Model training methods, devices, signal transmission systems, equipment, and storage media

Abstract

This application relates to a model training method, apparatus, signal transmission system, device, and storage medium. The method includes: processing the discrete probability distribution vector of a training source signal to obtain a differentiable continuous probability vector; determining a target constellation point from the constellation diagram corresponding to the original multi-level orthogonal amplitude modulation signal constellation point vector; obtaining a decoding vector based on the signal of the target constellation point; training an initial probability shaping model based on the differentiable continuous probability vector and the decoding vector to obtain a target probability shaping model; and using the target probability shaping model to solve for the probability distribution of the source signal, thus achieving probability shaping. In this embodiment, the probability distribution obtained after solving for the source signal using the target probability shaping model can match different channels and adapt to a wider range of signal-to-noise ratios.

Description

Technical Field

[0001] This application relates to the field of optical communication coding and modulation technology, and in particular to a model training method, apparatus, signal transmission system, device and storage medium. Background Technology

[0002] In recent years, probabilistic shaping (PS) technology has attracted much attention in the optical transmission research community. PS is an optimization technique for coding modulation, often combined with higher-order orthogonal amplitude modulation formats. The main principle of PS is to map the outer constellation points of a constellation to points closer to the constellation origin with a certain probability and rule, and then transmit the data.

[0003] Currently, probability shaping is achieved through schemes such as probability amplitude shaping, parallel distribution matching, or multidimensional distribution matching.

[0004] However, current probabilistic shaping schemes have the problem that the probability distribution obtained after probabilistic shaping only fits a small signal-to-noise ratio range under a specific channel (Gaussian channel) model. Summary of the Invention

[0005] Therefore, it is necessary to provide a model training method, apparatus, signal transmission system, device, and storage medium that can adapt the probability distribution obtained after probability shaping to a large signal-to-noise ratio range under various channel models, in order to address the above-mentioned technical problems.

[0006] Firstly, this application provides a model training method. The method includes:

[0007] The discrete probability distribution vector of the training source signal is processed to obtain a differentiable continuous probability vector;

[0008] Based on the constellation point vector of the original multi-level quadrature amplitude modulation signal, the target constellation point is determined from the constellation diagram corresponding to the constellation point vector;

[0009] Based on the signals from the target constellation points, a decoding vector is obtained;

[0010] Based on the differentiable continuous probability vector and the decoding vector, an initial probability shaping model is trained to obtain a target probability shaping model.

[0011] In one embodiment, determining the target constellation point from the constellation diagram corresponding to the constellation point vector based on the original multi-level quadrature amplitude modulation (M-QAM) signal includes:

[0012] Calculate the product of the differentiable continuous probability vector and the constellation point vector to obtain the product result corresponding to each constellation point in the constellation diagram;

[0013] The target constellation point is determined from the constellation map based on the product results corresponding to each constellation point.

[0014] In one embodiment, determining the target constellation point from the constellation map based on the product results corresponding to each constellation point includes:

[0015] The constellation point corresponding to the product result that is not equal to zero is taken as the target constellation point.

[0016] In one embodiment, obtaining the decoding vector based on the signal from the target constellation point includes:

[0017] Based on the energy constraint conditions of the signals at the target constellation points, the energy of the signals at the target constellation points is constrained to obtain the first signal;

[0018] The first signal is transmitted through the channel to obtain the second signal;

[0019] The second signal is decoded to obtain the decoded vector.

[0020] In one embodiment, the method further includes:

[0021] The weights of each convolutional layer in the initial encoding module of the initial probability shaping model are initialized to obtain the encoding module;

[0022] The training source signal is encoded using the encoding module to obtain the discrete probability distribution vector, wherein the discrete probability distribution vector satisfies the Maxwell-Boltzmann distribution.

[0023] In one embodiment, processing the discrete probability distribution vector of the training source signal to obtain a differentiable continuous probability vector includes:

[0024] The discrete probability distribution vector is sampled using a sampler to obtain the sampling result;

[0025] The sampling results are quantized using an estimator to obtain the differentiable continuous probability vector.

[0026] In one embodiment, training an initial probability shaping model based on the differentiable continuous probability vector and the decoded vector to obtain a target probability shaping model includes:

[0027] Based on the differentiable continuous probability vector and the decoding vector, the initial probability shaping model is trained using the gradient descent method to obtain the target probability shaping model.

[0028] Secondly, this application also provides a model training apparatus. The apparatus includes:

[0029] The processing module is used to process the discrete probability distribution vector of the training source signal to obtain a differentiable continuous probability vector;

[0030] The determination module is used to determine the target constellation point from the constellation diagram corresponding to the constellation point vector based on the constellation point vector of the original multi-level quadrature amplitude modulation signal;

[0031] The first acquisition module is used to obtain a decoding vector based on the signal of the target constellation point;

[0032] The second acquisition module is used to train an initial probability shaping model based on the differentiable continuous probability vector and the decoding vector to obtain a target probability shaping model.

