Transmitter, receiver and method for wireless communication

Abstract

The disclosure discloses a transmitter, a receiver and a method for wireless communication. The transmitter comprises an encoder, a normalizer, a binarizer and a radio frequency circuit. The encoder is configured to encode a control message into a channel dimension vector by using a training model according to a channel dimension and a semantic column number, wherein the control message comprises at least one of control plane media access control layer information and control plane physical layer information. The normalizer is configured to perform normalization processing on the channel dimension vector to generate a normalized channel dimension vector. The binarizer is configured to binarize the normalized channel dimension vector to generate a fixed point number. The radio frequency circuit is configured to modulate the fixed point number into a radio frequency signal and transmit the radio frequency signal. According to the invention, the fault-tolerant coding and decoding technology based on semantic communication is adopted, and the capability of resisting channel transmission errors can be provided for transmission carried out in various wireless channel environments.

Description

Technical Field

[0001] The present disclosure relates to wireless communication technology, and more particularly to a transmitter, a receiver and a method for wireless communication. Background Art

[0002] 5G New Radio (NR) is a recently developed radio access technology that supports high throughput, low latency, and high capacity communications. In 5G New Radio and similar wireless communication systems, the transmitter usually uses channel coding to encode data to generate the ability to resist channel transmission errors, so that the receiver can successfully decode it into correct data. However, since channel noise can easily cause an increase in the bit error rate (BER), if the bit error rate is too high, the receiver will not be able to fully restore the data, resulting in a waste of transmission resources and a burden of retransmission. Summary of the invention

[0003] In view of the technical problems existing in the prior art, the present disclosure proposes to adopt error-tolerant coding and decoding technology based on semantic communication, which can provide the ability to resist channel transmission errors for transmissions performed in various wireless channel environments.

[0004] One aspect of the present disclosure is a transmitter for wireless communication, comprising an encoder, a normalizer, a binarizer, and a radio frequency circuit. The encoder is configured to encode a control message into a channel dimension vector using a training model and based on a channel dimension and a number of semantic fields, wherein the control message includes at least one of control plane media access control layer (MAC) information and control plane physical layer (Physical Layer) information. The normalizer is configured to normalize the channel dimension vector to generate a normalized channel dimension vector. The binarizer is configured to binarize the normalized channel dimension vector to generate a fixed-point number. The radio frequency circuit is configured to modulate the fixed-point number into a first radio frequency signal and send the first radio frequency signal.

[0005] According to an embodiment of the present disclosure, the control message includes downlink control information (DCI).

[0006] According to another embodiment of the present disclosure, the training model includes a bidirectional long short-term memory (BiLSTM) neural network model.

[0007] According to another embodiment of the present disclosure, the encoder includes a Joint Source-Channel Coding (JSCC) encoder.

[0008] According to another embodiment of the present disclosure, the fixed-point number includes a plurality of redundant sign bits, a plurality of redundant integer bits, and a plurality of fractional bits.

[0009] According to another embodiment of the present disclosure, the encoder is configured to generate at least one of the channel dimension, the bit length of the fixed-point number, and the number of semantic fields according to a channel environment where the transmitter is located.

[0010] According to another embodiment of the present disclosure, before the RF circuit sends the first RF signal to the receiving end, the encoder encodes a setting message including the channel dimension, the bit length of the fixed-point number, and at least one of the semantic field numbers into setting parameter data, and the RF circuit modulates the setting parameter data into a second RF signal and sends the second RF signal to the receiving end.

[0011] Another aspect of the present disclosure is a receiver for wireless communication, comprising a radio frequency circuit, a de-binarizer, and a decoder. The radio frequency circuit is configured to receive a first radio frequency signal and demodulate the first radio frequency signal into a fixed-point number. The de-binarizer is configured to perform a de-binarization process on the fixed-point number to generate a channel dimension vector. The decoder is configured to use a training model and decode the channel dimension vector into a control message according to the channel dimension and the number of semantic fields, wherein the control message includes a control plane media access control layer message or a control plane physical layer message.

[0012] According to one embodiment of the present disclosure, the training model includes a long short-term memory (LSTM) neural network model.

[0013] Another aspect of the present disclosure refers to a wireless communication method, which is used at the transmitting end and includes: using a training model and encoding a control message into a channel dimension vector based on the channel dimension and the number of semantic fields, wherein the control message includes at least one of control plane media access control layer information and control plane physical layer information; normalizing the channel dimension vector to generate a normalized channel dimension vector; binarizing the normalized channel dimension vector to generate a fixed-point number; and modulating the fixed-point number into a radio frequency signal and sending the radio frequency signal.

