A semantic communication method based on deep neural network

Abstract

The present invention discloses a semantic communication method based on a deep neural network, comprising: first inputting sentence information into a semantic encoder to output semantic information; using a channel encoder to further encode the semantic information into channel information; inputting the channel information into a channel feature extractor to extract channel features; simultaneously inputting the channel features and channel information into a channel decoder, combining the two features through a bilinear transformation and decoding to recover the estimated semantic information; and using a semantic decoder to recover the estimated sentence information. Simulation results show that the proposed semantic communication network outperforms traditional benchmarks and existing semantic communication systems in terms of semantic transmission accuracy under frequency-selective channels, and is robust to different frequency-selective channels. Based on pre-trained channel encoders and decoders, the deep neural network is trained end-to-end, wherein cross entropy is used as the loss function. The training efficiency can be effectively improved by using different loss function methods.

Description

Technical Field

[0001] The present invention belongs to the field of semantic communication technology, and specifically relates to a semantic communication method based on deep neural network. Background Art

[0002] Traditional wireless communications use model-based design, requiring mathematical models or expert knowledge. However, mathematical models rarely fully reflect reality, and individual modules are typically designed and optimized independently, making it difficult to achieve global system optimization. Therefore, wireless communications require a new paradigm to address this problem, and semantic communication based on deep learning (DL) has been seen as a potential solution to the bottlenecks of traditional communication systems. Unlike traditional communication systems that primarily focus on error-free transmission at the bit or symbol level, semantic communication systems process and transmit data at the semantic level to improve the accuracy of the delivered semantics. This communication architecture realizes the second-level communication system envisioned by Shannon and Weaver, namely the transmission of desired meaning, making communication more intelligent and efficient.

[0003] A typical semantic communication model consists of two layers: the semantic layer and the transport layer. The semantic layer extracts and interprets semantic information from messages by encoding and decoding it based on shared background knowledge. The transport layer transmits signals in a specific manner to ensure accurate transmission of semantic information. Natural language processing (NLP) has developed mature methods for extracting and processing semantic information at the semantic layer, enabling machines to understand human grammar and text, and ultimately, natural language. The Transformer model is a leading example in this field. It boldly abandons the recurrent neural network (RNN) and convolutional neural network (CNN) architectures previously used in NLP, instead employing an attention mechanism based on a fully connected neural network (FCN). This attention mechanism assigns more weight to important parts of a neural network, similar to how humans pay attention to key points. This enables it to understand sentences of most lengths. Compared to the other two methods, the Transformer focuses more on useful semantic information that improves the performance of various NLP tasks. It combines the parallel computing capabilities of CNNs with the ability of RNNs to extract semantic information from long sentences. The existing semantic communication framework based on this model can achieve good results under the additive white Gaussian noise (AWGN) channel.

[0004] However, it is worth noting that existing research lacks discussion on wireless channel losses in the transmission layer. In actual wireless communication scenarios, the impact of wireless channels, such as fading and multipath effects, can significantly impair transmission efficiency and accuracy. Therefore, various methods have been proposed to overcome channel losses in traditional communication systems. The traditional method is to insert pilots to estimate the fading channel and use the estimated channel parameters to recover the transmitted data. However, almost all current semantic communication systems adopt a deep neural network (DNN)-based structure and an end-to-end training strategy. This pilot-driven paradigm is incompatible with end-to-end communication systems. The use of pilots greatly complicates the design because the receiver must take two heterogeneous sources (i.e., received data and estimated channel parameters) as input and solve the inverse problem without any prior knowledge. In addition, the inserted pilots themselves are also additional overhead, which reduces transmission efficiency. Summary of the Invention

[0005] Purpose of the invention: By introducing a convolutional neural network channel feature extraction and data recovery module into the receiver, the channel information is implicitly extracted directly from the received signal and the semantic information is recovered. This effectively adapts to frequency-selective channels, saves the overhead caused by pilot signals, and improves transmission efficiency. A semantic communication method based on deep neural networks is proposed.

