Rate-free coding semantic communication transmission method

Abstract

The invention discloses a rateless coding semantic communication transmission method, which combines a DeepSC architecture with a Luby Transform (LT) coding and decoding scheme, realizes high-rate transmission of a DJSCC technology, can realize complete self-adaptability without channel prior state information for different channel conditions, and can realize high-speed transmission of the DJSCC technology, thereby realizing high-speed transmission of the DJSCC technology, realizing high-speed transmission of the DJSCC technology, realizing high-speed transmission of the DJSCC technology and realizing high-speed transmission of the DJSCC technology. And the adaptability of the semantic communication system to different channels and the effectiveness of semantic information transmission under a lossy channel are effectively improved. Meanwhile, a bilingual evaluation substitution score (BLEU score) in the field of machine translation is introduced into the semantic communication, and the BLEU score is used as an index for evaluating the performance of the semantic communication.

Description

Technical Field

[0001] The present invention relates to the technical field of source-channel joint coding, and is particularly applicable to a rate-free coding semantic communication transmission method. Background Art

[0002] The development of modern natural language processing and other artificial intelligence technologies and the improvement of physical computing power have provided a foundation and possibility for the study of semantic communication. A commonly used technology in semantic communication research is DJSCC, which is the Deep Joint Source Channel Coding method, which is a combination of Joint Source Channel Coding (JSCC) and a deep neural network learning algorithm. It can train the network with the goal of maximizing transmission rate and accuracy, extract more advanced features from complex data, and learn deep patterns and rules in large amounts of data.

[0003] In the existing public literature, Xie et al. published an article titled "Deep Learning Enabled Semantic Communication Systems" in Volume 69 of IEEE Transactions on Signal Processing in 2021. They proposed an end-to-end DJSCC communication system DeepSC based on the transformer architecture, which can transmit text. The mutual information is taken into account in the loss function design process of this model, which effectively improves the data transmission rate. Compared with traditional communication methods, it can have higher semantic transmission accuracy under low signal-to-noise ratio conditions. The overall neural network model architecture of DeepSC is shown in the attached figure. Figure 2 However, the information dimension output by DJSCC technology based on DeepSC architecture is fixed. When some information dimensions are interfered, the overall performance will be degraded, resulting in the loss of spectrum resources.

[0004] In order to improve the channel environment adaptability of the DeepSC architecture, Zhou et al. published an article titled "Adaptive bitrate control insemantic communication with incremental knowledge-based HARQ" in Volume 3 of IEEE Open Journal of the Communications Society in 2022, proposing a semantic communication system with self-adjustable code rate to adapt to channel conditions. It can select the appropriate code rate for the current channel conditions for semantic coding by training deep neural networks. However, this scheme requires the receiver to transmit the CSI back to the transmitter as the input of the neural network, which is not true channel adaptation. The channel adaptation mechanism should not require prior state information of the channel. Summary of the invention

[0005] The purpose of the present invention is to provide a rate-free coding semantic communication transmission method to solve the problem that the DJSCC technology under the DeepSC architecture cannot adapt to the channel environment, which will lead to overall performance degradation when some information dimensions are interfered.

[0006] To achieve the above object, the present invention adopts the following technical solutions: The rate-free coding semantic communication transmission method of the present invention comprises the following steps: S1, quantize the vector X output by the DeepSC semantic communication model encoder; S2, LT encoding is performed on the quantized vector Q, and LT decoding is performed at the receiving end after passing through the binary erasure channel. ; S3, the decoded vector After dequantization, the data is input into the DeepSC semantic communication model decoder to achieve rateless coding semantic communication.

[0007] Preferably, the sign part, the integer part and the fractional part of the vector X are quantized using bits respectively.

[0008] Preferably, vector Q implements LT encoding and decoding by bit-level cascading or data packet cascading.

[0009] Preferably, the bit-level cascading method flattens the vector Q, and each source symbol corresponds to 1 bit in the vector Q, thereby obtaining a source symbol set of the LT encoder.

[0010] Preferably, the data packet cascade method flattens the vector Q, and each source symbol corresponds to a bit sequence obtained by quantizing an element in the vector X, thereby obtaining a source symbol set of the LT encoder.

[0011] The advantage of the present invention is that it combines the DeepSC architecture with the Luby Transform (LT) coding and decoding scheme to achieve high-speed transmission of the DJSCC technology while achieving full adaptability to different channel conditions without the need for channel prior state information, effectively improving the adaptability of the semantic communication system to different channels and the effectiveness of semantic information transmission under lossy channels. At the same time, the present invention introduces the bilingual evaluation understudy score (BLEU score) in the field of machine translation into semantic communication as an indicator for evaluating the performance of semantic communication. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 It is a schematic diagram of the process of the method of the present invention.

[0013] Figure 2 It is a flow chart of the existing DeepSC communication system.

