Physical Layer Structure of Wireless Communication Based on Deep Learning

Abstract

The present invention provides a physical layer structure for wireless communication based on deep learning, including: an encoder part, a channel simulation part, a decoder part, a random interference generator part, an interference feature extraction part, and a subcarrier constraint part. The present invention realizes reliable communication under the condition of severe cross-protocol interference. The encoder end will automatically adjust the modulation mode of the signal according to the interference in the current environment, so that it has the ability to resist interference. At the receiving end, the decoder of the autoencoder is used to restore the transmitted signal. The interference feature extractor and subcarrier constraint loaded at the encoder and decoder ends will be used to guide the encoder to perform adaptive modulation mode control, so as to adopt different modulation modes for different types of interference and achieve the effect of environmental adaptability. This design has strong universality and can find universal and effective anti-interference communication methods, thus guiding the communication of a large number of Internet of Things devices and providing support for the development of the Internet of Things field.

Description

Technical Field

[0001] The present invention relates to the technical field of communication technologies, and specifically, to a physical layer structure of wireless communication based on deep learning. Background Art

[0002] In the era of the Internet of Things, the objects of communication have extended from people to objects, which means that the diversity of network devices has increased greatly. It can be a smart phone or a bus card. At the same time, with the diversification of application scenarios, the network environment has become highly dynamic and complex, and communication modes have become increasingly diverse.

[0003] In the patent document with the publication number CN113746628A, a physical layer key generation method and system based on deep learning are disclosed. By collecting the estimated value pairs of the channels of both legitimate communication parties within the coherence time, fusing the estimated value pairs respectively obtained by both legitimate communication parties to obtain a pair of training data, obtaining several pairs of training data within several coherence times, establishing a key generation network at both communication parties, and the key generation network includes a feature extraction network and a decoding network; using the training data to train the key generation network to achieve network deep learning training, and completing the training of the key generation network by sharing the Pearson correlation coefficient and mean of each dimension of the consistent feature vectors output by the feature extraction networks of both communication parties. Using the trained key generation network to generate feature vectors according to the communication value, and quantifying the generated feature vectors using a key quantization algorithm to obtain the key sequences of both communication parties.

[0004] Existing communication technologies usually rely on relatively idealized and rigid network models, and it is difficult to achieve reliable transmission in highly dynamic and complex scenarios. Taking the physical layer of communication as an example, to achieve wireless signal transmission, the primary task is to model the channel. As the network environment becomes more and more complex, the mathematical models used to describe the channel become more and more complex. From the initial free space transmission model to the final spatio-temporal channel model, the channel model needs to consider path fading, shadow fading, multipath effect, Doppler phenomenon, and the influence of multiple antennas on the channel at the same time. Coupled with the mutual interference among a large number of devices today, this makes the channel more difficult to estimate, and people can no longer find an accurate mathematical model to describe and characterize such a highly dynamic and complex channel. With the advent of the "AI +" era, this patent provides a method of using a deep learning model (mainly an autoencoder) to replace the traditional mathematical model for channel modeling, enabling the channel model to adapt to the changing channel environment, thereby guiding the Internet of Things devices to perform highly reliable communication in the presence of interference.

[0005] Therefore, it is necessary to propose a new technical solution to improve the above technical problems. Summary of the Invention

[0006] Aiming at the defects in the prior art, the purpose of the present invention is to provide a physical layer structure for wireless communication based on deep learning.

[0007] A physical layer structure for wireless communication based on deep learning provided by the present invention includes an encoder, a channel simulator, a decoder, a random interference generator, an interference feature extractor, and a subcarrier constraint.

[0008] The encoder encodes and power-constrains the signal according to the transmitted signal and the input interference features.

[0009] The channel simulator superimposes the effects existing in the simulated real channel on the outputs of the encoder and the random interference generator during training.

[0010] The decoder decodes according to the output of the encoder after channel simulation and the input interference features.

[0011] The interference feature extractor extracts the features of the interference according to the output of the random interference generator after channel simulation.

[0012] The random interference generator simulates various interferences that may occur in the environment as an interference source.

[0013] The subcarrier constraint constrains the subcarriers output by the encoder according to the input interference information.

[0014] Preferably, the encoding of the encoder is based on the one-dimensional convolutional layer and the fully connected layer of the neural network, performs dimension transformation on the input signal, and the output dimension is equal to the number of subcarriers, and modulates the signal on each subcarrier.