[0006] In order to achieve the above object, the present invention adopts the following technical solution: a semantic communication method based on deep neural network, comprising the following steps:

[0007] Step A: For the text information input into the transmitter, it is first encoded using a semantic encoder to obtain semantic information, where the semantic encoder is based on a deep neural network;

[0008] Step B: Input the semantic information into the channel encoder, and then further encode it through the channel encoder to obtain channel information, where the channel encoder is based on a deep neural network;

[0009] Step C: transmitting the channel transmission information to the receiver through the channel;

[0010] Step D: After the channel information passes through the channel and enters the receiver, the channel information is first input into the channel feature extractor, which implicitly obtains the channel state information. The channel feature extractor is based on a convolutional neural network, which includes at least one convolutional layer and one fully connected layer.

[0011] Step E: Input the channel state information and the information passing through the channel into a channel decoder to obtain recovered channel information, wherein the channel decoder is based on a deep neural network;

[0012] Step F: Input the recovered channel information into the semantic decoder for decoding to recover the original text information.

[0013] Further, the semantic encoder in step A includes an embedding layer and a Transformer layer, and the specific steps for the semantic encoder to encode text information into semantic information are as follows:

[0014] The text information will be batch - input into the transmitter for transmission, and the text information is denoted as In the semantic encoder in the transmitter, sentences with lengths less than L in the text information are padded to L with special symbols, and then the text information is converted into dense word vectors through the embedding layer in the semantic encoder, and the dense word vectors are denoted as where D is the encoding length designed by the embedding layer for each word, that is, the embedding dimension. Then, the dense word vectors pass through several Transformer layers in the semantic encoder to output intermediate vectors, and the intermediate vectors are denoted as where M is the output dimension of the Transformer layer. Then, a forward feedback network is used to compress the intermediate vectors into semantic information with a lower dimension, and the semantic information is denoted as where B is the number of sentences input in the same batch, L is the maximum length of all sentences, N < M, the forward feedback network consists of two linear layers, and the forward feedback network uses ReLU as the activation function and contains a dropout layer.

[0015] Further, the channel encoder in step B is completed by a convolutional neural network, which includes at least one convolutional layer, and step B specifically includes:

[0016] First, adjust the dimension of the semantic information to After that, it is divided into n segments, and the length of each segment is not less than the input length W of the channel encoder. Then, the n - segment semantic information is encoded into n - segment channel information represented by complex numbers [x1, x2, …, x n , where where n needs to satisfy nW≥LN to ensure that all semantic information is channel - encoded, and when the equality does not hold, N needs to be padded with zeros to LN. The power of the transmission vector x i should be normalized to 1.

[0017] Further, the channel feature extractor in step D is completed by a convolutional neural network, which includes four convolutional layers and two fully - connected layers, and step D specifically includes:

[0018] The channel information [x1, x2, …, x n is divided into n segments of information [y1, y2, …, y n passing through the channel after passing through the channel, where the information passing through the channel Then [y1,y2,…,y n ] is input into the channel feature extractor, which then implicitly obtains the channel state information Where K is the number of channel taps, l c is the length of the second fully connected layer of the channel feature extractor, and the channel is a frequency selective channel with an additive white Gaussian noise channel.

[0019] Furthermore, the channel decoder in step E includes a bilinear transformer and a convolutional neural network, and step E specifically includes:

[0020] Channel State Information and information passing through the channel Input channel decoder, channel state information is converted into and information passing through the channel Combined using bilinear transformations to get and Then z i Reorganized into Then enter the convolutional neural network of the channel decoder to obtain the restored channel information, where Represents the vector outer product.

