Optical fiber terahertz communication system optimization method based on artificial intelligence

Abstract

The invention discloses an optical fiber terahertz communication system optimization method based on artificial intelligence, and belongs to the technical field of communication. According to the method, four-dimensional constellation shaping is constructed by utilizing additional dimensions provided by two independent sidebands, and an end-to-end optimization framework is divided into a training stage and a deployment stage; in the training stage, firstly, an equalizer based on a pruning neural network and an equalized channel model based on a mixed density network are trained in sequence through receiving and transmitting of a random training sequence; then connecting a constellation shaper at a transmitting end with a demapper at a receiving end through an equalized channel model with fixed parameters, thereby considering full-link residual distortion and noise after equalization, and carrying out back propagation gradient to learn a globally optimal four-dimensional joint geometric probability constellation diagram to further approach the Shannon limit; and in the deployment stage, transmission of the bit stream is realized through coordination of the four-dimensional geometric shaping constellation table and the probability distribution obtained in the training stage and an equalizer at the receiving end and a demapper.

Description

Technical Field

[0001] The present invention belongs to the field of communication technology, and in particular to an optimization method for an optical fiber terahertz communication system based on artificial intelligence. Background Art

[0002] The fiber-optic THz communication system realizes the seamless integration of fiber-optic networks and THz wireless links, and has larger spectrum resources and lower latency. In recent years, artificial intelligence technology has been continuously applied to other disciplines in its development process, and has achieved performance advantages that surpass traditional algorithms. At the same time, the large bandwidth advantage of the fiber-optic THz system enables it to quickly generate a large amount of data, which makes it very suitable for the application of various data-driven methods. An important reason why deep learning methods are of special significance in the field of communications is that they can jointly design the various modules of the entire communication system end-to-end, thereby approaching the global optimum, which is different from the traditional communication system design concept, which generally considers using different indicators to optimize different communication modules, thereby achieving multiple local optimality.

[0003] In recent years, some two-dimensional constellation shaping schemes have been adopted in wireless and fiber-optic communication systems, such as using autoencoders with mutual information as the loss function to achieve end-to-end joint geometric probability shaping over additive white Gaussian noise (AWGN) and fading channels [Stark, Maximilian, AitAoudia, and Jakob Hoydis. "Joint learning of geometric and probabilistic constellation shaping." 2019IEEE Globecom Workshops (GC Wkshps). IEEE, 2019.】, further considering the bit label, using generalized mutual information as the loss function, and avoiding sampling from the probability by uniformly sampling from the constellation points, we can achieve end-to-end joint geometric probabilistic constellation shaping for any channel model [Aoudia, Ait, and JakobHoydis. "Joint learning of probabilistic and geometric shaping for coded modulation systems." GLOBECOM 2020-2020IEEE Global Communications Conference. IEEE, 2020.】. However, these methods require the probability distribution of channel conditions to be known a priori and are not suitable for more complex channels, such as fiber terahertz channels.

[0004] Compared with two-dimensional geometric shaping, four-dimensional geometric shaping can obtain a larger minimum Euclidean distance between constellation points by sacrificing a small amount of spectral efficiency. In addition, since the four-dimensional space can provide a larger search range, a constellation diagram with better nonlinear tolerance can be designed, thereby achieving higher coding gain [Chen, Bin, et al. "Geometrically-shaped multi-dimensional modulation formats in coherent optical transmission systems." Journal of Lightwave Technology 41.3(2022):897-910.]. However, in order to achieve four-dimensional geometric shaping, existing optical fiber communication systems usually use two polarizations of light as the additional two dimensions, that is, optical polarization multiplexing and optical polarization diversity are required, which greatly increases the complexity of the system architecture. In addition, the double-sideband signal generated by intensity modulation based on the single-arm Mach-Zehnder modulator will introduce frequency selective fading under the influence of fiber dispersion. The use of independent sideband modulation can solve this problem well. Independent sideband modulation generates two different sidebands with independent information around the same carrier, which can naturally provide two additional dimensions for the four-dimensional constellation shaping scheme, and is very suitable for constructing four-dimensional constellation shaping. Summary of the invention

[0005] The purpose of the present invention is to overcome the shortcomings of the prior art and provide an artificial intelligence-based fiber terahertz communication system optimization method, which utilizes the extra dimension provided by two independent sidebands. The overall optimization framework adopts an end-to-end framework, considers the full link damage mechanism (including the loss, dispersion, nonlinear effect of the optical fiber link and the atmospheric loss and multipath interference of the terahertz link on the transmission signal), and adaptively designs a globally optimal four-dimensional joint geometric probability constellation shaping scheme to further approach the Shannon limit and achieve higher reliability and effectiveness of the communication system.