[0014] The beneficial effect of the present disclosure is at least that, by adopting the error-tolerant coding and decoding technology based on semantic communication, the ability to resist channel transmission errors can be provided for transmissions performed in various wireless channel environments. In particular, compared with the existing wireless communication system, the present disclosure embodiment adopts a control message transmission mechanism based on semantic communication instead of a retransmission mechanism. In an environment with a high bit error rate, the existing wireless communication system may need to retransmit the control message multiple times to correctly decode the control message, but the control message transmission based on the present disclosure embodiment can still extract semantic features through semantic decoding, and then correctly restore the control message without retransmission, so the efficiency of transmission resource utilization can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] For a more complete understanding of the embodiments and their advantages, reference is now made to the following description taken in conjunction with the accompanying drawings, in which:

[0016] Figure 1 is a schematic diagram of a wireless communication system according to an embodiment of the present disclosure;

[0017] Figure 2 A schematic diagram of wireless communication between a transmitting end and a receiving end according to an embodiment of the present disclosure;

[0018] Figure 3 A block diagram of each stage of the model training method according to an embodiment of the present disclosure;

[0019] Figure 4A is a functional block diagram of a transmitter according to an embodiment of the present disclosure;

[0020] Figure 4B is a functional block diagram of a receiver according to an embodiment of the present disclosure;

[0021] Figure 5 A block diagram of each stage of a control message transmission method according to an embodiment of the present disclosure;

[0022] Figure 6 is a flow chart of a wireless communication method according to an embodiment of the present disclosure;

[0023] Figure 7 FIG. 4 is a block diagram of a device according to an embodiment of the present disclosure.

[0024]

Explanation of symbols

[0025] 10: Wireless communication system

[0026] 20: Model training method

[0027] 30: Control message transmission method

[0028] 40: Wireless communication method

[0029] 110: Transmission end

[0030] 120: receiving end

[0031] 210:Transmitter

[0032] 211: Encoder

[0033] 212: Normalizer

[0034] 213: Binarizer

[0035] 214,221:RF circuit

[0036] 220: Receiver

[0037] 222: Debinarizer

[0038] 223:Decoder

[0039] 300: Installation

[0040] 310: Processor

[0041] 320: Memory

[0042] 330: Transceiver

[0043] B21-B27,B31-B39: Blocks

[0044] BS: Base Station

[0045] CN: Core Network

[0046] NW: Network side

[0047] RFS1,RFS2:RFS signal

[0048] S41-S44: Steps

[0049] UE: User Equipment DETAILED DESCRIPTION

[0050] The following is a detailed discussion of the embodiments of the present disclosure. However, it is understood that the embodiments provide many applicable concepts that can be implemented in a variety of specific contexts. The embodiments discussed and disclosed are for illustration only and are not intended to limit the scope of the present disclosure.

[0051] The terms used in this article are only for describing specific embodiments and are not intended to limit the scope of the patent application. Unless otherwise limited, the singular forms of "a", "an" or "the" may also be used to represent plural forms.

[0052] It is understood that, although the terms "first" and "second" used in this article may be used to describe different elements, these elements should not be limited by these terms. These terms are only used to distinguish one element from other elements. The operation of "determining" used in this article can be replaced by the operation of "generating" or "calculating".

[0053] Figure 1 Schematic diagram of a wireless communication system 10 according to an embodiment of the present disclosure. The wireless communication system 10 may be, for example, a fifth generation (5G) (hereinafter referred to as 5G) new radio communication system, a beyond 5G (B5G) communication system, a sixth generation (6G) and / or other similar wireless communication systems (e.g., the evolution of any of the above systems). In the wireless communication system 10, a user equipment (UE) UE is connected to a network end NW via a radio access network (RAN). The network end NW includes a base station (BS) BS and a core network (CN) CN, wherein the base station BS is configured to provide an interface for the user equipment UE to access the radio access network, and the core network CN is configured to provide network services to each user equipment UE and has a variety of core network functions.

[0054] In the wireless communication system 10, each user equipment UE is connected to the base station BS via a wireless interface. In the example of the 5G new radio communication system, the core network CN is also called the 5G core network (5GCore, 5GC) or the fourth generation (Fourth Generation, 4G) mobile communication technology Evolved Packet Core (Evolved Packet Core, EPC) network supporting 5G functions, the base station BS is also called the next generation base station (Next Generation NodeB, gNB), the evolved version of the Evolved Universal Terrestrial Radio Access Network (Evolved Universal Terrestrial Radio Access Network, E-UTRAN) Evolved version of the base station (Evolved NodeB, eNB) or the next generation evolved base station (ng-eNB), and the radio access network can be called the next generation radio access network (Next Generation RAN, NG-RAN) or the evolved version of the universal terrestrial radio access network supporting 5G functions.

[0055] Figure 2 FIG. 1 is a schematic diagram of wireless communication between a transmitting end 110 and a receiving end 120 according to an embodiment of the present disclosure. Figure 2In the embodiment, the transmission end 110 and the receiving end 120 are connected to each other through a wireless channel. The transmission end 110 and the receiving end 120 can perform unidirectional or bidirectional message transmission, wherein the message transmission from the transmission end 110 to the receiving end 120 is downlink transmission, and the message transmission from the receiving end 120 to the transmission end 110 is uplink transmission. Figure 1 Taking the wireless communication system 10 as an example, the transmitting end 110 and the receiving end 120 may be a base station BS and any user equipment UE in the wireless communication system 10 respectively.

[0056] In the disclosed embodiment, the transmitter 110 and the receiver 120 may use machine learning to generate a training model, such as deep learning, reinforcement learning, deep reinforcement learning, etc., and use the training model to control message transmission and reception. Figure 3 The block diagram of each stage of the model training method 20 specifically illustrates the operation of the training stage using the transmission end 110 and the receiving end 120. The blocks of each stage of the model training method 20 can be compiled into a program for execution.