[0014] Figure 3 This is a table of specific parameter settings for the neural network in the DeepSC semantic communication model in Example 2 of the present invention.

[0015] Figure 4 This is an illustration of the distribution of bit-level cascaded LT coding symbol values ​​in Embodiment 2 of the present invention.

[0016] Figure 5 It is a performance comparison chart of the bit-level cascaded LT coding DeepSC in the method described in the present invention and the existing DeepSC communication system.

[0017] Figure 6 This is an illustration of the distribution of the cascaded LT coding symbol values ​​of the data packets in Embodiment 2 of the present invention.

[0018] Figure 7 It is a performance comparison chart of the data packet cascade LT coding DeepSC in the method described in the present invention and the existing DeepSC communication system.

[0019] Figure 8 It is a performance comparison chart of the LT-coded DeepSC communication system with packet-level cascading and bit-level cascading in the method of the present invention. DETAILED DESCRIPTION

[0020] The technical solutions in the embodiments of the present invention are described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0021] Glossary: The operation process of DeepSC semantic communication model: First, you need to build a word list based on the data set, such as , through this word list, different words can be mapped to different index values. Figure 2 From the DeepSC semantic communication model shown, it can be seen that the input of the DeepSC semantic communication model is a batch of text index vectors S of the same length, whose size is BxL, where B is the batch size and L is the length of the current text, that is, the number of words it contains. Each element in the vector S represents the index value of a word in the word table v. The text index vector S can then be converted into a semantic feature vector E through the embedding layer, whose size is BxLxE, where E is the vector dimension of each word expanded through the embedding process. It can be seen that the embedding layer can extract the semantic features of different words and present the features in the form of semantic vectors. Furthermore, the semantic feature vector After passing through the semantic encoder, the semantic coding vector M can be obtained. The vector size is BxLxV, where V is the output dimension of the semantic encoder. Through the semantic encoder, the semantic information contained in the vector S will be extracted into the semantic coding vector M by the transformer architecture. This step is equivalent to the source coding process in JSCC. Subsequently, after passing through the channel coding layer, the coding vector X of size BxLx2N can be output. In order to transmit the signal in the form of complex real and imaginary parts, the vector X is further deformed to a size of BxNLx2. After channel transmission, the vector X is received by the receiving end as vector Y. After channel decoding, the vector , and then the original text index vector can be restored at the receiving end through the semantic decoder .

[0022] LT (Luby Transform, LT) code is a typical rateless code that performs well in channel adaptation.

[0023] Example 1 like Figure 1 As shown, the rate-free coding semantic communication transmission method of the present invention comprises the following steps: S1, quantize the vector X output by the DeepSC semantic communication model encoder.

[0024] The vector X output by the DeepSC semantic communication model encoder is of size BxLx2N. Each element in vector X is a signed floating point number. Assume that the floating point number is f, and let its sign, integer, and decimal parts be , and , then . Assume that the three parts are respectively bits are used for quantization, the size of the resulting quantization vector Q is That is, the sign part, integer part and fractional part of vector X are quantized using bits respectively.

[0025] S2, LT encoding is performed on the quantized vector Q, and LT decoding is performed at the receiving end after passing through the Binary Erasure Channel (BEC) .

[0026] For the quantized vector For example, the sub-quantization vector size for direct LT coding is , that is, after LT encoding, it will generate bit data stream. When performing LT encoding and decoding, the quantization result can be converted into the source symbol set of the LT encoder in two ways: , which are bit-level cascading and packet-level cascading respectively.

[0027] For bit-level cascading, each source symbol of the LT encoder contains 1 bit of information, which corresponds to 1 bit in the vector Q. Since the size of the quantized vector Q is , so flattening the quantized vector yields a length of The LT encoder source symbol set.

[0028] For packet-level concatenation, the amount of information contained in each source symbol of the LT encoder is bits, which can correspond to an element in vector X. After flattening the quantized vector Q, the source symbol set size of the LT encoder is .

[0029] S3, the decoded vector Dequantize and restore the vector It is then input into the DeepSC semantic communication model decoder to achieve rateless coding semantic communication.

[0030] Example 2 Specific application effect of the technical solution of the present invention The DeepSC semantic communication model is a combination of Joint Source Channel Coding (JSCC) and a deep neural network learning algorithm, and needs to be trained with a deep neural network learning algorithm.

[0031] The dataset used for training the DeepSC semantic communication model of the present invention is the European Parliament Proceedings Parallel Corpus, which is a very standardized parallel corpus currently available for free on the Internet. The corpus is mainly extracted from the minutes of the European Parliament, including 11 European languages ​​such as Romance languages ​​(French, Italian, Spanish, Portuguese), Germanic languages ​​(English, Dutch, German, Greek and Finnish), etc.