[0021] Furthermore, the semantic decoder in step F includes an embedding layer and a Transformer layer, and step F specifically includes:

[0022] First, combine the recovered channel information into Discard the last nW-LN filled data in step B and reshape them to get Will The semantic decoder is then used to recover the original text information through the forward feedback network. The original text information is recorded as The feedforward network consists of two linear layers, and the feedforward network uses ReLU as the activation function. The feedforward network includes a dropout layer.

[0023] Furthermore, the training steps of the deep neural network used in the semantic communication method specifically include:

[0024] Step 1: On a channel with uniformly distributed PDP, the channel encoder and channel decoder are trained using binary input and binary cross-entropy loss function.

[0025] Step 2: The channel encoder and decoder are further trained using the mean square error (MSE) loss function on a channel with the same power delay (PDP) distribution as in the actual communication scenario.

[0026] Step 3: Based on the channel encoder and channel decoder trained in steps 1 and 2, the entire semantic communication network including the semantic encoder and semantic decoder is trained end-to-end.

[0027] Furthermore, the step 1 specifically includes:

[0028] The binary input n is input to the channel encoder, and then the output of the channel encoder reaches the channel decoder through a channel with uniform distribution PDP. The output of the channel decoder is processed by the Sigmoid function, and the final output is expressed as Finally, the channel encoder and decoder are combined through stochastic gradient descent and binary cross entropy loss function For training, binary cross entropy loss function Specifically:

[0029]

[0030] The binary input refers to a small batch of randomly generated bit vectors, n i and n and The i-th element, n i ∈{0,1}.

[0031] Furthermore, the step 2 specifically includes:

[0032] The randomly generated data m that follows the standard normal distribution is input into the channel encoder. Then the output of the channel encoder reaches the channel decoder through the channel with the same power delay PDP distribution in the actual communication scenario. The output of the channel decoder is not processed by the Sigmoid function. The final output is expressed as Finally, the channel encoder and channel decoder are connected by stochastic gradient descent and minimum mean square error loss function For training, the minimum mean square error loss function Specifically:

[0033]

[0034] Furthermore, step 3 specifically includes: based on the channel encoder and channel decoder trained in steps 1 and 2, a small batch of semantic sentences s are selected from the data set as input, and then after passing through the entire network in steps A to F, an estimated semantic sentence is obtained, the estimated semantic sentence is input into the semantic decoder, and finally the semantic decoder decodes the estimated semantic sentence to restore the original semantic sentence Then use cross entropy loss To train the entire network from step A to F, the cross entropy loss Specifically:

[0035]

[0036] Where s(w l )and Respectively indicate that the lth word appears in s and The actual probability and predicted probability in w l Represents the lth word.

[0037] Beneficial effects: A data recovery module is composed of a semantic encoder, a pre-trained channel encoder, a pre-trained channel decoder and a semantic decoder to encode and decode the input text information. At the same time, a channel feature extractor is used in the transmission process to implicitly obtain channel state information, thereby implicitly extracting channel information and recovering semantic information directly from the received signal, thereby effectively adapting to frequency selective channels, saving the overhead caused by pilot signals, and improving transmission efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a schematic diagram of the framework flow of the present invention;

[0039] Figure 2 It is a network structure diagram of the present invention;

[0040] Figure 3 It is a line graph of the relationship between BLEU score and SNR on the AWGN channel and the frequency selective channel of the present invention;

[0041] Figure 4 It is a line graph of the 1-gram BLEU scores of channels of different PDPs of the present invention. DETAILED DESCRIPTION

[0042] The present invention will be further explained below with reference to the accompanying drawings.

[0043] The present invention proposes a semantic communication method based on deep neural networks, which uses deep neural networks to perform semantic extraction and coding compression at the semantic layer and realize channel feature extraction at the communication layer, and can show better efficiency than traditional communication systems under frequency-selective channels.