[0057] In the stage corresponding to block B21, a set of reference data sets to be established is selected (which may consist of all semantic descriptions of each semantic field of the control message), and the parameters of the training model are set, which may include but are not limited to channel dimension, batch size, learning rate, feature dimension, training round (Epoch), signal-to-noise ratio, loss function, etc. Then, each semantic sequence corresponding to each semantic field of the reference data set is converted into an integer label sequence as an input sequence, wherein the conversion includes corresponding each semantic description (such as a word or a phrase) in each semantic sequence to an integer label index, and converting all integer label indexes in sequence to an integer label sequence. Different semantic descriptions correspond to different integer label indexes. The channel dimension can be set according to the number of semantic sequences in the reference data set. For example, the smaller the number of semantic sequences, the smaller the channel dimension can be (for example, 8, but not limited to this); the larger the number of semantic sequences, the larger the channel dimension can be (for example, 256, but not limited to this).

[0058] In the stage corresponding to block B22, the input sequence is converted into word embedding vectors with high feature dimensions by a word embedding layer (such as Word2Vec). Through these word embedding vectors, the correlation between different semantic descriptions can be obtained.

[0059] In the stage corresponding to block B23, the encoder of the transmission end 110 compresses the word embedding vector length of each input sequence into a compressed sequence, and inputs the compressed sequence into the training model for processing. The training model used includes (for example, it can be) a bidirectional long short-term memory (Bi-LSTM) neural network model, other suitable neural network models, such as: recurrent neural network (RNN), transformer, bidirectional encoder representation technology based on transformer (BERT), generative pre-trained transformer (GPT), large language model (LLM) or a combination thereof. In the present disclosed embodiment, the training model used by the transmission end 110 is a bidirectional long short-term memory neural network model as an example, which is used to perform bidirectional long short-term memory network encoding in the training phase.

[0060] The bidirectional long short-term memory neural network model performs a forward transfer of the bidirectional long short-term memory on the compressed sequence to generate the hidden state of the compressed sequence. Specifically, after each forward long short-term memory training, the final state of the forward long short-term memory is used as the initial state of the reverse long short-term memory for subsequent forward long short-term memory training. During the training, the forward and reverse hidden states are respectively transferred to the corresponding reverse long short-term memory unit and forward long short-term memory unit, the gradient of the training model is calculated, and the weight of the training model is updated through back propagation. Through the bidirectional long short-term memory training method, the contextual information between the previous and subsequent semantic narratives and the dependencies between the semantic narratives in different input sequences can be extracted.

[0061] In the disclosed embodiment, an activation function may be used in the training model to learn nonlinear relationships and increase nonlinear features.

[0062] Next, the forward and backward hidden states in the output of the bidirectional long short-term memory neural network model are extracted, and these hidden states are connected through a fully connected layer to extract features, and these features are converted into channel dimension vectors represented by floating point numbers. In the stage corresponding to block B24, the transmission end 110 performs normalization processing to normalize the floating point numbers of each channel dimension vector within a fixed range to avoid excessive differences in floating point numbers between different features and affect the prediction ability of the training model. In normalizing the output of each layer in the training model, the sum of the square values ​​of all channel dimension vectors is first calculated, and then the square root of the sum of the square values ​​is calculated, and then each channel dimension vector is multiplied by it to calculate the corresponding normalized channel dimension vector.

[0063] In the stage corresponding to block B25, after the input normalized channel dimension vector is offset, a channel dimension vector containing the offset is generated and transmitted to the receiving end 120 via the simulated wireless channel. The simulated wireless channel may be an additive white Gaussian noise (AWGN) channel or other suitable channels.

[0064] In some embodiments, the encoder of the transmission end 110 includes a joint source-channel coding (JSCC) encoder (hereinafter referred to as the JSCC encoder), which encodes the control message by combining source coding and channel coding to take into account the representation of the semantic sequence and the noise resistance of the physical transmission, thereby providing better control message transmission reliability.

[0065] In the stage corresponding to block B26, after the receiving end 120 receives the channel dimension vector, the decoder therein gradually decodes the received channel dimension vector. The training model used by the decoder can be a long short-term memory (LSTM) neural network model, such as a unidirectional long short-term memory neural network model, other suitable neural network models or a combination thereof. In the disclosed embodiment, the training model used by the decoder is an example of a unidirectional long short-term memory neural network model, which is used to perform unidirectional long short-term memory network decoding in the training stage. In the stage corresponding to block B27, the decoder processes the channel dimension vector to sequentially generate semantic descriptions, and generates the next semantic description at each time point according to the previous input and hidden state, until all the semantic descriptions are decoded, and the generation of the semantic sequence is completed and used as the output sequence. For example, the decoder can use the Softmax function distribution to represent the generation probability of each candidate semantic sequence at the current time point, and use the candidate semantic sequence with the maximum value as the semantic sequence generated at that time.

[0066] In some embodiments, the Dropout function can be used to randomly set the output of some neurons to zero during the decoding process. In short, the output of some neurons can be randomly set to zero for each batch to force other nodes to learn more features, thereby reducing the dependence on a single node and improving the generation capacity of the training model.

[0067] Finally, the decoder of the receiving end 120 calculates the loss based on the cross-entropy (CE) between the candidate semantic sequence and the input sequence, and performs a reverse propagation on the calculated loss to update the model weights to calculate the gradients of all parameters in the training model. In some embodiments, the gradient can be clipped to a maximum value of 0.1 to avoid the problem of gradient explosion. The smaller the cross entropy, the smaller the loss, that is, the better the training model.