[0033] Thirdly, this application also provides a signal transmission system, which includes a transmitting device, an optical transmitting module, a wireless signal transmitting module, a signal receiving module, and a receiving device;

[0034] The transmitting device is configured to obtain a source signal, input the source signal into the target probability shaping model described in any of the above-mentioned items to obtain a probability shaping signal, obtain a signal to be transmitted based on the probability shaping signal, and transmit the signal to be transmitted to the optical transmitting module.

[0035] The optical transmitting module is used to output an optical signal according to the signal to be transmitted, and to input the optical signal to the wireless signal transmitting module;

[0036] The wireless signal transmitting module is used to send a wireless signal to the signal receiving module according to the optical signal;

[0037] The signal receiving module is used to receive the wireless signal and send the wireless signal to the receiving device;

[0038] The receiving device is used to demodulate the wireless signal to obtain the source signal.

[0039] Fourthly, this application also provides a computer device. The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to perform the following steps:

[0040] The discrete probability distribution vector of the training source signal is processed to obtain a differentiable continuous probability vector;

[0041] Based on the constellation point vector of the original multi-level quadrature amplitude modulation signal, the target constellation point is determined from the constellation diagram corresponding to the constellation point vector;

[0042] Based on the signals from the target constellation points, a decoding vector is obtained;

[0043] Based on the differentiable continuous probability vector and the decoding vector, an initial probability shaping model is trained to obtain a target probability shaping model.

[0044] Fifthly, this application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program thereon, which, when executed by a processor, performs the following steps:

[0045] The discrete probability distribution vector of the training source signal is processed to obtain a differentiable continuous probability vector;

[0046] Based on the constellation point vector of the original multi-level quadrature amplitude modulation signal, the target constellation point is determined from the constellation diagram corresponding to the constellation point vector;

[0047] Based on the signals from the target constellation points, a decoding vector is obtained;

[0048] Based on the differentiable continuous probability vector and the decoding vector, an initial probability shaping model is trained to obtain a target probability shaping model.

[0049] Sixthly, this application also provides a computer program product. The computer program product includes a computer program that, when executed by a processor, performs the following steps:

[0050] The discrete probability distribution vector of the training source signal is processed to obtain a differentiable continuous probability vector;

[0051] Based on the constellation point vector of the original multi-level quadrature amplitude modulation signal, the target constellation point is determined from the constellation diagram corresponding to the constellation point vector;

[0052] Based on the signals from the target constellation points, a decoding vector is obtained;

[0053] Based on the differentiable continuous probability vector and the decoding vector, an initial probability shaping model is trained to obtain a target probability shaping model.

[0054] The aforementioned model training method, apparatus, signal transmission system, device, and storage medium process the discrete probability distribution vector of the training source signal to obtain a differentiable continuous probability vector. Based on the constellation point vector of the original multi-level orthogonal amplitude modulation signal, target constellation points are determined from the constellation diagram corresponding to the constellation point vectors. Based on the signal of the target constellation point, a decoding vector is obtained. Then, based on the differentiable continuous probability vector and the decoding vector, an initial probability shaping model is trained to obtain a target probability shaping model. The probability distribution of the source signal is solved using the target probability shaping model to achieve probability shaping. In this embodiment, the probability distribution obtained after solving the source signal using the target probability shaping model can match different channels and adapt to a wider range of signal-to-noise ratios. Attached Figure Description

[0055] Figure 1 This is a schematic flowchart of a model training method provided in an embodiment of this application;

[0056] Figure 2 This is a flowchart illustrating a method for determining target constellation points provided in an embodiment of this application;

[0057] Figure 3 This is a flowchart illustrating a method for obtaining a decoding vector provided in an embodiment of this application;

[0058] Figure 4 This is a flowchart illustrating a method for obtaining a discrete probability distribution vector provided in an embodiment of this application;

[0059] Figure 5 This is a flowchart illustrating a method for obtaining a differentiable continuous probability vector provided in an embodiment of this application;

[0060] Figure 6 This is a schematic diagram of the structure of a communication system provided in an embodiment of this application;

[0061] Figure 7 This embodiment provides a three-dimensional constellation diagram of the probability distribution of MQAM signals.

[0062] Figure 8 This is an adaptively optimized MQAM constellation diagram under the AWGN channel provided in the embodiments of this application;

[0063] Figure 9 This is a performance comparison chart of the present invention's solution and other reference solutions under an AWGN channel;

[0064] Figure 10 This is a performance comparison chart of the present invention's solution and other reference solutions under Rayleigh fading channels;

[0065] Figure 11 This is a performance comparison chart of the present invention's solution and a reference solution in a nonlinear optical fiber channel for 64QAM.