[0044] The following is an embodiment of the present invention providing a semantic communication method based on a deep neural network, comprising the following steps:

[0045] Step A: For the text information input into the transmitter, it is first encoded using a semantic encoder to obtain semantic information, where the semantic encoder is based on a deep neural network;

[0046] Step B: Input the semantic information into the channel encoder, and then further encode it through the channel encoder to obtain channel information, where the channel encoder is based on a deep neural network;

[0047] Step C: Transmit the channel transmission information to the receiver through the channel;

[0048] Step D: After the channel information passes through the channel, it enters the receiver. The channel information is first input into the channel feature extractor, and the channel feature extractor implicitly obtains the channel state information, where the channel feature extractor is based on a convolutional neural network and at least includes a convolutional layer and a fully connected layer;

[0049] Step E: Input the channel state information and the information passing through the channel into the channel decoder to obtain the restored channel information, where the channel decoder is based on a deep neural network;

[0050] Step F: Input the restored channel information into the semantic decoder for decoding to restore the original text information.

[0051] Please refer to Figure 1-2 , in this embodiment, for the text information input into the transmitter in step A, it is first encoded by a semantic encoder, where the semantic encoder is based on a deep neural network. Among them, the semantic encoder includes an embedding layer and a Transformer layer. Step A is specifically: the text data will be batch-input into the framework for transmission, denoted as B is the number of sentences input in the same batch, L is the maximum length of all sentences, and sentences with a length less than L are padded to L with special symbols. It is converted into a dense word vector through the embedding layer of the semantic encoder where D is the encoding length designed by the embedding layer for each word, that is, the embedding dimension. The dense word vector passes through several Transformer layers to output an intermediate vector where M is the output dimension of the Transformer layer. The number of Transformer layers that the dense word vector passes through is 3, and the Transformer adopts a multi-head mechanism, and the number of multi-heads is 8.

[0052] For the semantic information obtained after encoding by the semantic encoder in step B, it is input into the channel encoder for further encoding, where the channel encoder is based on a deep neural network. Among them, the specific method of inputting the semantic information M encoded by the semantic encoder into the pre-trained channel encoder is: first, use a forward feedback network to compress M into a semantic information vector with a lower dimension ​​​​], where the length of each segment is greater than or equal to the channel encoder input length W. n must satisfy nW ≥ LN to ensure that all semantic information is channel-encoded. If the equality does not hold, N must be padded to LN with zeros. The feedforward network consists of two linear layers, preferably using ReLU as the activation function. The network preferably includes a dropout layer.

[0053] The channel encoder is implemented by a convolutional neural network and contains at least one convolution layer. Specifically, n segments of semantic information will be encoded into n segments of complex number representation transmission vectors [x1, x2, ..., x n ],in In addition, the transmission vector x i The power should be normalized to 1. The number of layers of the convolutional neural network is preferably 4, and the activation function is preferably ReLU.

[0054] In step C, the channel transmission information is transmitted to the receiver through the channel. The channel is a frequency selective channel with additive white Gaussian noise, and the number of channel taps is K.

[0055] In step D, after the channel information passes through the channel, the information passing through the channel is obtained, and then the information passing through the channel is input into the channel feature extractor. The channel feature extractor implicitly obtains the channel state information, and the channel state information is used to recover the channel information. The channel feature extractor is based on a convolutional neural network, which includes at least one convolution layer and one fully connected layer. The signal after passing through the channel is divided into n segments, which are the information passing through the channel [y1, y2, ..., y n ],in

[0056] The channel feature extractor is implemented by a convolutional neural network, which contains four convolutional layers and two fully connected layers. The number of input features of the first fully connected layer is L+K-1. The activation function used is ReLU. The channel state information is where l c is the length of the second fully connected layer. The channel passed is a frequency selective channel with additive white Gaussian noise, and the number of channel taps is K.

[0057] In step E, the channel state information and channel transmission information received by the receiver are input into a pre-trained channel decoder, where the channel decoder is based on a neural network. The channel state information and channel transmission information are c i and y i The channel decoder consists of a bilinear transformer and a convolutional neural network. i and y i After inputting the pre-trained channel decoder, this c i and y iCombined using bilinear transformations to get in represents the vector outer product, z i Reorganized into Then enter the convolutional neural network. The convolutional neural network contains at least one convolution layer.