[0015] The power constraint normalizes the encoded output and constrains the transmission power.

[0016] Preferably, the normalization refers to dividing the signal amplitude on each subcarrier by the maximum value of the signal amplitude on that subcarrier.

[0017] Preferably, the effects in the channel simulator include carrier frequency offset CFO, sampling frequency offset SFO, Gaussian noise, frequency conversion, and pulse shaping.

[0018] Preferably, the decoding in the decoder is based on the one-dimensional convolutional layer, the fully connected layer, and the regularization layer of the neural network, performs dimension transformation on the input, and outputs the transmitted signal.

[0019] Preferably, the feature extraction of the interference feature extractor includes the interference feature extraction before the encoder and the feature extraction before the decoder.

[0020] The feature extraction is based on the one-dimensional convolutional layer, pooling layer, and fully connected layer of the neural network, which perform dimensional transformation on the input interference to find the low-dimensional representation of the interference signal.

[0021] Preferably, the interference feature extractor before the encoder is trained with the system bit error rate as the constraint, and the interference feature extractor before the decoder is trained with the system bit error rate and the output of the interference feature extractor before the encoder as the constraints;

[0022] The inputs of the interference feature extractor before the encoder and the interference feature extractor before the decoder are staggered in time to extract features that are independent of time for the interference.

[0023] Preferably, the output of the interference feature extractor is connected to the encoder and the decoder to guide the encoder and the decoder and synchronize the information of the encoder and the decoder.

[0024] Preferably, the subcarrier constraint is based on the one-dimensional convolutional layer and fully connected layer of the neural network, inputs the output of the random interference generator after channel simulation, outputs the subcarrier frequency band selection vector of the encoder, and multiplies it with the output of the encoder.

[0025] Preferably, the dimension of the subcarrier frequency band selection vector is the same as the output of the encoder, and it is a vector containing only 0 / 1.

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] Through such a design, the present invention realizes reliable communication under the condition of severe cross-protocol interference; the present invention has strong universality and can find a universal and effective anti-interference communication method, thereby guiding the communication of a large number of Internet of Things devices and providing support for the development of the Internet of Things field. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] By reading the following detailed description of the non-limiting embodiments with reference to the accompanying drawings, other features, objects, and advantages of the present invention will become more apparent:

[0029] Figure 1 It is the working flowchart of the physical layer structure of the wireless communication based on the deep autoencoder in the present invention;

[0030] Figure 2 It is the time-frequency spectrum diagram of the interference generated by the random interference generator in the present invention;

[0031] Figure 3 It is the simulation training effect diagram;

[0032] Figure 4 It is the input interference constellation diagram;

[0033] Figure 5 It is the constellation diagram at the output end of the encoder. Detailed implementation manners

[0034] The present invention will be described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but do not limit the present invention in any form. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several changes and improvements can still be made. These all belong to the protection scope of the present invention.

[0035] Example 1:

[0036] A physical layer structure of wireless communication based on deep learning provided by the present invention includes an encoder, a channel simulator, a decoder, a random interference generator, an interference feature extractor, and a subcarrier constraint device; the encoder encodes and power-constrains the signal according to the transmitted signal and the input interference features; the channel simulator superimposes and simulates the effects existing in the real channel on the outputs of the encoder and the random interference generator during training; the decoder decodes according to the output of the encoder after channel simulation and the input interference features; the interference feature extractor extracts features of the interference according to the output of the random interference generator after channel simulation; the random interference generator simulates various interferences that may occur in the environment as an interference source; the subcarrier constraint device constrains the subcarriers output by the encoder according to the input interference information.

[0037] The encoding of the encoder is based on the one-dimensional convolutional layer and the fully connected layer of the neural network, performs dimensional transformation on the input signal, and the output dimension is equal to the number of subcarriers, and modulates the signal on each subcarrier; the power constraint normalizes the encoded output and constrains the transmission power.

[0038] Normalization means dividing the signal amplitude on each subcarrier by the maximum value of the signal amplitude on that subcarrier.

[0039] The effects in the channel simulator include carrier frequency offset CFO, sampling frequency offset SFO, Gaussian noise, frequency conversion, and pulse shaping.

[0040] The decoding in the decoder is based on the one-dimensional convolutional layer, the fully connected layer, and the regularization layer of the neural network, performs dimensional transformation on the input, and outputs the transmitted signal.