[0066] Figure 12 This is a performance comparison chart of the present invention's solution and a reference solution in a nonlinear optical fiber channel for 256QAM.

[0067] Figure 13 This is a relative gain plot estimated by the NLIN model for 64 / 256QAM at optimal transmit power and transmission distance of 500km to 5000km;

[0068] Figure 14 This is the constellation diagram of 64 / 256QAM after optimization by the NLIN model at the optimal transmit power and transmission distance of 5000km;

[0069] Figure 15 This is a schematic diagram of the structure of a model training device provided in an embodiment of this application;

[0070] Figure 16 This application provides an experimental apparatus for a signal transmission system.

[0071] Figure 17 The spectrum of the local oscillator signal obtained after the optical signal passes through the optical coupler;

[0072] Figure 18 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0073] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0074] In one embodiment, such as Figure 1 As shown, Figure 1 This is a flowchart illustrating a model training method provided in an embodiment of this application. The method is applied to a computer device and includes the following steps:

[0075] S101. Process the discrete probability distribution vector of the training source signal to obtain a differentiable continuous probability vector.

[0076] The training source signal can be an M-dimensional one-hot vector S. M =[s1,s2,...,s M [ ], symbolic label vectors or Gray-mapped symbols. M-dimensional "one-hot" vectors can be input into an encoding module composed of multi-layer neural networks (NNs) for encoding. The encoding module is defined as a function The encoding module outputs an M-dimensional discrete probability distribution vector P. S ={p1,p2,…,p M}

[0077] S102. Based on the constellation point vector of the original multi-level quadrature amplitude modulation signal, determine the target constellation point from the constellation diagram corresponding to the constellation point vector.

[0078] In this context, the constellation point vector of a Multiple Quadrature Amplitude Modulation (MQAM) signal can be a normalized constellation point vector C = [c1, c2, ..., c...]. MMQAM includes 16QAM, 64QAM, and 256QAM.

[0079] S103. Based on the signal of the target constellation point, obtain the decoding vector.

[0080] The process involves transmitting the signal from the target constellation point through a channel to obtain the received signal. This received signal is then input to a decoding module, which decodes the signal to obtain the decoded vector. The channel can be an AWGN, Rayleigh fading channel, or NLIN channel, etc.

[0081] S104. Based on the differentiable continuous probability vector and the decoding vector, train the initial probability shaping model to obtain the target probability shaping model.

[0082] In this step, the initial probability shaping model can be input based on the differentiable continuous probability vector and the decoding vector, and the weights of the initial probability shaping model can be adjusted to obtain the target probability shaping model.

[0083] After obtaining the target probability shaping model, it can be applied to mobile terminals or base stations to solve the probability distribution of the source signal and achieve probability shaping.

[0084] The method provided in this embodiment processes the discrete probability distribution vector of the training source signal to obtain a differentiable continuous probability vector. Based on the constellation point vector of the original multi-level orthogonal amplitude modulation signal, the target constellation point is determined from the constellation diagram corresponding to the constellation point vector. Based on the signal of the target constellation point, a decoding vector is obtained. Then, based on the differentiable continuous probability vector and the decoding vector, an initial probability shaping model is trained to obtain a target probability shaping model. The probability distribution of the source signal is solved using the target probability shaping model to achieve probability shaping. In this embodiment, the probability distribution obtained after solving the source signal using the target probability shaping model can match different channels and is suitable for a wider range of signal-to-noise ratios.

[0085] Reference Figure 2 , Figure 2 This is a flowchart illustrating a method for determining target constellation points according to an embodiment of this application. This embodiment relates to an optional implementation of determining target constellation points from the constellation diagram corresponding to the original constellation point vectors of a multi-level quadrature amplitude modulation (MQAM) signal. Based on the above embodiment, the method includes the following steps:

[0086] S201. Calculate the product of the differentiable continuous probability vector and the constellation point vector to obtain the product result corresponding to each constellation point in the constellation diagram.

[0087] S202. Based on the product results corresponding to each constellation point, determine the target constellation point from the constellation map.

[0088] The target constellation point can be determined from the constellation map based on the product of each constellation point, which can be achieved in the following way:

[0089] The constellation point corresponding to the product result that is not equal to zero is taken as the target constellation point.

[0090] Reference Figure 3 , Figure 3 This is a flowchart illustrating a method for obtaining a decoding vector according to an embodiment of this application. This embodiment relates to an optional implementation of how to obtain a decoding vector based on the signal of a target constellation point. Based on the above embodiment, S103 includes the following steps:

[0091] S301. Based on the energy constraint condition of the signal of the target constellation point, constrain the energy of the signal of the target constellation point to obtain the first signal.

[0092] To ensure that the expected energies of the constellations are equal, energy constraints are imposed, which can be expressed by, for example, the following formula:

[0093] in, Let x represent the probability of the s-th constellation point. s Let x+yj represent the two-dimensional complex coordinates of the s-th constellation point, and S represent the constellation point matrix vector.