[0006] To achieve the above-mentioned object of the invention, the present invention provides an optimization method for an optical fiber terahertz communication system based on artificial intelligence, which includes a training phase and a deployment phase, and specifically includes the following steps:

[0007] Training phase steps:

[0008] Step a1, the transmitting end randomly generates a transmitting symbol sequence, and after passing through the full-link optical fiber terahertz channel, a receiving symbol sequence is obtained at the receiving end;

[0009] Using the transmitted symbol sequence and the received symbol sequence as training data, an equalizer based on pruned neural network is trained;

[0010] Step a2: training a post-equalization channel model based on a mixture density network (MDN) based on the transmitted symbol sequence and the post-equalization symbol sequence, so as to be used for the subsequent connection of the constellation shaper (i.e., the four-dimensional joint geometric probability constellation shaper) at the transmitting end and the demapper at the receiving end, so as to realize end-to-end gradient direction propagation considering the full link impairment mechanism;

[0011] Among them, the post-equalization channel model is used to output the probability distribution describing the post-equalization channel conditions The composition parameters of The fitting of

[0012] The symbol sequence after equalization is: the transmitted symbol sequence is input into the trained equalizer based on the pruned neural network, and the symbol sequence after equalization is obtained based on the output thereof;

[0013] Step a3, simultaneously training a constellation shaper at the transmitting end and a demapper at the receiving end, wherein both the constellation shaper and the demapper are network models based on a neural network;

[0014] The constellation shaper includes a geometric shaper and a probability shaper, which are used to generate a four-dimensional geometric and probability constellation table to shape the transmitting end bit sequence to obtain a four-dimensional frequency domain transmission symbol sequence x;

[0015] The four-dimensional frequency domain transmission symbol sequence x is input into the trained equalized channel model to obtain the four-dimensional frequency domain reception symbol sequence Then the four-dimensional frequency domain received symbol sequence Input demapper to obtain the corresponding bit sequence estimate

[0016] Deployment phase steps:

[0017] Step b1, the trained constellation shaper generates a constellation table for shaping the information bit sequence of the transmitting end to obtain a four-dimensional frequency domain transmission symbol sequence x;

[0018] Step b2, the four-dimensional frequency domain transmission symbol sequence x passes through the same full-link optical fiber terahertz channel as in step a1, and obtains a corresponding reception symbol sequence at the receiving end, and then sends the reception symbol sequence to a trained pruned neural network-based equalizer to obtain an equalized reception symbol sequence;

[0019] Step b3: After equalization, the received symbol sequence is demapped to obtain an estimate of the received information bit sequence.

[0020] The implementation process of an artificial intelligence-based optical fiber terahertz communication system optimization method provided by the present invention is as follows: four-dimensional constellation shaping is constructed by utilizing the extra dimensions provided by two independent sidebands, and the end-to-end optimization framework is divided into a training stage and a deployment stage; in the training stage, firstly, an equalizer based on a pruned neural network and a post-equalization channel model based on a mixed density network are trained in sequence by sending and receiving a random training sequence; then, the constellation shaper at the transmitting end and the demapper at the receiving end are connected through the post-equalization channel model with fixed parameters (trained), so as to consider the residual distortion and noise of the whole link after equalization, and back-propagate the gradient to learn the globally optimal four-dimensional joint geometric probability constellation diagram, so as to further approach the Shannon limit; in the deployment stage, the transmission of the bit stream is realized through the coordination of the four-dimensional geometric shaping constellation table and probability distribution learned in the training stage and the equalizer and demapper at the receiving end.