[0058] After the above process, the channel decoder outputs the recovered channel information Among them, the number of convolutional layers is preferably 6, and the activation function is preferably ReLU.

[0059] In step F, the recovered channel information is input into the semantic decoder for decoding to recover the original text information. The specific steps are: first, merge the n segments of recovered channel information into Discard the last nW-LN filled data in step B and reshape them to get after, The semantic decoder is fed through the forward feedback network. The semantic decoder includes an embedding layer and a Transformer layer. The original text information is restored through the semantic decoder. The feedforward network consists of two linear layers using ReLU as the activation function, and includes a dropout layer. It passes through three Transformer layers, using a multi-head mechanism with eight heads.

[0060] The training process of the deep neural network in the above embodiment is as follows:

[0061] In step 1, the channel encoder and channel decoder are trained using binary input and binary cross entropy loss function on a channel with uniform power delay profile (PDP). The specific training steps are: a small batch of randomly generated bit vectors, denoted as n, each element n i ∈{0,1}, is input into the channel encoder. The encoder output will reach the channel decoder through a channel with uniform distribution PDP, and the output of the channel decoder is processed by the Sigmoid function. The final output is expressed as Finally, the channel encoder and decoder are constructed by stochastic gradient descent and binary cross entropy loss function. For training, binary cross entropy loss function The loss function is specifically:

[0062]

[0063] where n i and n and The binary input is a small batch of randomly generated bit vectors.

[0064] In step 2, the channel encoder and decoder are further trained using the mean square error (MSE) loss function on a channel with the same power delay (PDP) distribution as in the actual communication scenario. The specific training steps are as follows: the input is randomly generated data m that follows a standard normal distribution. The encoder output will reach the channel decoder through a channel with the same power delay (PDP) distribution as in the actual communication scenario. The output of the channel decoder is not processed by the Sigmoid function. The final output is expressed as Finally, the channel encoder and channel decoder are constructed by stochastic gradient descent and minimum mean square error loss function. For training, the minimum mean square error loss function The loss function is specifically:

[0065]

[0066] In step 3, based on the channel encoder and decoder trained in steps 1 and 2, the entire network from steps A to F is trained end-to-end, where cross entropy is used as the loss function. Based on the channel encoder and decoder trained in steps 1 and 2, a small batch of sentences s is selected from the dataset as input to obtain an estimated semantic sentence. The estimated semantic sentence then passes through the entire network from steps A to F and is input into the semantic decoder. Finally, the semantic decoder decodes the estimated semantic sentence to restore the original semantic sentence. Then use cross entropy loss To train the entire network from steps A to F. The training process can be expressed as:

[0067]

[0068] Where s(w l )and Respectively indicate that the lth word appears in s and The actual probability and predicted probability in w l Represents the lth word.

[0069] The implementation effect of the present invention is compared with other communication frameworks as shown in the figure Figure 3As shown. Among them, the BLEU (Bilingual Evaluation Understudy) score is an indicator for evaluating the quality of machine translation output. It calculates the score by comparing the similarity between machine translation and manual reference translation, and the value range is from 0 to 1. DeepSC is a semantic communication framework that already has good results in the industry. At the same time, as a benchmark, the DeepSC network and the traditional scheme have linear ZF channel equalization with perfect and imperfect CSI before the receiver to overcome the influence of frequency selective channels. The results show that the present invention is twice as good as DeepSC equalized with estimated CSI under frequency selective channels. At the same time, the present invention is also slightly better than DeepSC equalized with perfect CSI, and this benchmark is an ideal case. The present invention is superior to traditional benchmarks in all aspects of frequency selective channels, and performs particularly well in low signal-to-noise ratio environments.