[0094] S302, The first signal is transmitted through the channel to obtain the second signal.

[0095] S303. Decode the second signal to obtain the decoded vector.

[0096] In this implementation, the energy of the signal at the target constellation point is constrained by the energy constraint condition of the signal at the target constellation point to obtain the first signal, thereby constraining the energy of the signal at the target constellation point and reducing the requirements on the hardware system.

[0097] In some embodiments, a method for obtaining discrete probability distribution vectors is also provided, referring to... Figure 4 , Figure 4 This is a flowchart illustrating a method for obtaining a discrete probability distribution vector according to an embodiment of this application. The method includes the following steps:

[0098] S401. Initialize the weights of each convolutional layer in the initial encoding module of the initial probabilistic shaping model to obtain the encoding module.

[0099] S402. Use the encoding module to encode the training source signal to obtain the discrete probability distribution vector.

[0100] Sampling different constellation points generates training source signals, and the discrete probability distribution vectors satisfy the Maxwell-Boltzmann distribution. The Maxwell-Boltzmann distribution is also known as the Maxwell-Boltzmann (MB) distribution, which is: P(x) is the shaping factor. i ) represents the probability of sending to the i-th constellation point, x i This represents the i-th constellation point.

[0101] Reference Figure 5 , Figure 5 This is a flowchart illustrating a method for obtaining a differentiable continuous probability vector according to an embodiment of this application. This embodiment relates to an optional implementation of processing the discrete probability distribution vector of a training source signal to obtain a differentiable continuous probability vector. Based on the above embodiment, S101 includes the following steps:

[0102] S501. Use a sampler to sample the discrete probability distribution vector to obtain the sampling result.

[0103] In this step, the Gumbel-Max sampler can be used to sample the discrete probability distribution vector to obtain the sampling results.

[0104] S502. Use the estimator to quantize the sampling results to obtain a differentiable continuous probability vector.

[0105] In this step, the Gumbel-Softmax estimator can be used to quantize the sampling results to obtain the probability of the i-th sample of a continuously differentiable function, i.e., to obtain a differentiable continuous probability vector. The estimator generates a differentiable continuous probability vector, which approximates a "one-hot" vector. A "one-hot" vector is a matrix vector containing 0 and 1. Based on the dot product of the "one-hot" vector and another vector, the constellation point corresponding to 1 can be selected as the transmission signal. For example, if the "one-hot" vector is vector a, and vector a = [0,1,0,0], then the dot product of vector a = [0,1,0,0] and vector b = [1,2,3,4] is a.*b = [2]. That is, the second coordinate point in vector b is selected as the transmission signal, i.e., the second coordinate point in vector b is the target constellation point.

[0106]

[0107] g i = -log(-log(u)), u ~ U(0,1)

[0108] in, Let g represent the probability of the i-th sample of a continuously differentiable function.i To introduce a continuously differentiable function, u represents a variable that follows a uniform distribution, and τ represents the temperature parameter (slack variable).

[0109] As can be seen from the above description of this embodiment, compared with the prior art, the model training method proposed in this application is a general shaping coding method based on end-to-end adaptive learning probability distribution. This method can obtain a larger shaping gain, while reducing the nonlinear effect of the signal, and realizes the probability distribution of optimal channel matching in a wider range of signal-to-noise ratios.

[0110] To provide a clearer explanation of this application, the following is combined with... Figure 6 This section introduces model training methods. (Refer to...) Figure 6 , Figure 6 This is a schematic diagram of a communication system provided in an embodiment of this application. The communication system includes an encoding module 601, a weight initialization module 602, a signal distribution quantization module 603, a signal modulation module 604, a link transmission module 605, a decoding module 606, a loss function constraint module 607, and an adaptive optimization update module 608.

[0111] In this embodiment, the training source signal can be a 16QAM, 64QAM, or 256QAM signal. The encoding module 601 takes the M-dimensional training source signal as the original input signal and inputs it into the encoding module 601, which is composed of a multi-layer neural network, for encoding. This is defined as a function. The output is an M-dimensional discrete probability distribution vector P S ={p1,p2,…,p M}

[0112] For 16QAM, λ = 0.13, H(x) = 3.66; for 64QAM, λ = 0.05, H(x) = 5.27; for 256QAM, λ = 0.01, H(x) = 7.47. The initialized probability distribution is as follows: Figure 7 As shown, Figure 7 This embodiment provides a three-dimensional constellation diagram of the probability distribution of an MQAM signal, where H(x) represents the information entropy. The weight initialization module 602 initializes the weights of each layer of the neural network in the encoding module 601, so that the discrete probability distribution vector output by the encoding module 601 satisfies the Maxwell-Boltzmann (MB) distribution.

