An original AI communication system against power amplifier nonlinearity

Abstract

The application discloses an anti-power amplifier nonlinearity original AI communication system, which comprises a transmitting end, a channel and a receiving end, wherein the transmitting end comprises a trained learnable back-off mapper LBOM, which maps an input coded bit stream into a power back-off symbol stream; the LBOM performs joint optimization on constellation distribution and input power back-off parameters through training; the channel transmits the sending symbol distorted by the power amplifier nonlinearity; when the computing power of the receiving end is limited, the receiving end adopts a neural network demapper ND to demap the symbol and signal-to-noise ratio information output by the channel into a log likelihood ratio, and information bits are obtained through BP decoding; when the computing power of the receiving end is sufficient, the receiving end adopts an IDD module to receive the symbol and signal-to-noise ratio information output by the channel, and joint iterative decoding is realized through mutual transmission of extrinsic information between submodules to obtain information bits; the IDD module comprises a neural network demapper iDemapper submodule and a BP decoder iDecoder submodule.

Description

Technical Field

[0001] This invention belongs to the fields of radio communication, satellite communication, and deep space communication technology, and particularly relates to a native AI communication system that resists power amplifier nonlinearity. Background Technology

[0002] In bandwidth-constrained scenarios such as satellite communication, spacecraft telemetry and control links, and high-speed satellite-to-ground data transmission links, the application of higher-order modulation formats, such as 32QAM / APSK and 64QAM / APSK, can improve spectral efficiency. However, these traditional higher-order modulation formats are characterized by non-constant envelopes, resulting in high peak-to-average power ratios (PAPRs) and high sensitivity to power amplifier nonlinearity. In these scenarios, due to tight link budgets, power amplifiers typically need to operate in near-saturation regions. The introduced nonlinearity leads to constellation diagram distortion, significantly increasing the bit error rate. While higher power backoff can allow the power amplifier to operate in the linear region, it significantly reduces power efficiency, making it difficult to maintain the required transmit power under power-constrained conditions, resulting in insufficient link margin. Therefore, the conditions for enabling higher-order modulation in practical systems are extremely stringent, especially in satellite communication and spacecraft telemetry and control communications, leading to low spectral efficiency.

[0003] To address this issue, deep learning technology has offered new solutions in recent years. In the radio frequency domain, neural network-based digital predistortion (NN-DPD) fits the power amplifier nonlinearity and memory effect through deep networks, achieving higher accuracy than traditional polynomial models. However, its high computational complexity makes it difficult to meet the real-time requirements of satellite-borne systems. In the baseband domain, deep learning methods have been used for constellation geometry shaping and probabilistic shaping, optimizing nonlinear constellation distributions through end-to-end learning. However, it faces difficulties in joint optimization with structured channel coding such as LDPC / Turbo, and its generalization ability is insufficient, making it difficult to adapt to time-varying satellite channels. Summary of the Invention

[0004] To address the issue that high-order modulation requires high-power backoff to maintain power amplifier linearity, which results in limited transmit power and insufficient link margin, the present invention aims to overcome the aforementioned defects of the prior art and proposes a native AI communication system and method that resists power amplifier nonlinearity.

[0005] In view of this, the present invention provides a native AI communication system resistant to power amplifier nonlinearity, comprising a transmitter, a channel, and a receiver. The system is used when the computing power of the receiver is limited. The transmitter includes a trained learnable backoff mapper (LBOM) for mapping the input encoded bitstream to a power-back-off symbol stream; wherein, the LBOM is trained to jointly optimize the constellation distribution and the input power backoff parameters. The channel is used to transmit transmitted symbols after nonlinear distortion by the power amplifier; The receiver includes a neural network demapper (ND) for receiving the symbol and signal-to-noise ratio information output by the channel, performing demapping, and outputting the log-likelihood ratio for decoding.

[0006] As an improvement to the above system, the processing procedure of the learnable backtracking mapper (LBOM) includes: Register the constellation distribution and input power backoff value as parameters; Perform energy normalization on the constellation distribution; Select the corresponding constellation symbol based on the input bits; The selected symbol is power adjusted based on the learnable input power backoff value, using the following formula:

[0007] in, For the symbol stream after rollback, For the selected symbol, This is the input power backoff value. This represents the input power corresponding to the power amplifier's saturation point.