[0070] Figure 4 The proposed system is simulated on a frequency-selective channel with different exponentially distributed PDPs to test its robustness. The power of the first path is 1, while the power of the last branch is pK. Taking a signal-to-noise ratio of 9dB as an example, and with similar results for other signal-to-noise ratios, the network performance of the proposed method remains nearly unchanged, outperforming the baseline solution for all PDP distributions.

[0071] An embodiment of the present invention takes only 3.72 ms to run on a computer equipped with an Intel Core i5-12400F CPU @ 2.50 GHz and an NVIDIA GeForce GTX 3060, which is faster than the benchmark DeepSC using zero-forcing equalization, which takes 5.43 ms, and the traditional solution, which takes 10.6 ms, and has lower complexity.

[0072] The present invention is applicable to additive white Gaussian noise channels and frequency selective channels.

[0073] In an embodiment of the present invention, sentence information is first input into a semantic encoder to output semantic information; the semantic information is further encoded into channel information using a channel encoder; the channel information is input into a channel feature extractor to extract channel features; the channel features and channel information are simultaneously input into a channel decoder, and the two features are combined through bilinear transformation and then decoded to recover estimated semantic information; and the estimated sentence information is recovered using a semantic decoder.

[0074] The simulation results of the embodiments of the present invention show that the proposed semantic communication network outperforms the traditional benchmark and the existing semantic communication system in terms of semantic transmission accuracy under frequency selective channels, and is robust to different frequency selective channels. Based on the pre-trained channel encoder and decoder, the deep neural network is trained end-to-end, where the cross entropy is used as the loss function. The method of using different loss functions in three steps can effectively improve the training efficiency. The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, without departing from the principle of the present invention, several improvements and modifications can be made, and these improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A semantic communication method based on deep neural network, characterized in that: The following steps are involved: Step A: For the text information input into the transmitter, it is first encoded using a semantic encoder to obtain semantic information, where the semantic encoder is based on a deep neural network; Step B: Input the semantic information into the channel encoder, and then further encode it through the channel encoder to obtain channel information, where the channel encoder is based on a deep neural network; Step C: transmitting the channel transmission information to the receiver through the channel; Step D: After the channel information passes through the channel and enters the receiver, the channel information is first input into the channel feature extractor, which implicitly obtains the channel state information. The channel feature extractor is based on a convolutional neural network, which includes at least one convolutional layer and one fully connected layer. Step E: Input the channel state information and the information passing through the channel into a channel decoder to obtain recovered channel information, wherein the channel decoder is based on a deep neural network; Step F: Input the recovered channel information into the semantic decoder for decoding to restore the original text information; The semantic encoder in step A includes an embedding layer and a Transformer layer, and the semantic encoder encodes text information into semantic information in the following specific steps: The text information will be batch - input into the transmitter for transmission, and the text information is denoted as In the transmitter, the semantic encoder fills the sentences in the text information with a length less than L to L using special symbols, and then converts the text information into dense word vectors through the embedding layer in the semantic encoder. The dense word vectors are denoted as where D is the encoding length designed by the embedding layer for each word, that is, the embedding dimension. Then, the dense word vectors pass through several Transformer layers in the semantic encoder to output intermediate vectors, and the intermediate vectors are denoted as where M is the output dimension of the Transformer layer. Then, the intermediate vectors are compressed into semantic information with a lower dimension using a forward - feedback network, and the semantic information is denoted as where B is the number of sentences input in the same batch, L is the maximum length of all sentences, N < M, the forward - feedback network consists of two linear layers, and the forward - feedback network uses ReLU as the activation function. The forward - feedback network contains a dropout layer; The channel encoder in step B is implemented by a convolutional neural network, including at least one convolutional layer, and step B specifically includes: First adjust the dimension of semantic information to Then it is divided into n segments, the length of each segment is not less than the channel encoder input length W, and then the n segments of semantic information will be encoded into n segments of complex channel information [x1, x2, ..., x n ],in Where n needs to satisfy nW ≥ LN to ensure that all semantic information is channel coded, and when the equality does not hold, N needs to be padded with zeros to LN. The transmission vector x i The power of should be normalized to 1; The channel feature extractor in step D is implemented by a convolutional neural network, comprising four convolutional layers and two fully connected layers, and step D specifically includes: Channel information [x1,x2,…,x n ] is divided into n segments of information passing through the channel [y1,y2,…,y n ], where the information through the channel Then [y1,y2,…,y n ] is input into the channel feature extractor, which then implicitly obtains the channel state information Where K is the number of channel taps, l c is the length of the second fully connected layer of the channel feature extractor, and the channel is a frequency selective channel with an additive white Gaussian noise channel; The channel decoder in step E includes a bilinear transformer and a convolutional neural network, and step E specifically includes: Channel State Information and information passing through the channel Input channel decoder, channel state information is converted into and information passing through the channel Combined using bilinear transformations to get and Then z i Reorganized into Then enter the convolutional neural network of the channel decoder to obtain the restored channel information, where represents the vector outer product; The semantic decoder in step F includes an embedding layer and a Transformer layer, and step F specifically includes: First, combine the recovered channel information into Discard the last nW-LN filled data in step B and reshape them to get Will The semantic decoder is then used to recover the original text information through the forward feedback network. The original text information is recorded as The feedforward network consists of two linear layers, and the feedforward network uses ReLU as the activation function. The feedforward network includes a dropout layer.