[0068] In some embodiments, the decoder of the receiving end 120 includes a joint source channel codec (hereinafter referred to as a JSCC decoder), which decodes the channel dimension vector by combining source decoding and channel decoding to increase the reliability of the candidate semantic sequence obtained by decoding.

[0069] The above training process is to use the reference data set and all candidate semantic sequences to perform a complete training round. In some embodiments, the weights of the training model can be stored after every three training rounds. The training model with the smallest loss can be used for the transmitting end 110 and the receiving end 120 for actual wireless communication.

[0070] The training phase disclosed herein is to synchronously train the encoder and decoder when the encoder and decoder have the same reference data set, and synchronously train the artificial intelligence / machine learning model through the transmission end 110 and the receiving end 120, encode the semantic description or features of the control message, and transmit the encoded signal from the transmission end 110 to the receiving end 120 via the physical layer channel, and the receiving end 120 decodes the received signal into a control message, and can still correctly restore the semantics and features when the bit error rate is high (for example, 3%). Through the above encoding method, the data can be further compressed to reduce the amount of data transmitted. At the same time, the above encoding method extracts semantic features from the control message for channel coding, which also helps to improve the error resistance. Therefore, even if the transmission is distorted by noise or interference in the physical layer channel, the receiving end can still accurately restore the original control message by extracting semantic features through decoding.

[0071] Figure 4A and Figure 4B The following are functional block diagrams of a transmitter 210 and a receiver 220 according to an embodiment of the present disclosure. The transmitter 210 may be, for example, Figure 1 Base station BS or Figure 2The transmitting end 110 and correspondingly the receiver 220 may be, for example, Figure 1 Any user equipment UE or Figure 2 The receiving end 120. Figure 4A and Figure 4B As shown, the transmitter 210 includes an encoder 211, a normalizer 212, a binarizer 213 and a radio frequency circuit 214, and the receiver 220 includes a radio frequency circuit 221, a de-binarizer 222 and a decoder 223. Figure 5 The block diagram of each stage of the control message transmission method 30 specifically illustrates the operation of wireless communication using the transmitter 210 and the receiver 220. The blocks of each stage of the control message transmission method 30 can be compiled into a program for execution.

[0072] The encoder 211 is configured to select a reference data set, and the encoder 211 is configured to use the training model and encode the control message into a channel dimension vector according to the channel dimension and the number of semantic fields. The control message may include control plane media access control layer information, control plane physical layer information, or a combination of the above information. The control message may include downlink control information (Downlink Control Information, DCI) as shown in Table 1 or other control information for configuration of the receiving end 120, such as: System Information Block (System Information Block, SIB) 6 / 7 / 8, warning notification information broadcast by the earthquake and tsunami warning system / commercial mobile alert service, control elements (Control Elements, CE) of the media access control layer, etc.

[0073] Specifically, in the stage corresponding to block B31, the encoder 211 first combines the control message into a semantic sequence according to the channel dimension, and uses the training model to convert the semantic sequence into an integer tag sequence as an input sequence, wherein each integer tag index in the integer tag sequence corresponds to a specific semantic description. Then, the encoder 211 converts the integer tag sequence into a channel dimension vector using the training model and the number of semantic fields. In detail, in the stage corresponding to block B32, the encoder 211 converts the input sequence into a word embedding vector with high feature dimension through a word embedding layer (e.g., Word2Vec), and then in the stage corresponding to block B33, the encoder 211 compresses the word embedding vector length of each input sequence into a compressed sequence, and inputs the compressed sequence into the training model for processing. The training model may include a bidirectional long short-term memory neural network model, so that the encoder 211 can obtain the dependency between each feature in the corresponding channel dimension vector through the bidirectional long short-term memory network mechanism.

[0074] In some embodiments, the control message includes downlink control information. The 3GPP standard defines multiple formats of downlink control information, including downlink control information format 1_0 (DCI format 1_0), downlink control information format 1_1 (DCI format 1_1), and / or other formats that are not listed but are also applicable to the present disclosure. For example, the content of downlink control information format 1_0 is shown in Table 1.

[0075] Table 1

[0076]

[0077]

[0078] According to the content shown in Table 1, the downlink control information format 1_0 has 12 fields. If the control message includes downlink control information in the format shown in Table 1, these 12 fields are respectively used as semantic fields of the control message, that is, the number of semantic fields of the control message is 12. If the format of the control message is the format shown in Table 1, it includes the downlink control information format 1_0, and all semantic sequences corresponding to all fields can constitute a reference data set. For example, the reference data set constructed according to the content of Table 1 can include about one hundred semantic sequences. In addition, the number of bits in the "Downlink control information format identifier" field is 1, and its binary value can be "0" or "1", where the semantic sequence corresponding to "0" can be "From ue to gnb information", and the semantic sequence corresponding to "1" can be "This message will send downlink information". That is, this field corresponds to 2 (that is, 2 1 ) semantic sequences.