[0041] The feature extraction of the interference feature extractor includes interference feature extraction before the encoder and feature extraction before the decoder; the feature extraction is based on the one-dimensional convolutional layer, the pooling layer, and the fully connected layer of the neural network, performs dimensional transformation on the input interference, and finds the low-dimensional representation of the interference signal.

[0042] The interference feature extractor before the encoder is trained with the system bit error rate as the constraint, and the interference feature extractor before the decoder is trained with the system bit error rate and the output of the interference feature extractor before the encoder as the constraints; the inputs of the interference feature extractor before the encoder and the interference feature extractor before the decoder are staggered in time to extract features that are independent of the interference and time.

[0043] The output of the interference feature extractor is connected to the encoder and the decoder to guide the encoder and the decoder and synchronize the information of the encoder and the decoder.

[0044] The subcarrier constraint is based on the one-dimensional convolutional layer and the fully connected layer of the neural network, inputs the output of the random interference generator after channel simulation, outputs the subcarrier frequency band selection vector of the encoder, and multiplies it with the output of the encoder.

[0045] The dimension of the subcarrier frequency band selection vector is the same as the output of the encoder, and it is a vector containing only 0 / 1.

[0046] Example 2:

[0047] A physical layer structure of wireless communication based on deep learning provided by the present invention includes:

[0048] An encoder part, a channel simulation part, a decoder part, a random interference generator part, an interference feature extraction part, and a subcarrier constraint part.

[0049] The encoder part: encodes and power-constrains the signal according to the transmitted signal and the input interference features;

[0050] Specifically, the encoding: based on the one-dimensional convolutional layer and the fully connected layer of the neural network, performs dimension transformation on the input signal, and the output dimension is equal to the number of subcarriers, which is equivalent to modulating the signal on each subcarrier.

[0051] The power constraint: normalizes the encoded output to constrain the transmission power.

[0052] Specifically, the normalization refers to dividing the signal amplitude on each subcarrier by the maximum value of the signal amplitude on that subcarrier.

[0053] The channel simulation part: in training, superimposes some effects existing in the real channel on the outputs of the encoder and the random interference generator, including: carrier frequency offset CFO, sampling frequency offset SFO, Gaussian noise, frequency conversion, and pulse shaping.

[0054] The decoder part: decodes according to the output of the encoder after channel simulation and the input interference features;

[0055] Specifically, the decoding: similar to the encoding structure, is based on the one-dimensional convolutional layer, fully connected layer, and regularization layer of the neural network to perform dimensional transformation on the input and output the transmitted signal. The role of the regularization layer is to prevent overfitting during the training process.

[0056] The interference feature extractor part: extracts features of the interference according to the output of the channel simulation by the random interference generator, which is divided into interference feature extraction before the encoder and feature extraction before the decoder.

[0057] Specifically, the feature extraction: based on the one-dimensional convolutional layer, pooling layer, and fully connected layer of the neural network, performs dimensional transformation on the input interference to find the low-dimensional representation of the interference signal.

[0058] Specifically, the interference feature extraction before the encoder and the interference feature extraction before the decoder have the same network structure, and the differences are as follows:

[0059] The interference feature extractor before the encoder is trained with the system bit error rate as the constraint, and the interference feature extractor before the decoder is trained with the system bit error rate and the output of the interference feature extractor before the encoder as the constraints;

[0060] The inputs of the interference feature extractor before the encoder and the interference feature extractor before the decoder are staggered in time, aiming to extract features independent of the interference and time.

[0061] Specifically, the output of the interference feature extractor is connected to the encoder and the decoder, used to guide the encoder and the decoder, and at the same time synchronize the information of the encoder and the decoder.

[0062] The random interference generator part: simulates various interferences that may occur in the environment as the interference source, including:

[0063] Wi-Fi 802.11a / g;

[0064] Wi-Fi 802.11b / g;

[0065] Wi-Fi 802.11p;

[0066] BLE

[0067] LTE downlink RMC

[0068] LTE uplink RMC

[0069] OFDM

[0070] The subcarrier constraint part: constrains the subcarriers output by the encoder according to the input interference information to avoid the subcarrier frequency bands where the interference is located.

[0071] Specifically, the subcarrier constraint: based on the one-dimensional convolutional layer and fully connected layer of the neural network, it inputs the output of the channel simulation by the random interference generator, outputs the subcarrier frequency band selection vector of the encoder, and multiplies it with the output of the encoder.