2. The semantic communication method based on deep neural network according to claim 1, characterized in that The training steps of the deep neural network used in the semantic communication method specifically include: Step 1: On a channel with uniformly distributed PDP, the channel encoder and channel decoder are trained using binary input and binary cross-entropy loss function. Step 2: The channel encoder and decoder are further trained using the mean square error (MSE) loss function on a channel with the same power delay (PDP) distribution as in the actual communication scenario. Step 3: Based on the channel encoder and channel decoder trained in steps 1 and 2, the entire semantic communication network including the semantic encoder and semantic decoder is trained end-to-end.

3. The semantic communication method based on deep neural network according to claim 2, characterized in that The step 1 specifically includes: The binary input n is input to the channel encoder, and then the output of the channel encoder reaches the channel decoder through a channel with uniform distribution PDP. The output of the channel decoder is processed by the Sigmoid function, and the final output is expressed as Finally, the channel encoder and decoder are combined through stochastic gradient descent and binary cross entropy loss function For training, binary cross entropy loss function Specifically: The binary input refers to a small batch of randomly generated bit vectors, n i and n and The i-th element, n i ∈{0,1}.

4. The semantic communication method based on deep neural network according to claim 2, characterized in that The step 2 specifically includes: The randomly generated data m that follows the standard normal distribution is input into the channel encoder. Then the output of the channel encoder reaches the channel decoder through the channel with the same power delay PDP distribution in the actual communication scenario. The output of the channel decoder is not processed by the Sigmoid function. The final output is expressed as Finally, the channel encoder and channel decoder are connected by stochastic gradient descent and minimum mean square error loss function For training, the minimum mean square error loss function Specifically:

5. The semantic communication method based on deep neural network according to claim 2, characterized in that The step 3 specifically includes: based on the channel encoder and channel decoder trained in steps 1 and 2, a small batch of semantic sentences s are selected from the data set as input, and then the estimated semantic sentences are obtained after passing through the entire network of steps A to F, and the estimated semantic sentences are input into the semantic decoder. Finally, the semantic decoder decodes the estimated semantic sentences to restore the original semantic sentences. Then use cross entropy loss To train the entire network from step A to F, the cross entropy loss Specifically: Where s(w l )and Respectively indicate that the lth word appears in s and The actual probability and predicted probability in w l Represents the lth word.

Images