[0008] As an improvement to the above system, the neural network demapper ND takes as input two dimensions (I and Q) of symbol and signal-to-noise ratio information, and outputs as... m Bit log-likelihood ratio, where ND is a three-layer fully connected network, wherein, First layer: Input dimension is 3, output dimension is 128, activation function is ReLU; Second layer: Input dimension is 128, output dimension is 128, activation function is ReLU; The third layer has an input dimension of 128 and an output dimension of the number of modulation bits. m No activation function.

[0009] As an improvement to the aforementioned system, the system employs end-to-end training, jointly optimizing the transmitter constellation mapping and receiver demapping, with a loss function... for: , in, This is the function for calculating binary cross-entropy. , Indicates the independent variable. The LLR value is the codeword obtained by flattening the log-likelihood ratio of the ND output and regrouping it according to the LDPC encoding bit length n. The encoded bits are obtained by the transmitter through LDPC encoding of the input information bits.

[0010] On the other hand, this invention provides a native AI communication system resistant to power amplifier nonlinearity, comprising a transmitter, a channel, and a receiver. The system is used when the receiver has sufficient computing power, wherein... The transmitter includes a trained learnable backoff mapper (LBOM) for mapping the input encoded bitstream to a power-back-off symbol stream; wherein, the LBOM is trained to jointly optimize the constellation distribution and the input power backoff parameters. The channel is used to transmit transmitted symbols after nonlinear distortion by the power amplifier; The receiver includes an IDD module, which receives the symbol and signal-to-noise ratio information output by the channel. The sub-modules exchange external information to achieve joint iterative decoding, thereby obtaining information bits. The IDD module includes a neural network demapper sub-module and a BP decoder sub-module.

[0011] As an improvement to the above system, the processing procedure of the learnable backtracking mapper (LBOM) includes: Register the constellation distribution and input power backoff value as parameters; Perform energy normalization on the constellation distribution; Select the corresponding constellation symbol based on the input bits; The selected constellation symbol is power-adjusted based on the learnable input power backoff value, using the following formula:

[0012] in, For the symbol stream after rollback, For the selected symbol, This is the input power backoff value. This represents the input power corresponding to the power amplifier's saturation point.

[0013] As an improvement to the above system, the processing procedure of the IDD module includes:

[0014]

[0015]

[0016]

[0017]

[0018]

[0019] in, The received symbol represents two dimensions, I and Q. This indicates information related to the signal-to-noise ratio. This refers to the iDemapper submodule. This indicates the external information passed from the previous iDecoder submodule. This indicates the external information passed by the current-level iDemapper submodule. This indicates a backpropagation (BP) decoding with an internal iteration count of 1. i This indicates the current iteration number of the IDD module. Indicates the number of the previous iteration. This represents the output of the iDemapper submodule. This indicates the output of the iDecoder submodule.

[0020] As an improvement to the above system, the input of the iDemapper submodule includes symbol and signal-to-noise ratio information in both I and Q dimensions, as well as external information passed from the previous iDecoder submodule, and the output is... m Bit log-likelihood ratio; the iDemapper submodule is a three-layer fully connected network, wherein... First layer: Input dimension is 3+ m The output dimension is 128, and the activation function is ReLU; where, m The number of modulation bits; Second layer: Input dimension is 128, output dimension is 128, activation function is ReLU; The third layer has an input dimension of 128 and an output dimension of the number of modulation bits. m No activation function.

[0021] As an improvement to the aforementioned system, the system employs end-to-end training, jointly optimizing the transmitter constellation mapping and receiver demapping, with a loss function... for:

[0022] in, This is the function for calculating binary cross-entropy. , Indicates the independent variable. The encoded bits are obtained by the transmitter through LDPC encoding of the input information bits. The bit length for LDPC encoding.

[0023] As an improvement to the above system, the transmitter further includes: an LDPC encoding module deployed before the Learnable Backoff Mapper (LBOM) and a TWTA power amplifier module deployed after the LBOM. The TWTA power amplifier module uses a Saleh model for nonlinear modeling, and its amplitude... and phase Satisfy the following formula:

[0024]

[0025] in, , , , All are parameters. , These represent the signal amplitude and phase input to the TWTA power amplifier module, respectively.