[0079] In some variant embodiments, semantic fields with low bit numbers may be combined into combined semantic fields to reduce the number of semantic fields, thereby reducing the amount of data transmission. Taking Table 1 as an example, three fields with a bit number of 1, three fields with a bit number of 2, and two fields with a bit number of 3 in the downlink control information in the format of Table 1 may be combined into one combined field. In this way, the number of fields in the downlink control information is reduced to 5, and the number of semantic fields of the control message containing the downlink control information is reduced to 5. Since the number of bits of the signal transmitted by the transmitter 210 in the physical layer channel is determined by parameters such as the channel dimension and the number of semantic fields, reducing the number of semantic fields helps to reduce the amount of signal transmission (reduce the number of transmission bits), thereby improving the transmission performance of the wireless communication system.

[0080] In addition, the length of the semantic sequence and the number of semantic descriptions can be determined according to system design requirements. Each semantic field can generate multiple semantic sequences, where all semantic descriptions can be a data set used to generate a training model, i.e., a reference data set. In some embodiments, the encoder 211 can convert a K-bit semantic field into 2 K In the training phase, the transmitter and receiver use 2 K Specifically, the transmission end can establish all (i.e., 2 K For example, 1 bit of field information corresponds to 2 kinds of semantic sequences, 2 bits of field information corresponds to 4 kinds of semantic sequences, and so on. The transmitting end can receive 2 K Then, the similarity of the semantic sequences at the transmitting end and the receiving end can be evaluated by using, for example, a bilingual evaluation understudy (BLEU).

[0081] In some embodiments, similar to Figure 2 At the transmission end 110, the encoder 211 includes a JSCC encoder, which encodes the control message by combining source coding and channel coding.

[0082] In the stage corresponding to block B34, the normalizer 212 is configured to perform normalization processing on the channel dimension vector output by the encoder to generate a normalized channel dimension vector. The normalization processing performed by the normalizer 212 can balance the importance of each feature in the channel dimension vector output by the encoder 211, and can also help avoid the problem of gradient disappearance or gradient explosion in the training model, so that the output of each neuron in the training model is kept in a stable vector range.

[0083] In the stage corresponding to block B35, the binarizer 213 is configured to binarize the normalized channel dimension vector. The fixed-point number generated by the binarization process performed by the binarizer 213 is a variable value. The fixed-point number has a bit length, for example, a fixed-point number of 16 bits represents a bit length of 16. The binarization process can be performed using one of the following methods (1) to (2).

[0084] Method (1): Convert a floating point number with a value range of -1 to less than 1 (i.e. [-1,1)) to a binary fixed point number in the Q3 format (consisting of 4 binary digits), Q7 format (consisting of 8 binary digits), or Q15 format (consisting of 16 binary digits) of a binary fixed point format (i.e. Q format) to reduce the amount of data transmission. Taking the Q15 format as an example, after the floating point number -0.7850 is multiplied by 2 to the 15th power, the nearest integer is rounded, i.e. round(-0.7850×2 15 )=-25722, and then convert the obtained integer into a binary fixed-point number, that is, -25722=[1,0,0,1,1,0,1,1,1,0,0,0,0,1,1,0], where round(·) represents an integer function. Other similar formats (such as Q31 format) can also be used.

[0085] Method (2): Convert a floating point number between -2 and 2 (i.e. (-2,2)) to a 16-bit format, where the first 5 bits represent redundant sign bits, bits 6 to 10 are redundant integer bits, and the last 6 are decimal bits. In this 16-bit format, the first 5 bits are redundant sign bits with the same value (i.e. 11111 or 00000, the former represents a positive floating point number, while the latter represents a negative floating point number), and bits 6 to 10 are redundant integer bits with the same value (i.e. 11111 or 00000, the former represents an integer 1, while the latter represents an integer 0). For example, the fixed point number after converting the floating point number -0.7443 in this way is [1,1,1,1,1,0,0,0,0,0,1,0,1,1,1,1,1]. The advantage of this method is that, in addition to reducing the amount of transmission, it can also improve the channel error tolerance and strengthen the protection of the high bits of the floating point number. The number of converted bits, the number of repeated bits corresponding to the sign bit and the integer bit, and the number of fractional bits can be changed accordingly according to design requirements.

[0086] However, the disclosed embodiment may also choose to adopt other suitable binarization processing methods, and is not limited to the above-mentioned methods (1) to (2).

[0087] In the stage corresponding to block B36, the RF circuit 214 is configured to modulate the fixed point data into the RF signal RFS1 and transmit it in the wireless channel to send the RF signal RFS1 to the receiving end. The RF circuit 214 can perform single-output or multi-output RF signal transmission through a single or multiple antennas.

[0088] In some embodiments, before the RF circuit 214 sends the RF signal RFS1 to the receiving end, the encoder 211 encodes the setting message into setting parameter data, and the RF circuit 214 modulates the setting parameter data into the RF signal RFS2, and sends the RF signal RFS2 to the same receiving end via a wireless channel, so that the receiving end can demodulate and decode the RF signal RFS2 to obtain the setting message, and process the RF signal RFS1 according to the content of the setting message. The setting message may include channel dimension, bit length of fixed-point number, number of semantic fields and / or other content that can be used by the receiving end to demodulate and decode the RF signal RFS1.