[0113] The signal distribution quantization module 603 uses a Gumbel-Max sampler to quantize the discrete probability distribution and uses a Gumbel-Softmax estimator to obtain the probability of the i-th sample of the continuous differentiable function, that is, to obtain the differentiable continuous probability vector S.

[0114] The signal modulation module 604 modulates the signal based on the constellation point vector C = [c1, c2, ..., c2] obtained after normalization. M The dot product of the differentiable continuous probability vector S and the constellation point vector C is obtained. Based on this product, the target constellation point can be selected from the constellation diagram corresponding to the constellation point vector. To ensure that the expected energies of the constellation points are equal, an energy constraint condition is imposed on the target constellation point, as expressed by the following formula:

[0115]

[0116] In this embodiment, the link transmission module 605 can transmit the digital signal x of the target constellation point through an AWGN channel, a Rayleigh fading channel, or an analog nonlinear fiber channel (NLIN). The digital signal x can be a signal constrained by energy constraints or a signal from the target constellation point without energy constraints. The nonlinear fiber channel is a simulated 5-channel wavelength division multiplexing (WDM) system standard single-mode fiber (50 GHz channel spacing, 32 G symbols / s symbol rate, 20 spans × 100 km transmission distance, propagation attenuation of 0.2 dB / km, dispersion coefficient of 16.464 ps / nm / km, nonlinear coefficient of 1.3 Wkm⁻¹, power level scan range of -5 dBm to 5 dBm). The module obtains the received signal y (the received signal y is the second signal) and sends it to the decoding module 606 for processing. The decoding module 606 decodes the second signal to obtain the decoding vector r = [r1, r2, ..., rm].

[0117] The loss function constraint module 607 calculates the loss value based on the differentiable continuous probability vector and the decoding vector using the defined loss function, and adjusts the weights of the encoding module 601 and the decoding module 606 according to the loss value. The encoding module 601 re-encodes the M-dimensional training source signal based on the adjusted weights to obtain a new M-dimensional discrete probability distribution vector, and recalculates the loss value based on the new M-dimensional discrete probability distribution vector until the loss value no longer decreases, which means that the training has converged. At this time, the obtained generalized mutual information (GMI) is maximized. The estimated value of GMI is calculated as follows:

[0118]

[0119] c k,n For signal transmission bits, λ k,n The output of the neural network activation function can be used as an LLR estimate. To obtain the maximum GMI, the loss function is defined as follows:

[0120]

[0121] in, Represents a label, such as [0,0,1]. Represents information entropy; θ s θ represents the network weights of encoding module 601. D This represents the hyperparameters of the decoding module 606, where N represents the number of samples. This represents the label of the i-th position of the k-th symbol; the label can be 0 or 1. This represents the probability of the symbol x. Let X represent the probability of receiving the i-th bit of the k-th symbol, and let X represent the coordinate vector of the constellation points of QAM. E represents the probability vector obtained by the decoding module 606. x,y It expresses expectation.

[0122] The adaptive optimization update module 608 updates the weights W of the encoding module 601 in real time during backpropagation. n1 The weight W of the decoding module 606 n2 By minimizing the true posterior distribution and Learned receiver distribution The Kullback-Leibler (KL) distribution maximizes the system performance parameter GMI. Here, the true posterior distribution refers to the differentiable continuous probability vector S, and the receiver distribution refers to the decoded vector r mentioned above. The relative entropy, also known as Kullback-Leibler divergence or information divergence, is an asymmetric measure of the difference between two probability distributions.

[0123] This combination Figures 8 to 14 This section describes the system performance achievable through the embodiments of this application. (Refer to...) Figure 8 , Figure 8 This is an adaptively optimized MQAM constellation diagram under the AWGN channel provided in the embodiments of this application. Figure 8 The diagrams show the learned constellation diagrams for AWGN channels: (16QAM, @SNR=5dB), (64QAM, @SNR=13dB), and (256QAM, @SNR=18dB). SNR stands for Signal-to-noise ratio. The diagrams show that the probability distribution of low-energy symbols in the center is relatively high, while the probability distribution of high-energy symbols around the edges is relatively low, approximating a Gaussian distribution. Learning this probability distribution can reduce the impact of nonlinear effects and improve the system's transmission capacity.

[0124] Reference Figure 9 and Figure 10 , Figure 9 This is a performance comparison chart of the present invention's solution and other reference solutions under an AWGN channel. Figure 10 This is a performance comparison chart of the proposed solution and other reference solutions under Rayleigh fading channels. AWGN, short for Additive White Gaussian Noise, is the most basic noise and interference model. AOPDs Figure 9 and Figure 10 The figures compare the performance of the end-to-end adaptive optimization probability distribution scheme and other reference schemes under AWGN and Rayleigh fading channels, respectively. As can be seen from the figures, the probability distribution learned through training outperforms the reference scheme, matching the optimal probability distribution across the entire signal-to-noise ratio (SNR) range, and achieving shaping gains of 0.2 bits / 2D and 0.4 bits / 2D for 64 and 256QAM modulated signals, respectively. Here, AOPDs refers to Adaptive Optimization Probability Distributions.