[0026] Compared with the prior art, the advantages of the present invention are: This invention constructs two deep learning-based end-to-end communication models. The core idea is to replace key modules in traditional communication systems with learnable neural network components, and to jointly optimize the transmitter constellation mapping and receiver demapping through end-to-end training. Both models overcome the limitations of traditional modular design, achieving end-to-end joint optimization, significantly improving communication performance under nonlinear channels, and enhancing the power amplifier efficiency and bandwidth efficiency of existing communication systems. Attached Figure Description

[0027] Figure 1 This is a structural diagram of the LBOM module in Example 1; Figure 2 This is a structural diagram of the ND module in Example 1; Figure 3 This is the block diagram of the LBOM-ND model in Example 1. Figure 4 This is a structural diagram of the IDD module in Example 2; Figure 5 This is a structural diagram of the iDemapper module in Example 2; Figure 6 This is a block diagram of the LBOM-IDD model structure in Example 2; Figure 7 This is a comparison chart of BER (Best Per Hour) values. Figure 8 This is a comparison chart of TD. Detailed Implementation

[0028] This invention proposes a native AI communication system resistant to power amplifier nonlinearity, comprising two neural network communication models: a learnable backplane mapping-neural demapping model (LBOM-ND) and a learnable backplane mapping-iterative decoding model (LBOM-IDD). By replacing key modules in traditional communication systems with learnable neural network components and jointly optimizing the transmitter, channel, and receiver through end-to-end training, a deep neural network-based communication system is formed. This significantly reduces the impact of power amplifier nonlinearity on high-order modulation in existing communication systems, improving effective transmit power and link margin while ensuring signal transmission quality.

[0029] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0030] This invention presents two deep learning communication models: LBOM-ND and LBOM-IDD. The LBOM-ND model introduces a Learnable Back-Off Mapper (LBOM) at the transmitter end, jointly optimizing the constellation diagram and the power amplifier input back-off (IBO) parameters. At the receiver end, a Neural Demapper (ND) replaces the traditional demapper, performing end-to-end training by minimizing the cross-entropy loss between the transmitted bits and the received LLR, effectively suppressing the degrading effect of power amplifier nonlinearity on higher-order modulation. The LBOM-IDD model draws inspiration from the traditional Bit-Interleaved Coded Modulation with Iterative Decoding (BICM-ID) system, further fusing the neural demapper and the confidence propagation (BP) decoder into an Iterative Demapping Decoder (IDD), achieving iterative optimization of joint demapping and decoding through external information exchange.

[0031] The difference lies in whether the receiver performs joint iteration of soft demodulation and decoding. When the receiver's computing power is limited, the LBOM-ND model is used without joint iteration; when the receiver's computing power is sufficient and there is a need to further improve communication performance, the LBOM-IDD model can be used with joint iteration.

[0032] Example 1 Embodiment 1 of the present invention provides a native AI communication system that resists power amplifier nonlinearity. When the receiver's computing power is limited, the LBOM-ND model is adopted, and training and deployment are performed according to the following steps: 1.1 LBOM-ND Model Architecture The LBOM-ND model consists of six parts, such as Figure 3As shown. Transmitter: LDPC encoding, LBOM, and traveling wave tube amplifier (TWTA power amplifier) ​​modules; Channel: AWGN channel module; Receiver: ND and BP decoding modules. Among them, the LBOM and ND modules are learnable parametric modules.

[0033] 1.2 Detailed Implementation of the LBOM Module like Figure 1 As shown, the input to the LBOM module is an encoded bitstream. The output is the back-down mapped symbol stream. The module will distribute the constellations. and input power back-off Registration is used as a parameter to ensure that the average energy of the constellation data is one. Normalized constellation distribution is obtained by performing energy normalization. First, the encoded bitstream... Divide into groups, each m Grouped by position ( m (is the number of modulation bits); m A normalized constellation symbol is selected using a bit as the label, and then power back-off is performed on the selected symbol. (in TWTA power amplifier Saleh The input power corresponding to the model saturation point. , , , for Saleh (Model parameters), and the backtracked symbolic stream can be obtained. .