[0089] In addition, in some embodiments, the encoder 211 may determine the channel dimension, the bit length of the fixed-point number, the number of semantic fields, other related parameters and / or the binarization method based on the channel environment of the transmitter 210 (i.e., the environment of the wireless channel to which the transmitter 210 is connected, such as the channel environment quality, the noise interference level, the multipath attenuation and / or other factors), but is not limited thereto. In some embodiments, the encoder 211 may determine the bit length of the fixed-point number based on the channel environment of the transmitter 210, and the channel dimension may be determined based on the parameters of the training model set in the training stage, and the number of semantic fields may be determined based on the number of fields of the reference data set selected in the training stage.

[0090] In the receiver 220, the RF circuit 221 is configured to receive in the wireless channel and demodulate the received RF signal RFS1 into fixed-point data. Similar to the RF circuit 214 in the transmitter 210, the RF circuit 221 can also receive single-input or multi-input RF signals through a single or multiple antennas.

[0091] In the stage corresponding to block B37, the debinarizer 222 is configured to debinarize the fixed-point number with the bit length to generate the channel dimension vector. The debinarizer 222 corresponds to the binarizer 213. For example, if the binarizer 213 is configured to convert the floating-point number into a fixed-point number in the Q15 format, the debinarizer 222 converts the fixed-point number in the Q15 format into the floating-point channel dimension vector accordingly. A person with ordinary knowledge in the art can directly understand based on the description of the binarizer 213 that the debinarizer 222 debinarizes the fixed-point number in a reverse manner to convert the fixed-point number into the floating-point channel dimension vector.

[0092] In the stage corresponding to block B38, the decoder 223 is configured to use the training model and decode the channel dimension vector into a control message including a control plane media access control layer message or a control plane physical layer message in the channel dimension. The training model may include (for example, may be) a long short-term memory neural network model, such as a unidirectional long short-term memory neural network model, so that the decoder 223 can generate semantic descriptions in sequence through a unidirectional long short-term memory network mechanism, and generate the next semantic description at each time point according to the previous input and hidden state until all semantic descriptions are decoded. In the stage corresponding to block B39, the bilingual evaluation index is used to evaluate the semantic similarity to determine which of the restored semantic sequences is closest to the semantic sequence of the transmitting end, so as to complete the generation of the semantic sequence and serve as the output sequence, that is, to complete the transmission of the semantic message.

[0093] In some embodiments, similar to Figure 2 At the receiving end 120, the decoder 223 includes a JSCC decoder, which decodes the channel dimension vector by combining source decoding and channel decoding to obtain a control message.

[0094] Corresponding to the transmitter 210, in some embodiments, before the RF circuit 221 receives the RF signal RFS1, the RF circuit 221 receives the RF signal RFS2 via a wireless channel, and demodulates the RF signal RFS2 into setting parameter data, and the decoder 223 decodes the setting parameter data to obtain a setting message, and processes the RF signal RFS1 according to the content of the setting message. The setting message may include a channel dimension, a bit length of a fixed-point number, a number of semantic fields, and / or other content that can be used by the RF circuit 221, the debinarizer 222, and the decoder 223 to demodulate and decode the RF signal RFS1.

[0095] The following example illustrates wireless communication between the transmitter 210 and the receiver 220. In the transmitter 210, the encoder 211 first encodes the control message such as "This message will send downlink information." to generate integer tag indexes 7, 118, 41, 122, 110, 18 corresponding to the semantic descriptions "This", "message", "will", "send", "downlink", and "information" in the control message, and generates an integer tag index 2 corresponding to the "." in the control message to indicate the end of the control message, and converts all integer tag indexes into an integer tag sequence [7, 118, 41, 122, 110, 18, 2] in sequence. After the integer tag sequence [7, 118, 41, 122, 110, 18, 2] is converted, the channel dimension vector [-0.3484, 5.2276, 7.9998, -8.6602,

[0096] Next, the normalizer 212 performs normalization processing on the channel dimension vector to obtain a normalized channel dimension vector [-0.0376, 0.5645, 0.8627, -0.9325, -0.7065, -0.7850, 0.4853, -0.8428]. After that, the binarizer 213 performs binarization processing on the normalized channel dimension vector in Q15 format to obtain a fixed-point number [[1, 1, 1, 1, 1, 0, 1, 1, 0, 1, 1, 1, 1],

[0097] [0,1,0,0,1,0,0,0,0,1,0,0,0,0,1,0],

[0098] [0,1,1,0,1,1,1,0,0,1,1,0,1,1,0,1],

[0099] [1,0,0,0,1,0,0,0,0,1,0,0,1,1,0,0],

[0100] [1,0,1,0,0,1,0,1,1,0,0,1,0,0,1,0],

[0101] [1,0,0,1,1,0,1,1,1,0,0,0,0,1,1,0],

[0102] [0,0,1,1,1,1,1,0,0,0,0,1,1,1,0,1],

[0103] [1,0,0,1,0,1,0,0,0,0,1,0,0,0,0,0]], and the RF circuit 214 modulates the fixed-point number into a RF signal and transmits it to the receiver 220 through a wireless channel. In the receiver 220, the RF circuit 221 first demodulates the received RF signal into a fixed-point number [[1,1,1,1,1,0,1,1,0,0,1,0,1,1,1,1],