[0125] Reference Figure 11 and Figure 12 , Figure 11 This is a performance comparison chart of the present invention's solution and a reference solution in a nonlinear optical fiber channel for 64QAM. Figure 12 This is a performance comparison chart of the present invention's solution and the reference solution in a nonlinear optical fiber channel with 256QAM. Figure 11 and Figure 12 The figures show a performance comparison between 64, 256QAM and traditional schemes under the nonlinear channel model of a 5-channel WDM system. As can be seen from the figures, the learned probability distribution is superior to the shaping gain of 0.15 bits / 2D and 0.16 bits / 2D of the reference scheme. Figure 11 The figure shows the relative gain estimated by the nonlinear regression (NLIN) model at optimal transmit power and transmission distances of 500km to 5000km (5–10 km span). As can be seen from the figure, the proposed scheme 64,256QAM can achieve shaping gains as high as 0.17 bits / 2D and 0.15 bits / 2D. Figure 12 These are the constellation points learned at the optimal transmit power and transmission distance of 2000km. It can be seen that the probability distribution of the constellation points approximates a Gaussian-like distribution.

[0126] Reference Figure 13 and Figure 14 , Figure 13 This is a relative gain plot estimated by the NLIN model for 64 / 256QAM at optimal transmit power and transmission distance of 500km to 5000km. Figure 14 This is the constellation diagram of 64 / 256QAM after optimization by the NLIN model at the optimal transmit power and transmission distance of 5000km. Figure 11 and Figure 12 The figures show a performance comparison of 64QAM, 256QAM, and traditional schemes under the nonlinear channel model of a 5-channel Wavelength Division Multiplexing (WDM) system. As can be seen from the figures, the learned probability distribution is superior to the shaping gain of 0.15 bits / 2D and 0.16 bits / 2D of the traditional scheme. Figure 13 The figure shows the relative gain estimated by the nonlinear regression (NLIN) model at optimal transmit power and transmission distances of 500km to 5000km (5–10 km range). As can be seen from the figure, the proposed 64QAM and 256QAM schemes can achieve shaping gains as high as 0.17 bits / 2D and 0.15 bits / 2D, respectively. Figure 14 These are the constellation points learned at the optimal transmit power and transmission distance of 2000km. It can be seen that the probability distribution of the constellation points approximates a Gaussian distribution.

[0127] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0128] Based on the same inventive concept, this application also provides a probability shaping device for implementing the model training method described above. The solution provided by this device is similar to the implementation described in the above method; therefore, the specific limitations in one or more probability shaping device embodiments provided below can be found in the limitations of the model training method described above, and will not be repeated here.

[0129] In one embodiment, such as Figure 15 As shown, Figure 15 This is a schematic diagram of a model training device provided in an embodiment of this application. The device 1500 includes:

[0130] The processing module 1501 is used to process the discrete probability distribution vector of the training source signal to obtain a differentiable continuous probability vector.

[0131] The determination module 1502 is used to determine the target constellation point from the constellation diagram corresponding to the constellation point vector based on the constellation point vector of the original multi-level quadrature amplitude modulation signal;

[0132] Decoding module 1503 is used to obtain a decoding vector based on the signal of the target constellation point;

[0133] Training module 1504 is used to train an initial probability shaping model based on the differentiable continuous probability vector and the decoding vector to obtain a target probability shaping model.

[0134] In some embodiments, the determining module 1502 includes:

[0135] The calculation unit is used to calculate the product of the differentiable continuous probability vector and the constellation point vector to obtain the product result corresponding to each constellation point in the constellation diagram;

[0136] The determining unit is used to determine the target constellation point from the constellation map based on the product results corresponding to each constellation point.

[0137] In some embodiments, the determining unit is specifically used to take the constellation point corresponding to the product result that is not equal to zero as the target constellation point.

[0138] In some embodiments, the decoding module 1503 is specifically used to constrain the energy of the signal of the target constellation point based on the energy constraint condition of the signal of the target constellation point to obtain a first signal; transmit the first signal through a channel to obtain a second signal; and perform decoding processing on the second signal to obtain the decoding vector.

[0139] In some embodiments, the apparatus may further include:

[0140] The weight initialization module is used to initialize the weights of each convolutional layer in the initial encoding module of the initial probability shaping model to obtain the encoding module;

[0141] The encoding module is used to encode the training source signal to obtain the discrete probability distribution vector, wherein the discrete probability distribution vector satisfies the Maxwell-Boltzmann distribution.

[0142] In some embodiments, the processing module 1501 is specifically used to sample the discrete probability distribution vector using a sampler to obtain a sampling result; and to quantize the sampling result using an estimator to obtain the differentiable continuous probability vector.