[0034] During training, normalized constellation distribution and input power back-off Joint optimization will be performed, and the transmitter will learn the most suitable normalized constellation distribution and the corresponding optimal input power back-off for the nonlinear power amplifier.

[0035] 1.3 Detailed Implementation of the ND Module like Figure 2 As shown, the input to the ND module is the received symbol y (containing two real numbers in I and Q dimensions) and signal-to-noise ratio related information. The output is the m-bit log-likelihood ratio. . PEbN0 Defined as:

[0036]

[0037]

[0038]

[0039] in m The number of modulation bits; r For encoding bitrate; The noise power spectral density; This represents the average output power of the power amplifier. OBO represents the saturated output power of the power amplifier; OBO represents the output power back-off of the power amplifier. The saturated output energy of a TWTA amplifier within one symbol time is equal to the saturated output power of the amplifier. Multiplication sign time Due to simulation training It is usually set to 1, so in terms of value:

[0040] The ND module consists of a three-layer fully connected network: the network structure is as follows:

[0041] 1.4 LBOM-ND Model Training During training, a set of encoding, modulation, and power amplifier parameters are selected. The encoding can be any LDPC code, the modulation can be of any order, and the power amplifier can be a TWTA power amplifier with any parameters. Saleh Model or other types of power amplifiers Saleh Model. This example uses WIFI standard 1 / 2 code rate 1296 code length LDPC code (information bit length k=648, encoding bit length n=1296), 64th order modulation (m=6, M=64), parameters are as follows. , , , TWTA amplifier Saleh The model, AWGN channel, and BP decoding internal iteration count are set to 40.

[0042] The model input data consists of information bits in frames. ; and in units of transmitted symbols , Random and uniform sampling is performed within the interval [a, b]. The interval [a, b] is selected based on a 4 dB range centered on the 1e-6 bit error rate of the model's BER-PEbN0 curve. This range can be determined after training based on the model... BER- PEbN0 The curve is adjusted to adjust the sampling range; in this example, the range [8, 12] is selected. During training, the internal calculation process of the model is as follows: Sending end: Input information bit b is obtained through LDPC encoding. flatten it again The back-mapped symbol stream is obtained by performing back-mapping through the LBOM module. The transmitted symbol stream is then distorted by a power amplifier. (Note that the selection of frame number B needs to be flexible to meet the requirements.) (For integers), the specific implementation of the power amplifier module is as follows: First, calculate the amplitude and phase of the input symbol. , According to the amplitude of the input signal Calculate the amplitude of the output symbol and phase The transmitted symbol stream can then be obtained. x The formula is:

[0043]

[0044]

[0045] Channel: Transmit Symbol x The received symbol is obtained after adding noise through the AWGN channel. The formula is: ,in The power spectral density of Gaussian white noise can be expressed by: Calculated; This represents a standard Gaussian distribution with a mean of 0 and a variance of 1.

[0046] Receiver: Received symbols y and the corresponding The value is fed into the ND module to calculate the m-bit LLR value. After flattening and regrouping according to n, the LLR values ​​of the codewords can be obtained. ; After decoding, information bit extraction, and decision-making by the BP decoding module, the received bits are obtained in units of frames. .

[0047] The model loss function is defined as follows: ,in This is the function for calculating binary cross-entropy. .

[0048] Before training, set the constellation parameters of the LBOM module. Initialize to 64QAM to accelerate model convergence and improve model training stability.

[0049] The Adam optimizer was used during training; the training lasted for 200 epochs, with the learning rate gradually decreasing from 1e-3 to 1e-5; the training sample size was 10,000 frames, and the training batch size was 100 frames; the test sample size was 20,000 frames, and the test batch size was 800 frames; after each epoch, the model was tested, and the model parameters with the smallest loss on the test set were saved. Multiple training iterations were performed, and the optimal result was selected.

[0050] 1.5 LBOM-ND Model Deployment Once the LBOM-ND model is trained, the normalized constellation distribution S and the optimal input power backoff (IBO) learned by the model are embedded into the transmitter, and the ND module learned by the model is used in the receiver for demapping to complete the application of the LBOM-ND model.