[0104] [0,1,0,0,0,0,0,0,0,1,0,1,0,0,0,1,0],

[0105] [0,1,1,0,1,1,1,0,0,1,1,0,1,1,0,1],

[0106] [1,0,0,0,1,0,0,0,0,1,0,0,1,1,0,0],

[0107] [1,0,1,0,0,1,0,1,1,0,0,1,0,0,1,1],

[0108] [0,0,0,1,1,0,1,1,1,0,0,0,0,1,1,0],

[0109] [0,0,1,1,1,1,1,0,0,0,0,1,1,1,0,1],

[0110] [1,0,0,1,0,1,0,0,0,0,1,0,0,0,0,0]]. Due to the influence of noise, multipath fading and / or other factors in the wireless channel, the fixed-point number obtained after demodulation has 4 different bits from the fixed-point number generated by the binarizer 213 in the transmitter 210. Then, the debinarizer 222 debinarizes the fixed-point number to obtain the channel dimension vector [-0.0376, 0.5025, 0.8627, -0.9352, -0.7065,

[0111] 0.2150, 0.4853, -7.8044]. Finally, the decoder 223 decodes the channel dimension vector to generate an integer tag sequence [7, 118, 41, 122, 110, 18, 2]. Since the designs of the transmitter 210 and the receiver 220 disclosed in the present invention have partial fault tolerance, even if the channel dimension vector in the receiver 220 is not completely the same as the channel dimension vector in the transmitter 210, the decoder 223 can still decode the same integer tag sequence as the transmitter 210. Since different integer tag indexes correspond to different semantic descriptions, the control message obtained after the decoding process of the decoder 223 is the same as the control message in the transmitter 210, that is, the correct message transmission is completed.

[0112] The above-mentioned transmitter 210 and receiver 220 can be applied to base stations or access points and user equipment in wireless network systems (such as cellular networks, local area networks, etc. or other suitable wireless network systems) to transmit control messages. Taking the 5G new radio access network of the OpenAirInterface (OAI) Alliance as an example, the encoder 211, normalizer 212 and binarizer 213 in the transmitter 210 can be embedded between the media access control layer and the physical layer of the control plane in the new radio protocol stack (Protocol Stack) of the OAI next-generation base station, and the debinarizer 222 and decoder 223 in the receiver 220 can be embedded between the media access control layer and the physical layer in the new radio protocol stack of the OAI next-generation user equipment. In the control message transmission performed in the channel environment where the set channel dimension, the bit length of the fixed-point number, and the number of semantic fields are 8, 16, and 12 respectively, and the bit error rate is about 3%, the control message transmission success rate between the transmitter 210 and the receiver 220 can reach 100%. Compared to the conventional transmission method of 1 / 3 aggregation level and coding rate required to achieve 100% control message transmission success rate under the same channel environment, the control message transmission between the transmitter 210 and the receiver 220 disclosed in the present invention can significantly reduce the number of data bits required. If the control message encoding is performed by merging low-bit semantic fields (taking Table 1 as an example, 3 1-bit fields, 3 2-bit fields and 2 3-bit fields are merged into 1 combined semantic field), a 100% control message transmission success rate can be achieved under similar channel environments, and the number of data bits required for control message transmission can be further reduced. In addition, if 2 K The control message is encoded by using a semantic sequence. Taking Table 1 as an example, the maximum value of the number of bits K of the semantic field is 5. Under the condition that the bit length of the fixed-point number and the number of semantic fields are both fixed values, since the reference data set is small (i.e., the total number of semantic sequences included is small, i.e., 2 K Therefore, compared with the above two methods, this method can further reduce the channel dimension, reduce the number of data bits required for message transmission, and increase the success rate of control message transmission.

[0113] It should be noted that the above model training can be performed by the encoder of any transmitting end and the decoder of any receiving end, and is not limited to the encoder of the transmitting end and the decoder of the receiving end that use the same model training for control message transmission. In other words, the transmitting end and the receiving end (e.g. Figure 2 The transmission end 110 and the receiving end 120 for controlling the transmission of the message may be different from the transmission end and the receiving end for controlling the transmission of the message (for example Figure 4A The conveyor 210 and Figure 4A receiver 220).

[0114] Figure 6 4 is a flow chart of a wireless communication method 40 according to an embodiment of the present disclosure. The wireless communication method 40 is applicable to a transmitting end of a wireless communication system, such as Figure 1 A base station BS or a user equipment UE in a wireless communication system 10, Figure 4A The process of the wireless communication method 40 can be compiled into a program for execution, and includes the following steps: Step S41: Encode the control message into a channel dimension vector using a training model and according to the channel dimension and the number of semantic fields. The control message may include control plane media access control layer information, control plane physical layer information and / or other control plane information. Step S42: Normalize the channel dimension vector to generate a normalized channel dimension vector. Step S43: Binarize the normalized channel dimension vector to generate a fixed-point number. Step S44: Modulate the fixed-point number into a radio frequency signal and send the radio frequency signal. For example, if the wireless communication method 40 is used for Figure 4A If the transmitter 210 is a transmitter 210, steps S41-S44 may be respectively performed by the encoder 211, the normalizer 212, the binarizer 213 and the radio frequency circuit 214 in the transmitter 210. The detailed description of each step S41-S44 in the wireless communication method 40 may refer to the aforementioned description of the encoder 211, the normalizer 212, the binarizer 213 and the radio frequency circuit 214, which will not be repeated here.