[0143] In some embodiments, the training module 1504 is specifically used to train the initial probability shaping model using gradient descent based on the differentiable continuous probability vector and the decoding vector, to obtain the target probability shaping model.

[0144] Each module in the aforementioned model training device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0145] It should be noted that processing the discrete probability distribution vector of the training source signal to obtain a differentiable continuous probability vector can be implemented by the aforementioned processing module 1501, or by... Figure 6 The signal distribution quantization module 603 implements this. Based on the constellation point vector of the original multi-level quadrature amplitude modulation signal, determining the target constellation point from the constellation diagram corresponding to the constellation point vector can be implemented by the determination module 1502, or by... Figure 6 The signal modulation module 604 in the middle implements this. Based on the signal of the target constellation points, the decoding vector can be obtained by the decoding module 1503, or by... Figure 6 The decoding module 606 in the middle is implemented. Based on the differentiable continuous probability vector and the decoding vector, the initial probability shaping model is trained to obtain the target probability shaping model. This can be implemented by the training module 1504, or by... Figure 6 The loss function constraint module 607 and the adaptive optimization update module 608 are implemented in the middle.

[0146] Based on the above target probability shaping model, the source signal can be probabilistically shaped to obtain a probabilistically shaped signal, which is then processed as follows: Figure 16 The signal transmission system shown processes the probability-shaping signal. (Refer to...) Figure 16 , Figure 16 This application provides an experimental apparatus for a signal transmission system. The signal transmission system includes a transmitting device 1601, an optical transmitting module 1602, a wireless signal transmitting module 1603, a signal receiving module 1604, and a receiving device 1605.

[0147] The transmitting device 1601 is used to obtain a source signal, input the source signal into the target probability shaping model of any of the above to obtain a probability shaped signal, obtain the signal to be transmitted based on the probability shaped signal, and transmit the signal to be transmitted to the optical transmitting module.

[0148] The optical transmitting module 1602 is used to process the signal to be transmitted using photon extrapolation beat frequency to obtain a wireless signal, and input the wireless signal to the wireless signal transmitting module;

[0149] The wireless signal transmitting module 1603 is used to amplify the wireless signal and send the amplified wireless signal to the signal receiving module.

[0150] Signal receiving module 1604 is used to receive the amplified wireless signal and send the amplified wireless signal to the receiving device;

[0151] The receiving device 1605 is used to demodulate the amplified wireless signal to obtain the source signal.

[0152] Among them, AWG: Arbitrary Waveform Generator; EA: Electrical Amplifier; ECL: External Cavity Laser; I / Q MOD: Modulator; OC: Optical Coupler; DL: Direct Link; VOA: Tunable Optical Attenuator; PBC: Polarization-Optical Coupler; SSMF: Standard Single-Mode Fiber; EDFA: Erbium-Doped Amplifier; TOF: Tunable Resonant Optical Module; PC: Polarization Controller; AIPM: Antenna Integrated Optical Module; HA: Horn Antenna; LNA: Low Noise Amplifier; DSO: Digital Oscilloscope.

[0153] Figure 16 A demonstration of photon-assisted terahertz optical-wireless fusion transmission experiment was provided, verifying the effectiveness of the proposed end-to-end adaptive optimization probabilistic shaping coding scheme. The implementation process is as follows: The optimized probabilistically shaped signal is input to the optical transmitter module 1602, where it is processed by an arbitrary waveform generator with a sampling rate of 92 Gsa / s to generate a baseband electrical signal. Parallel amplifiers are used to amplify the I and Q signal components respectively. Then, a modulator modulates the baseband electrical signal onto an optical signal, which is transmitted through a 20km fiber optic link and sent to the wireless signal transmitter module 1603. The signal is then received by the signal receiver module 1604 via the wireless transmission link. Finally, the signal is captured by the receiver device 1605, where it undergoes demodulation to recover the original source signal and calculates performance parameters such as GMI and BER (Bit Error Ratio). The wireless signal can be either a terahertz signal or a millimeter-wave signal. The optimized probabilistically shaped signal is then processed by… Figure 16 The system shown, after processing, can improve the reliability of the recovered original source signal and reduce the bit error rate. The spectrum of the local oscillator signal obtained after the optical signal passes through the optical coupler (OC) can be referenced... Figure 17 As shown.

[0154] It should be noted that the above Figure 16The transmitting device 1601, after obtaining the probability-shaped signal using the target probability shaping model, can perform upsampling and filtering on the probability-shaped signal to obtain the signal to be transmitted, and then send the signal to be transmitted to the optical transmitting module 1602. The signal to be transmitted is an electrical signal; upsampling and filtering the probability-shaped signal reduces interference from the wireless channel, ensuring signal transmission security. EA is used to amplify the electrical signal, and the IQ modulator is used to modulate the electrical signal onto the optical signal. OC is used to separate or combine two polarized optical signals. PBC is used to combine two polarized optical signals into one or separate them. EDFA is used to amplify the signal. AIPM integrates a Uni-Traveling-Carrier Photodiode (UTC-PD) and an antenna for beat frequency and signal transmission.