[0051] Example 2 Embodiment 2 of the present invention provides a native AI communication system that resists power amplifier nonlinearity. When the receiver has sufficient computing power, the LBOM-IDD model is used, and training and deployment are performed according to the following steps: 2.1 LBOM-IDD Model Architecture The LBOM-IDD model consists of five parts, such as Figure 6 As shown. Transmitter: LDPC encoding, LBOM, and TWTA power amplifier modules; Channel: AWGN channel module; Receiver: IDD module. The internal structure of the IID module is as follows. Figure 4 As shown, iDemapper is as follows Figure 5 The neural network demapping module shown is an iDecoder, which is a backpropagation decoder with one iteration. The two modules exchange extrinsic information to achieve iterative decoding. The LBOM module and the iDemapper submodule are learnable parametric modules.

[0052] 2.2 Detailed Implementation of the LBOM Module like Figure 1 As shown, the input to the LBOM module is an encoded bitstream. The output is the back-down mapped symbol stream. The module will distribute the constellations. and input power back-off Registration is used as a parameter to ensure that the average energy of the constellation data is one. Normalized constellation distribution is obtained by performing energy normalization. First, the encoded bitstream... Divide into groups, each m Grouped by position ( m (is the number of modulation bits); mA normalized constellation symbol is selected using a bit as the label, and then power back-off is performed on the selected symbol. (in TWTA power amplifier Saleh The input power corresponding to the model saturation point. , , , for Saleh (Model parameters), and the backtracked symbolic stream can be obtained. .

[0053] During training, normalized constellation distribution and input power back-off Joint optimization will be performed, and the transmitter will learn the most suitable normalized constellation distribution and the corresponding optimal input power back-off for the nonlinear power amplifier.

[0054] 2.3 Detailed Implementation of the iDemapper Submodule like Figure 5 As shown, the iDemapper submodule and the ND module ( Figure 2 (As shown) are all neural network demappers, the difference being that the iDemapper module has an additional set of ( m External information input The module's total input is the received symbol y (containing two-dimensional real numbers, I and Q). and external information passed from the previous iDecoder. The output is the m-bit log-likelihood ratio. .

[0055] The iDemapper module consists of a three-layer fully connected network: the network structure is as follows:

[0056] 2.4 Detailed Implementation of the IDD Module like Figure 4 As shown, the IDD module consists of two sub-modules, iDemapper and iDecoder, which exchange external information to achieve joint iterative decoding. The formula is:

[0057]

[0058]

[0059]

[0060]

[0061]

[0062] in This indicates a backpropagation (BP) decoding with an internal iteration count of 1. i This indicates the current iteration number of the IDD module.

[0063] 2.5 LBOM-IDD Model Training During training, a set of encoding, modulation, and power amplifier parameters are selected. The encoding can be any LDPC code, the modulation can be of any order, and the power amplifier can be a TWTA power amplifier with any parameters. Saleh Model or other types of power amplifiers Saleh Model. This example uses WIFI standard 1 / 2 code rate 1296 code length LDPC code (information bit length k=648, encoding bit length n=1296), 64th order modulation (m=6, M=64), parameters are as follows. , , , TWTA amplifier Saleh Model, AWGN channel.

[0064] The model input data consists of information bits in frames. ; and in units of transmitted symbols , Randomly and uniformly sample within the interval [a, b], the selection of which is based on the model. BER- The curve represents a 4dB range centered on the 1e-3 bit error rate, which can be determined after training based on the model. PEbN0 BER- The curve is adjusted to adjust the sampling range; in this example, the range [6, 10] is selected. During training, the internal calculation process of the model is as follows: Sending end: Input information bit b is obtained through LDPC encoding. flatten it again The back-mapped symbol stream is obtained by performing back-mapping through the LBOM module. The transmitted symbol stream is then distorted by a power amplifier. (Note that the selection of frame number B needs to be flexible to meet the requirements.) (For integers), the specific implementation of the power amplifier module is as follows: First, calculate the amplitude and phase of the input symbol. , According to the amplitude of the input signal Calculate the amplitude of the output symbol and phase The transmitted symbol stream can then be obtained. x The formula is:

[0065]

[0066]

[0067] Channel: Transmit Symbol x The received symbol is obtained after adding noise through the AWGN channel. The formula is: ,in The power spectral density of Gaussian white noise can be expressed by: Calculated; This represents a standard Gaussian distribution with a mean of 0 and a variance of 1.