[0032] Set the batch size B=64, text length L=21, semantic encoder output dimension parameter N=6, and the weight of DeepSC's mutual information loss function in the DeepSC semantic communication model. , whose value range is [0,1]. Setting the specific parameters of the DeepSC semantic communication model neural network is as follows Figure 3 shown.

[0033] For the signed floating point numbers in the vector X output by the DeepSC semantic communication model encoder, As far as we are concerned, we only need to extract the sign information of each element in vector X, so we set quantization to one bit, that is, n 1 =1. The inventors found through observation that the maximum absolute value of the integer part i of each element in the vector X does not exceed 7, so three bits are used for quantization, that is, n 2 =3. The fractional part d of each element in vector X has little effect on the value of the entire element, so the quantization fineness of this part can be reduced, and four bits are set to be used for quantization, that is, n 3 =4.

[0034] The LT coding symbol degree distribution used in the present invention can be generated by various degree distribution generation methods. Taking the decreasing ripple size distribution (DRSD) as an example, when using the bit-level cascade method, the length of the degree distribution is , its degree distribution is as follows Figure 4 As shown, the degree values ​​with probability less than 0.01 have been omitted in the table.

[0035] Based on the degree distribution DRSD, the present invention compares the semantic transfer performance of the constructed and trained bit-level DeepSC cascade LT model with that of the DeepSC model without LT, and uses the bilingual evaluation understudy score (BLEU score) in the field of machine translation as the performance indicator for judging the DeepSC system and the DeepSC cascade LT system.

[0036] The Bilingual Evaluation Understudy score (BLEU score) measures the performance of machine translation results by comparing the differences between words in two sentences. To a certain extent, it can characterize the similarity of semantic information between the two sentences.

[0037] For the DeepSC model without LT, the retransmission mechanism is considered in the test comparison. Assuming that the second transmission can achieve the best performance (that is, the highest BLEU score that can be achieved when the erasure rate of the corresponding BEC is 0), the relationship between the complementary value of the BLEU score obtained by comparing the DeepSC model without LT and the bit-level cascaded LT coding DeepSC model and the full recovery decoding overhead is shown in the figure. Figure 5 shown.

[0038] Depend on Figure 5 It can be seen that the LT-free DeepSC scheme has the best performance at the BEC erasure rate, which is also the case where it is better than the bit-level DeepSC cascade LT scheme. However, in actual non-ideal situations, the erasure rate cannot be 0. At this time, the LT-free DeepSC scheme must use a retransmission to make up for the gap between the current BLEU score and the optimal BLEU score that the model can achieve, which will cause the full recovery decoding overhead to increase exponentially.

[0039] The bit-level cascaded LT coding DeepSC model in the present invention can fill the gaps in the original solution. The gap between the current BLEU score and the optimal BLEU score is flexibly achieved by receiving a small amount of more coded symbols (compared to the number of source symbols). In summary, when semantic transmission is performed in the bit-level cascaded LT-coded DeepSC model, compared with the DeepSC model without LT, the bit-level cascaded LT-coded DeepSC model proposed in the present invention can reach the optimal transmission performance point at a lower full recovery decoding overhead, and eliminates the dependence on the retransmission mechanism, effectively reducing the decoding delay and enhancing the adaptability to different channel conditions.

[0040] In addition, when using the packet-level cascaded LT coding DeepSC model, the number of source symbols used for LT encoding and decoding is , DRSD is still used as the degree distribution of LT coding symbols, and its degree value distribution is as follows Figure 6 As shown, the degree values ​​with probability less than 0.01 have been omitted in the table.

[0041] The relationship between the complementary BLEU scores obtained by comparing the DeepSC model without LT and the DeepSC model with packet-level cascaded LT coding and the full recovery decoding overhead is shown in Figure 7 shown.

[0042] Depend on Figure 7 It can be seen that similar to the bit-level transmission method, the packet-level transmission method can also reach the optimal transmission performance point at a lower full recovery decoding overhead. By eliminating the reliance on the retransmission mechanism, the decoding delay is effectively reduced and the adaptability to different channel conditions is enhanced. Then, the relationship between the complementary values ​​of the BLEU scores of the packet-level cascaded LT-coded DeepSC model and the bit-level cascaded LT-coded DeepSC model and the full recovery decoding overhead is further compared. The simulation results are shown in Figure 2. Figure 8 shown.

[0043] Depend on Figure 8 It can be seen that the DeepSC cascaded LT system proposed in the present invention has adaptive characteristics to the channel at both bit-level and packet-level transmission, but the number of source symbols transmitted at the bit-level transmission is greater. Therefore, according to the performance difference of DRSD for different numbers of source symbols, it can be known that the full recovery decoding overhead required under the same circumstances will be lower.