[0021] Furthermore, in step a1, the transmitting end randomly generates a transmission symbol sequence, and after passing through the full-link optical fiber terahertz channel, the receiving symbol sequence obtained at the receiving end includes:

[0022] (a1.1) The transmitter generates a random symbol sequence that follows a uniform probability distribution. After serial-to-parallel conversion, two independent data channels are obtained. They are modulated by orthogonal frequency division multiplexing (OFDM) and pulse-shaped and then up-converted to a center frequency of f s , and obtain two independent double-sideband OFDM signals;

[0023] (a1.2) The two double-sideband signals provide driving input for two Mach-Zehnder modulators (MZM) operating in push-pull mode to modulate a c The single-mode laser generates a continuous wavelength optical carrier to obtain two double-sideband optical signals. The left and right sidebands are extracted by two optical filters respectively, and the independent sideband (ISB) optical signal is obtained by an optical power combiner.

[0024] The ISB optical signal generated by (a1.3) is sent into a single-mode optical fiber for transmission. At the receiving end of the optical fiber link, it is amplified by an erbium-doped fiber amplifier and then mixed with a signal with an operating frequency of f. c+t The continuous wavelength optical carrier generated by the single-mode laser is combined. Based on the principle of optical heterodyne beat frequency, the combined signal is converted into a center frequency f by a single carrier photodiode (UTC-PD). t ISB terahertz signal;

[0025] (a1.4) The generated ISB terahertz signal is radiated into free space by the terahertz antenna. At the receiving end of the terahertz link, the local RF source is mixed with the received terahertz signal after N times of frequency multiplication, converting the ISB terahertz signal into an ISB baseband signal. Then, two equal-power ISB baseband signals are obtained through a power divider, and the independent left and right sidebands are separated through two filters.

[0026] (a1.5) The two independent left and right sidebands are respectively subjected to center frequency down-conversion, pulse shaping and OFDM demodulation to obtain the received symbol sequence.

[0027] Furthermore, the pruned neural network based equalizer may adopt a multi-layer perceptron model structure (ie including a multi-layer perceptron layer), and the number of neurons in the input layer of the pruned neural network equalizer is twice the number of subcarriers of the input transmitted symbol sequence.

[0028] Since intercarrier interference (ICI) is mainly concentrated on the adjacent subcarriers of the current subcarrier, the pruning technology is used to retain only the connections of the neurons corresponding to the subcarriers of the input layer adjacent to the neurons corresponding to the first hidden layer in the first layer of the model structure (this layer is called the corresponding pruning layer), and only point-by-point operations are performed in the subsequent network layers; based on the above pruning idea, the equalizer of the present invention adopts a dual-branch heterogeneous form, one of which is a linear branch, which only considers linear distortion, that is, does not contain an activation function, and the other is a nonlinear branch, which considers nonlinear distortion, that is, contains a nonlinear activation function. That is, the linear branch includes several layers of linear layers, and the nonlinear branch is several layers of linear layers with nonlinear activation functions, and the number of layers of the two branches is the same; at the same time, a self-attention mechanism is introduced, and a weight with a value range of [0,1] is learned from the input of the equalizer to weight the output of the linear branch, so that the neural network can better learn the linear and nonlinear distortion itself. That is, the input data of the equalizer is sent to the self-attention layer to learn the self-attention weight of the linear branch, and the output result of each subcarrier output by the linear branch is weighted, and then added to the output result of the corresponding subcarrier output by the nonlinear branch to obtain the output of the equalizer based on the pruned neural network;

[0029] Among them, each neuron of the first layer of the linear branch and the nonlinear branch is connected only with the neurons corresponding to the subcarrier of the input layer adjacent to the subcarrier represented by the current neuron. From the second layer onwards, the linear branch and the nonlinear branch only perform point-by-point calculation operations, that is, the neurons of adjacent layers are connected only with the neurons representing the same subcarrier.

[0030] Furthermore, the weight of each network layer of the linear branch and the nonlinear branch during operation is the weight W of the network layer and the corresponding mask matrix W mThe point product of , thereby achieving pruning of each network layer;

[0031] Among them, the mask matrix W of the first layer m Based on the number of subcarriers N sub and inter-subcarrier interference ICI length N ICI Determine, 0 means the subcarrier at the current position is not interfered, 1 means the subcarrier at the current position is interfered, and the number of elements 1 in each row except the first and last two rows is: 2×N ICI +2, the value 2 in this expression is because the real and imaginary parts are separated;

[0032] The linear branch and the nonlinear branch start from the second layer, and the mask matrix W during operation is m It is a block diagonal matrix to achieve the purpose of performing only point-by-point operations in subsequent network layers.