[0072] Specifically, the subcarrier frequency band selection vector: a vector with the same dimension as the output of the encoder and only containing 0 / 1.

[0073] Aiming at the defects in the prior art, the purpose of the present invention is to provide a physical layer structure for wireless communication based on deep learning.

[0074] According to a physical layer structure for wireless communication based on deep learning provided by the present invention, it includes the following parts:

[0075] An encoder part, a channel simulation part, a decoder part, a random interference generator part, an interference feature extraction part, and a subcarrier constraint part.

[0076] Part 1: Encoder:

[0077] The encoder is based on the one-dimensional convolutional layer and fully connected layer of the neural network, performs dimensional transformation and power constraint on the transmitted signal and the input interference feature, and outputs a dimension equal to the number of subcarriers, which is equivalent to modulating the signal with interference on each subcarrier. Finally, power constraint is performed: that is, normalizing the encoded output, that is, dividing the signal amplitude on each subcarrier by the maximum value of the signal amplitude on that subcarrier to constrain the transmission power.

[0078] Part 2: Channel simulation:

[0079] The channel simulation part simulates some effects that may affect the communication quality in the real channel during training, including carrier frequency offset CFO, sampling frequency offset SFO, Gaussian noise, frequency conversion, and pulse shaping.

[0080] The specific process is as follows:

[0081] For the output of the encoder (frequency domain), after superimposing the sampling frequency offset SFO, perform IFFT to convert it to the time domain, perform upsampling, pulse shaping, and upconversion, and then superimpose the carrier frequency offset CFO and a certain degree of Gaussian noise. This is used as Figure 1 the output of channel A therein.

[0082] For the 64-point output of the random interference generator (time domain, corresponding to a transmission time of about 5 μs), perform FFT to convert it to the frequency domain, then superimpose the sampling frequency offset SFO, and then perform IFFT to convert it to the time domain, and superimpose the carrier frequency offset CFO. This is used as Figure 1 the output of channel C therein.

[0083] Similarly, for the 500-point output of the random interference generator (in the time domain, corresponding to an observation time of approximately 25 μs), perform FFT to convert it to the frequency domain, then superimpose the sampling frequency offset SFO, and then perform IFFT to convert it back to the time domain, and superimpose the carrier frequency offset CFO. Use this as Figure 1 the output of channel B in

[0084] Superimpose the outputs of channel A and channel C and then perform down-conversion and down-sampling to serve as the first input to the decoder.

[0085] Part 3: Decoder:

[0086] Similar to the encoding structure, the decoder is based on the one-dimensional convolutional layer, fully connected layer, and regularization layer of the neural network to perform dimensional transformation on the input and output the transmitted signal. The role of the regularization layer is to prevent overfitting during the training process.

[0087] The output of the decoder is the finally demodulated information at the receiving end.

[0088] Part 4: Interference feature extractor:

[0089] Interference feature extractor part: Extract the features of the interference according to the output of the random interference generator after channel simulation, which is divided into interference feature extraction before the encoder and feature extraction before the decoder.

[0090] The feature extraction is based on the one-dimensional convolutional layer, pooling layer, and fully connected layer of the neural network to perform dimensional transformation on the input interference and find the low-dimensional representation of the interference signal.

[0091] The output dimension of the interference feature extractor should be less than or equal to the specific number of interference types in the environment, so as to guide the encoder and decoder to make the same modulation and demodulation responses to the same large category of interference.

[0092] At the same time, the interference feature extraction before the encoder and the interference feature extraction before the decoder have the same network structure, and their differences are as follows:

[0093] The interference feature extractor before the encoder is trained with the system bit error rate as the constraint, and the interference feature extractor before the decoder is trained with the system bit error rate and the output of the interference feature extractor before the encoder as the constraints;

[0094] The inputs of the interference feature extractor before the encoder and the interference feature extractor before the decoder are staggered in time, aiming to extract features independent of time for the interference.

[0095] Specifically, the input of the interference feature extractor is Figure 1 the output of channel B in , that is, the interference observation value considering the influence of the real channel. The output of the interference feature extractor is the second input to the encoder and decoder.

[0096] Part 5: Random interference generator:

[0097] The random interference generator part: Simulates various interferences that may occur in the interference source environment, including:

[0098]

[0099] Part 6: Sub - carrier constraint:

[0100] The sub - carrier constraint is to constrain the sub - carriers output by the encoder according to the input interference information to avoid the sub - carrier frequency bands where interference exists.