[0044] Example 3 Specific deployment case of the bit-level cascaded LT coding DeepSC model in the technical solution of the present invention Assume that data is transmitted between a gNB and a user equipment (UE), taking data transmission from a gNB to a UE as an example. Assume that DeepSC has been trained and its transmitter and receiver models have been deployed on the gNB and UE respectively.

[0045] The specific implementation steps can be described as follows.

[0046] 1. The vector output by the DeepSC encoder deployed on the gNB Quantize to get vector .

[0047] Let the text index vector of a batch to be sent by gNB be , each batch has sentences, and the maximum length of each sentence is (If the length is insufficient, gap filling is performed; if it exceeds the maximum length, truncation is performed.) After inputting it into the DeepSC transmitter model deployed on the gNB, a size of Vector Quantize the vector and set the number of quantization bits to be So the vector Each element value in implements Bit quantization, that is, the final quantization vector Size .

[0048] 2. Quantized vector Flatten and LT encode, and then perform LT decoding on the UE after BEC to obtain the vector .

[0049] For vector For example, the sub-quantization vector size for direct LT coding is , that is, each batch will have LT encoding is performed on each source symbol set, and each source symbol set is obtained by flattening a sub-quantization vector at the bit level or the data packet level. Perform bit-level LT coding, so the source symbol set size used for LT coding is Therefore, it is necessary to perform LT encoding on 2016 source symbols and send the encoded symbols to the UE via BEC and then perform LT decoding. This process needs to be repeated Second-rate.

[0050] After the UE receives the coded symbols and completes the LT decoding, it will obtain a set of 64 source symbols, and then perform an "inverse flattening" operation to reconstruct the quantized vector at the receiving end. .

[0051] 3. The vector obtained after UE decodes LT Dequantize and restore the vector Then input it into the DeepSC decoder and finally reconstruct the text index vector .

[0052] Example 4 Specific deployment case of the packet-level cascaded LT coding DeepSC model in the technical solution of the present invention Assume that data is transmitted between a gNB and a user equipment (UE), taking data transmission from a gNB to a UE as an example. Assume that DeepSC has been trained and its transmitter and receiver models have been deployed on the gNB and UE respectively.

[0053] The specific implementation steps can be described as follows.

[0054] 1. The vector output by the DeepSC encoder deployed on the gNB Quantize to get vector .

[0055] Let the text index vector of a batch to be sent by gNB be , each batch has sentences, and the maximum length of each sentence is (If the length is insufficient, gap filling is performed; if it exceeds the maximum length, truncation is performed.) After inputting it into the DeepSC transmitter model deployed on the gNB, a size of Vector Quantize the vector and set the number of quantization bits to be So the vector Each element value in implements Bit quantization, that is, the final quantization vector Size .

[0056] 2. Quantized vector Flatten and LT encode, and then perform LT decoding on the UE after BEC to obtain the vector .

[0057] For vector For example, the sub-quantization vector size for direct LT coding is , that is, each batch will have LT encoding is performed on each source symbol set, and each source symbol set is obtained by flattening a sub-quantization vector at the bit level or the data packet level. Perform packet-level LT coding, so the source symbol set size used for LT coding is Therefore, it is necessary to perform LT encoding on the 252 source symbols and send the encoded symbols to the UE via BEC and then perform LT decoding. This process needs to be repeated After the UE receives the coded symbols and completes the LT decoding, it will obtain a set of 64 source symbols, and then perform an "inverse flattening" operation to reconstruct the quantized vector at the receiving end. .

[0058] 3. The vector obtained after UE decodes LT Dequantize and restore the vector Then input it into the DeepSC decoder and finally reconstruct the text index vector .

Claims

1. A rateless coding semantic communication transmission method, characterized in that: The following steps are involved: S1, quantize the vector X output by the DeepSC semantic communication model encoder; S2, LT encoding is performed on the quantized vector Q, and LT decoding is performed at the receiving end after passing through the binary erasure channel. ; S3, the decoded vector After dequantization, the data is input into the DeepSC semantic communication model decoder to achieve rateless coding semantic communication.

2. A rateless coding semantic communication transmission method according to claim 1, characterized in that: The sign part, integer part and fractional part of vector X are quantized using bits respectively.

3. A rateless coding semantic communication transmission method according to claim 1, characterized in that: Vector Q uses bit-level cascading or packet cascading to implement LT encoding and decoding.

4. A rateless coding semantic communication transmission method according to claim 3, characterized in that: The bit-level cascading method flattens the vector Q, and each source symbol corresponds to 1 bit in the vector Q, thereby obtaining a source symbol set of the LT encoder.

5. A rateless coding semantic communication transmission method according to claim 3, characterized in that: The data packet cascade method flattens the vector Q, and each source symbol corresponds to a bit sequence obtained by quantizing an element in the vector X, thereby obtaining a source symbol set of the LT encoder.

Images