[0068] Receiver: Received symbols y and the corresponding The value is fed into the IDD module, and after joint iterative decoding, the a posteriori LLR value is obtained. After removing the check bits and making a decision, the received bits in frames can be obtained. .

[0069] The model loss function is defined as follows:

[0070] in This is the function for calculating binary cross-entropy. .

[0071] Before training, the constellation parameters and IBO in the LBOM module are initialized to the optimal constellation parameters and optimal IBO obtained from the LBOM-ND model training; the iDemapper submodule parameters are initialized to the optimal ND module parameters obtained from the LBOM-ND model training, wherein the parameters from the input layer to the first hidden layer are padded with 0 from [3,128] to obtain [3+m,128].

[0072] The Adam optimizer was used during training; the training lasted for 200 epochs, with the learning rate gradually decreasing from 1e-3 to 1e-5; the training sample size was 10,000 frames, and the training batch size was 100 frames; the test sample size was 20,000 frames, and the test batch size was 800 frames; after each epoch, the model was tested, and the model parameters with the smallest loss on the test set were saved. Multiple training iterations were performed, and the optimal result was selected.

[0073] 2.6 LBOM-IDD Model Deployment Once the LBOM-IDD model is trained, the normalized constellation distribution S and the optimal input power backoff IBO learned by the model are simply embedded into the transmitter, and the iDemapper submodule learned by the model is used for demapping in a receiver that supports iterative decoding (BICM-ID receiver) to complete the application of the LBOM-IDD model.

[0074] Model performance analysis: Under WIFI 1 / 2 bitrate 1296 code length LDPC encoding, PEbN0 The BER curves for four communication models are presented (the horizontal axis PEbN0 reflects the performance of the communication system under the same noise level), where the gray line represents the BER performance baseline of a pure AWGN channel without TWTA power amplifier and 64QAM modulation. Figure 7 Figure 8 This paper demonstrates the total degradation (TD) of these four communication models compared to the gray baseline at a bit error rate (BER) level of 1e-5. As shown in the figure, LBOM-ND reduces the minimum TD by approximately 1.9 dB compared to 64QAM, and by approximately 0.6 dB compared to 64QAM with DPD. LBOM-IDD reduces the minimum TD by approximately 2.5 dB compared to 64QAM, and by approximately 1.3 dB compared to 64QAM with DPD. While maintaining communication quality, it simultaneously reduces output power back-off and signal distortion, and improves power amplifier efficiency and link margin.

[0075] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to the embodiments, those skilled in the art should understand that modifications or equivalent substitutions to the technical solutions of the present invention do not depart from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A native AI communication system resistant to power amplifier nonlinearity, comprising a transmitter, a channel, and a receiver, characterized in that, The system is used when the computing power of the receiving end is limited, wherein, The transmitter includes a trained learnable backoff mapper (LBOM) for mapping the input encoded bitstream to a power-back-off symbol stream; wherein, the LBOM is trained to jointly optimize the constellation distribution and the input power backoff parameters. The channel is used to transmit transmitted symbols after nonlinear distortion by the power amplifier; The receiver includes a neural network demapper (ND) for receiving the symbol and signal-to-noise ratio information output by the channel, performing demapping, and outputting the log-likelihood ratio for decoding.

2. The native AI communication system with anti-power amplifier nonlinearity according to claim 1, characterized in that, The processing procedure of the learnable backtracking mapper (LBOM) includes: Register the constellation distribution and input power backoff value as parameters; Perform energy normalization on the constellation distribution; Select the corresponding constellation symbol based on the input bits; The selected symbol is power adjusted based on the learnable input power backoff value, using the following formula: in, For the symbol stream after rollback, For the selected symbol, This is the input power backoff value. This represents the input power corresponding to the power amplifier's saturation point.

3. The native AI communication system with anti-power amplifier nonlinearity according to claim 1, characterized in that, The input to the neural network demapper ND is symbol and signal-to-noise ratio information in two dimensions (I and Q), and the output is... m Bit log-likelihood ratio, where ND is a three-layer fully connected network. in, First layer: Input dimension is 3, output dimension is 128, activation function is ReLU; Second layer: Input dimension is 128, output dimension is 128, activation function is ReLU; The third layer has an input dimension of 128 and an output dimension of the number of modulation bits. m No activation function.