[0101] Specifically, based on the one - dimensional convolutional layer and fully - connected layer of the neural network, the output of the random interference generator after channel simulation is input, and a sub - carrier frequency band selection vector of the encoder is output. The dimension of this vector is the same as the output of the encoder and only contains 0 / 1, representing whether the sub - carrier is selected to transmit information, and it is multiplied by the output of the encoder.

[0102] Those skilled in the art can understand this embodiment as a more specific illustration of Embodiment 1.

[0103] Those skilled in the art know that in addition to implementing the system and its various devices, modules, and units provided by the present invention in the form of pure computer - readable program code, the method steps can be logically programmed to enable the system and its various devices, modules, and units provided by the present invention to be implemented in the form of logic gates, switches, application - specific integrated circuits, programmable logic controllers, and embedded microcontrollers, etc. to achieve the same functions. Therefore, the system and its various devices, modules, and units provided by the present invention can be considered as a kind of hardware component, and the devices, modules, and units included therein for implementing various functions can also be regarded as the structures within the hardware component; or the devices, modules, and units for implementing various functions can be regarded as both software modules for implementing the method and the structures within the hardware component.

[0104] The specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the above - mentioned specific embodiments. Those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essence of the present invention. Without conflict, the embodiments of the present application and the features in the embodiments can be combined arbitrarily.

Claims

1. A physical layer structure for wireless communication based on deep learning, characterized in that, It includes an encoder, a channel simulator, a decoder, a random interference generator, an interference feature extractor, and a subcarrier constrainer; The encoder encodes and power-constrains the signal according to the transmitted signal and the input interference features; The channel simulator superimposes the effects existing in the real channel on the outputs of the encoder and the random interference generator during training; The decoder decodes according to the output of the encoder after channel simulation and the input interference features; The interference feature extractor extracts the features of the interference according to the output of the random interference generator after channel simulation; The random interference generator simulates various interferences that may occur in the environment as an interference source; The subcarrier constrainer constrains the subcarriers output by the encoder according to the input interference information; The encoding of the encoder is based on the one-dimensional convolutional layer and the fully connected layer of the neural network, performs dimensional transformation on the input signal, and the output dimension is equal to the number of subcarriers, and modulates the signal on each subcarrier; The power constraint normalizes the encoded output and constrains the transmission power; The decoding in the decoder is based on the one-dimensional convolutional layer, the fully connected layer, and the regularization layer of the neural network, performs dimensional transformation on the input, and outputs the transmitted signal; The feature extraction of the interference feature extractor includes the interference feature extraction before the encoder and the feature extraction before the decoder; The feature extraction is based on the one-dimensional convolutional layer, the pooling layer, and the fully connected layer of the neural network, performs dimensional transformation on the input interference, and finds the low-dimensional representation of the interference signal; The interference feature extractor before the encoder is trained with the system bit error rate as the constraint, and the interference feature extractor before the decoder is trained with the system bit error rate and the output of the interference feature extractor before the encoder as the constraints; The inputs of the interference feature extractor before the encoder and the interference feature extractor before the decoder are staggered in time to extract features independent of time for the interference.

2. The physical layer structure of wireless communication based on deep learning according to claim 1, wherein The normalization refers to dividing the signal amplitude on each subcarrier by the maximum value of the signal amplitude on that subcarrier.

3. The physical layer structure of wireless communication based on deep learning according to claim 1, characterized in that, The effects in the channel simulator include carrier frequency offset CFO, sampling frequency offset SFO, Gaussian noise, frequency conversion, and pulse shaping.

4. The physical layer structure of the wireless communication based on deep learning according to claim 1, wherein The output of the interference feature extractor is connected to the encoder and the decoder to guide the encoder and the decoder and synchronize the information of the encoder and the decoder.

5. The physical layer structure of wireless communication based on deep learning according to claim 1, wherein The subcarrier constraint is based on the one-dimensional convolutional layer and the fully connected layer of the neural network, inputs the output of the random interference generator after channel simulation, outputs the subcarrier frequency band selection vector of the encoder, and multiplies it with the output of the encoder.

6. The physical layer structure of the wireless communication based on deep learning according to claim 5, wherein The dimension of the subcarrier frequency band selection vector is the same as the output of the encoder, and it is a vector containing only 0 / 1.

Images