dimensional complex valued tensor into the feature extraction module; The feature extraction module includes a one-dimensional complex-valued convolution layer, a first activation function layer and a complex-valued fully connected layer, which extracts sample features from the three-dimensional complex-valued tensor; the classification module converts the extracted sample features into a bit stream signal as a predicted transmission signal of the OFDM system.

5. The method according to claim 4, characterized in that The feature extraction module also includes a two-dimensional convolution layer and a second activation function layer, which extracts sample features from the three-dimensional complex-valued tensor, including: inputting the three-dimensional complex-valued tensor into the one-dimensional complex-valued convolution layer, the complex-valued fully connected layer and the first activation function layer in sequence, and inputting the output of the first activation function layer into the complex-valued fully connected layer and the first activation function layer in sequence again to extract complex-valued local features; inputting the complex-valued local features into the two-dimensional convolution layer and the second activation function layer to obtain real-valued local features; performing deep feature fusion on the complex-valued local features and the real-valued local features to obtain a feature map, and performing feature extraction on the feature map to obtain the sample features.

6. The method according to claim 5, characterized in that The first activation function layer and the second activation function layer use the Leaky ReLU function as an activation function.

7. The method according to claim 3, characterized in that The channel equalization model includes a time-frequency conversion module, a channel equalization module and a signal recovery module; the time-frequency conversion module is used to convert the received signal of the OFDM communication system into a frequency domain signal; the channel equalization module includes several layers of complex-valued fully connected layers, the frequency domain signal tensor is reduced in dimension and input into several layers of complex-valued fully connected layers, the pilot signal channel estimation, channel interpolation and channel estimation of the entire coherent time slot are performed on the frequency signal in sequence, and then the estimated channel tensor is stretched in four dimensions and input into a two-dimensional complex-valued convolution layer for processing; The signal recovery module is used to perform signal recovery based on the output signal of the signal equalization module.

8. The method according to claim 7, characterized in that The time-frequency transformation module includes a Flatten layer, a complex-valued fully connected layer, and a one-dimensional complex-valued convolutional layer. The received signal of the OFDM system is reduced in dimension by the Flatten layer and then input into the complex-valued fully connected layer. The output of the complex-valued fully connected layer is stretched in four dimensions and then input into the one-dimensional complex-valued convolutional layer to output the frequency domain signal.

9. The method according to claim 7, characterized in that: The signal recovery module includes a residual network, two one-dimensional complex-valued convolutional layers and a complex-valued fully connected layer. The residual module is used to obtain the output signal of the signal equalization module, and the output signal of the residual network and its autocorrelation function are respectively input into the two one-dimensional complex-valued convolutional layers, and the outputs of the two one-dimensional complex-valued convolutional layers are deeply fused to obtain a feature map; the feature map tensor is input into the complex-valued fully connected layer after dimensionality reduction, and the output of the complex-valued fully connected layer is used as the output of the channel equalization model.

10. The method according to claim 1, characterized in that Also includes: The trained OFDM system receiver model is subjected to a BER test using the test set data. When the BER is higher than a BER threshold, the signal detection model and the OFDM system receiver model are retrained until the BER is lower than the BER threshold or the number of training iterations reaches a threshold.

Images