[0155] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 18 As shown, Figure 18 This is an internal structural diagram of a computer device in one embodiment. The computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements a model training method.

[0156] Those skilled in the art will understand that Figure 18 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0157] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method provided in the above embodiments.

[0158] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the method provided in the above embodiments.

[0159] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps of the method provided in the above embodiments.

[0160] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0161] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0162] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0163] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A model training method, characterized in that, The method comprises: processing a discrete probability distribution vector of a training source signal to obtain a differentiable continuous probability vector; determining a target constellation point from a constellation diagram corresponding to a constellation point vector of an original multi-ary quadrature amplitude modulation signal based on the constellation point vector; obtaining a decoding vector based on a signal of the target constellation point; training an initial probability shaping model based on the differentiable continuous probability vector and the decoding vector to obtain a target probability shaping model; wherein the determining the target constellation point from the constellation diagram corresponding to the constellation point vector of the original multi-ary quadrature amplitude modulation signal comprises: calculating a product of the differentiable continuous probability vector and the constellation point vector to obtain a product result corresponding to each constellation point in the constellation diagram; determining the target constellation point from the constellation diagram according to the product result corresponding to each constellation point; the determining the target constellation point from the constellation diagram according to the product result corresponding to each constellation point comprises: taking a constellation point corresponding to a product result not equal to zero as the target constellation point.

2. The method of claim 1, wherein, the obtaining the decoding vector based on the signal of the target constellation point comprises: constraining an energy of the signal of the target constellation point based on an energy constraint condition of the signal of the target constellation point to obtain a first signal; obtaining a second signal after transmitting the first signal through a channel; decoding the second signal to obtain the decoding vector.

3. The method of claim 1, wherein, The method further comprises: initializing weights of each convolution layer in an initial encoding module in the initial probability shaping model to obtain an encoding module; encoding the training source signal by using the encoding module to obtain the discrete probability distribution vector, wherein the discrete probability distribution vector satisfies a Maxwell-Boltzmann distribution.

4. The method according to any one of claims 1 to 3, characterized in that, The processing the discrete probability distribution vector of the training source signal to obtain the differentiable continuous probability vector comprises: sampling the discrete probability distribution vector by using a sampler to obtain a sampling result; quantizing the sampling result by using an estimator to obtain the differentiable continuous probability vector.

5. The method according to any one of claims 1 to 3, characterized in that, The training the initial probability shaping model based on the differentiable continuous probability vector and the decoding vector to obtain the target probability shaping model comprises: training the initial probability shaping model by using a gradient descent method based on the differentiable continuous probability vector and the decoding vector to obtain the target probability shaping model.

6. A signal transmission system, characterized by The signal transmission system comprises a sending device, an optical emission module, a wireless signal emission module, a signal receiving module and a receiving device; The sending device is configured to obtain a source signal, input the source signal into the target probability shaping model in any one of claims 1 to 5 to obtain a probability shaping signal, obtain a to-be-sent signal according to the probability shaping signal, and send the to-be-sent signal to the optical emission module; The optical emission module is configured to process the to-be-sent signal by using a photonic extrapolation beat frequency to obtain a wireless signal, and input the wireless signal into the wireless signal emission module; The wireless signal emission module is configured to amplify the wireless signal, and send the amplified wireless signal to the signal receiving module; and The signal receiving module is configured to receive the amplified wireless signal, and output a signal obtained by decoding the amplified wireless signal to the receiving device. The signal receiving module is configured to receive the amplified wireless signal and send the amplified wireless signal to the receiving device. The receiving device is configured to perform demodulation processing on the amplified wireless signal to obtain the source signal.

7. A model training apparatus characterized by comprising: The apparatus comprises: The processing module is configured to process a discrete probability distribution vector of a training source signal to obtain a differentiable continuous probability vector; The determining module is configured to determine a target constellation point from a constellation diagram corresponding to a constellation point vector of an original multi-ary quadrature amplitude modulation signal; The first obtaining module is configured to obtain a decoding vector based on a signal of the target constellation point; The second obtaining module is configured to train an initial probability shaping model based on the differentiable continuous probability vector and the decoding vector to obtain a target probability shaping model; The method comprises: calculating a product of the differentiable continuous probability vector and the constellation point vector to obtain a product result corresponding to each constellation point in the constellation diagram; determining the target constellation point from the constellation diagram according to the product result corresponding to each constellation point; The method comprises: taking a constellation point corresponding to a product result not equal to zero as the target constellation point.

8. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that, The processor executes the computer program to implement the steps of the method of any one of claims 1 to 5.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method of any one of claims 1 to 5.

Images