[0115] As can be seen from the above description, the disclosed embodiment adopts the error-tolerant coding and decoding technology based on semantic communication to provide the ability to resist channel transmission errors for transmissions performed in various wireless channel environments. In particular, compared with the existing wireless communication system, the disclosed embodiment adopts a control message transmission mechanism based on semantic communication instead of a retransmission mechanism. In an environment with a high bit error rate, the existing wireless communication system may need to retransmit the control message multiple times to correctly decode the control message, but the control message transmission based on the disclosed embodiment can still extract semantic features through semantic decoding, and then correctly restore the control message without retransmission, so the efficiency of transmission resource utilization can be improved.

[0116] Figure 7 FIG. 3 is a block diagram of a device 300 according to an embodiment of the present disclosure. Figure 1 Each user equipment UE and / or base station BS, Figure 2 The transmitting end 110 and / or the receiving end 120, Figure 4A The transmission machine 210, and / or Figure 4BThe receiver 220 may have a block diagram similar to that of the device 300. The device 300 includes a processor 310, a memory 320, and a transceiver 330. The processor 310 may be, for example, a conventional processor (Conventional Processor), a digital signal processor (Digital Signal Processor, DSP), a microprocessor or an application-specific integrated circuit (Application-Specific Integrated Circuit, ASIC), but is not limited thereto. The memory 320 may be any data storage device that can be read and executed by the processor 310. The memory 320 may be, for example, a Subscriber Identity Module (SIM), a Read-Only Memory (ROM), an Erasable Programmable ROM (EPROM), an Electronically Erasable Programmable ROM (EEPROM), a Random Access Memory (Random Access Memory, RAM), a CD-ROM (CD-ROM), a hard disk, a solid state drive, a flash memory, or other data storage device suitable for storing program code, but is not limited thereto. The transceiver 330 may be a radio frequency circuit for wirelessly communicating with other entities based on the operation result of the processor 310 .

[0117] The processor 310 may be used to perform operations such as generation of training models and encoding / decoding of control messages disclosed herein. For example, if the device 300 is Figure 4A The transmitter 210 of FIG. 1 is a transmitter 210, and the transceiver 330 corresponds to the RF circuit 214 to communicate with the receiving end (eg Figure 4B The receiver 220 of the transmitter 210) performs wireless communication, and the processor 310 can be used to generate a training model for the transmitter 210 and / or perform operations performed by components such as the encoder 211, the normalizer 212 and the binarizer 213, and the memory 320 can be used to store programs for performing operations such as generating a training model, encoding, normalization and / or binarization.

[0118] Although the present disclosure has been disclosed as above by way of embodiments, it is not intended to limit the present disclosure. Any person having ordinary knowledge in the technical field may make slight changes and modifications without departing from the spirit and scope of the present disclosure. Therefore, the protection scope of the present disclosure shall be determined by the definition of the attached claims.

Claims

1. A transmitter for wireless communication, characterized in that: Include: An encoder configured to encode a control message into a channel dimension vector using a training model and according to a channel dimension and a semantic field number, wherein the control message includes at least one of a control plane medium access control layer information and a control plane physical layer information; a normalizer configured to normalize the channel dimension vector to generate a normalized channel dimension vector; a binarizer configured to binarize the normalized channel dimension vector to generate a certain number of points; and A radio frequency circuit is configured to modulate the fixed-point number into a first radio frequency signal and send the first radio frequency signal.

2. The conveyor according to claim 1, characterized in that The control message contains downlink control information.

3. The conveyor according to claim 1, characterized in that The training model includes a bidirectional long short-term memory neural network model.

4. The conveyor according to claim 1, characterized in that The encoder comprises a joint source-channel coding encoder.

5. The conveyor according to claim 1, characterized in that The fixed-point number includes a plurality of redundant sign bits, a plurality of redundant integer bits and a plurality of fractional bits.

6. The conveyor according to claim 1, characterized in that The encoder is configured to generate at least one of the channel dimension, a bit length of the fixed-point number, and the number of semantic fields according to a channel environment where the transmitter is located.

7. The conveyor according to claim 6, characterized in that Before the RF circuit sends the first RF signal to a receiving end, the encoder encodes a setting message including the channel dimension, the bit length of the fixed-point number and at least one of the semantic field number into a setting parameter data, and the RF circuit modulates the setting parameter data into a second RF signal and sends the second RF signal to the receiving end.

8. A receiver for wireless communication, characterized in that: Include: A radio frequency circuit configured to receive a first radio frequency signal and demodulate the first radio frequency signal to a certain number of points; a de-binarizer configured to de-binarize the fixed-point number to generate a channel dimension vector; and A decoder is configured to use a training model and decode the channel dimension vector into a control message according to a channel dimension and a semantic field number, wherein the control message includes a control plane media access control layer message or a control plane physical layer message.

9. The receiver according to claim 8, characterized in that The training model includes a long short-term memory neural network model.

10. A wireless communication method, used at a transmitting end, characterized in that: The wireless communication method comprises: Encoding a control message into a channel dimension vector using a training model and according to a channel dimension and a semantic field number, wherein the control message includes at least one of a control plane medium access control layer information and a control plane physical layer information; Normalizing the channel dimension vector to generate a normalized channel dimension vector; Binarizing the normalized channel dimension vector to generate a certain number of points; and The fixed-point number is modulated into a radio frequency signal and the radio frequency signal is sent.

Images