4. The native AI communication system with anti-power amplifier nonlinearity according to claim 3, characterized in that, The system employs end-to-end training, jointly optimizing the transmitter constellation mapping and receiver demapping, with a loss function... for: ， in, This is the function for calculating binary cross-entropy. , Indicates the independent variable. The LLR value is the codeword obtained by flattening the log-likelihood ratio of the ND output and regrouping it according to the LDPC encoding bit length n. The encoded bits are obtained by the transmitter through LDPC encoding of the input information bits.

5. A native AI communication system resistant to power amplifier nonlinearity, comprising a transmitter, a channel, and a receiver, characterized in that, The system is used when the receiving end has sufficient computing power, wherein... The transmitter includes a trained learnable backoff mapper (LBOM) for mapping the input encoded bitstream to a power-back-off symbol stream; wherein, the LBOM is trained to jointly optimize the constellation distribution and the input power backoff parameters. The channel is used to transmit transmitted symbols after nonlinear distortion by the power amplifier; The receiver includes an IDD module, which receives the symbol and signal-to-noise ratio information output by the channel. The sub-modules exchange external information to achieve joint iterative decoding, thereby obtaining information bits. The IDD module includes a neural network demapper sub-module and a BP decoder sub-module.

6. The native AI communication system with anti-power amplifier nonlinearity according to claim 5, characterized in that, The processing procedure of the learnable backtracking mapper (LBOM) includes: Register the constellation distribution and input power backoff value as parameters; Perform energy normalization on the constellation distribution; Select the corresponding constellation symbol based on the input bits; The selected constellation symbol is power-adjusted based on the learnable input power backoff value, using the following formula: in, For the symbol stream after rollback, For the selected symbol, This is the input power backoff value. This represents the input power corresponding to the power amplifier's saturation point.

7. The native AI communication system with anti-power amplifier nonlinearity according to claim 5, characterized in that, The processing procedure of the IDD module includes: in, The received symbol represents two dimensions, I and Q. This indicates information related to the signal-to-noise ratio. This refers to the iDemapper submodule. This indicates the external information passed from the previous iDecoder submodule. This indicates the external information passed by the current-level iDemapper submodule. This indicates a backpropagation (BP) decoding with an internal iteration count of 1. i This indicates the current iteration number of the IDD module. Indicates the number of the previous iteration. This represents the output of the iDemapper submodule. This indicates the output of the iDecoder submodule.

8. The native AI communication system with anti-power amplifier nonlinearity according to claim 5, characterized in that, The input to the iDemapper submodule includes symbol and signal-to-noise ratio information in both I and Q dimensions, as well as external information passed from the previous iDecoder submodule. The output is... m Bit log-likelihood ratio; the iDemapper submodule is a three-layer fully connected network. in, First layer: Input dimension is 3+ m The output dimension is 128, and the activation function is ReLU; where, m The number of modulation bits; Second layer: Input dimension is 128, output dimension is 128, activation function is ReLU; The third layer has an input dimension of 128 and an output dimension of the number of modulation bits. m No activation function.

9. The native AI communication system with anti-power amplifier nonlinearity according to claim 8, characterized in that, The system employs end-to-end training, jointly optimizing the transmitter constellation mapping and receiver demapping, with a loss function... for: in, This is the function for calculating binary cross-entropy. , Indicates the independent variable. The encoded bits are obtained by the transmitter through LDPC encoding of the input information bits. The bit length for LDPC encoding.

10. The native AI communication system with anti-power amplifier nonlinearity according to claim 1 or claim 5, characterized in that, The transmitter also includes: an LDPC encoding module deployed before the Learnable Backoff Mapper (LBOM) and a TWTA power amplifier module deployed after the LBOM. The TWTA power amplifier module uses the Saleh model for nonlinear modeling, and its amplitude... and phase Satisfy the following formula: in, , , , All are parameters. , These represent the signal amplitude and phase input to the TWTA power amplifier module, respectively.

Images