[0033] Furthermore, the pruned neural network based equalizer uses mean square error as the loss function during training.

[0034] Furthermore, each neuron in the first hidden layer of the equalized channel model is connected only to neurons corresponding to subcarriers in the input layer adjacent to the subcarrier represented by the current neuron, and starting from the second layer, neurons in adjacent layers are connected only to neurons representing the same subset of subcarriers.

[0035] Furthermore, the output of the equalized channel model is the probability distribution describing the equalized channel condition The latter includes the residual linear and nonlinear distortion and noise of the equalized channel. The equalized channel model learns multiple Gaussian distributions and their weighted combinations by transmitting symbols to achieve the conditional probability distribution of the equalized channel. The essence of the output of the equalized channel model is the probability distribution, so the negative log-likelihood function is selected as the loss function during training.

[0036] Further, in step a3, when the constellation shaper of the transmitter and the demapper of the receiver are trained simultaneously, the training algorithm framework of the configured constellation shaper includes a geometric shaper, a probability shaper, a symmetric quadrant constraint layer, a power normalization layer, a Softmax function layer and a sampler;

[0037] The geometry reshaper is used to transform the sequential one-hot symbol table into an unnormalized constellation table in the first quadrant of the four-dimensional space. The mapping of Represents the bit sequence corresponding to the symbol;

[0038] The probability shaper is used to convert an arbitrarily specified constant to the corresponding A set of unnormalized log probabilities The mapping of

[0039] Symmetric quadrant constraint layers are used to and Perform a mirror operation to obtain a complete four-dimensional unnormalized constellation table and the corresponding unnormalized log probability

[0040] The Softmax function layer is used to output the symmetric quadrant constraint layer. Perform the Softmax operation to obtain the probability distribution of the four-dimensional constellation diagram

[0041] The power normalization layer is based on probability distribution Output of the symmetric quadrant constraint layer Perform power normalization operation to obtain a four-dimensional geometric shaping constellation table

[0042] Sampler by probability distribution The one-hot symbol sequence is obtained by adopting it, and then combined with the four-dimensional geometric shaping constellation table Multiply to obtain a four-dimensional frequency domain transmission symbol sequence x;

[0043] When the preset training convergence conditions are met (such as the number of training times reaches the preset upper limit, or the loss function value converges), based on the currently obtained four-dimensional geometric shaping constellation table and the unnormalized log probability The trained constellation shaper is obtained for 4D geometric and probabilistic constellation shaping in the deployment phase.

[0044] Preferably, the sampler can perform sampling based on the Gumbel-Softmax technique.

[0045] Furthermore, in step a3, when the constellation shaper at the transmitting end and the demapper at the receiving end are trained simultaneously, maximizing generalized mutual information (GMI) is used as the optimization goal of the training.

[0046] Furthermore, the loss function used when simultaneously training the constellation shaper at the transmitter and the demapper at the receiver is:

[0047]

[0048] Where E(·) is the mathematical expectation and m is the bit sequence estimate. The sequence length, b i Estimation of the bit sequence The i-th symbol of represents the conditional probability distribution corresponding to the demapper, Hw (X) represents the source entropy, and the subscript w represents the trainable parameters of the constellation shaper and demapper.

[0049] The technical solution provided by the present invention brings at least the following beneficial effects:

[0050] (1) The pruned neural network-based equalization method adopted in the present invention naturally takes into account the fact that ICI is generally concentrated on the adjacent subcarriers of the current subcarrier, reducing the spatial and temporal complexity of the neural network equalization algorithm, so that its parameter quantity and calculation amount increase linearly with the increase in the number of subcarriers, rather than exponentially. In addition, a dual-branch heterogeneous neural network architecture is adopted, and a self-attention mechanism is further introduced to help the network itself better focus on the compensation of linear and nonlinear damage;

[0051] (2) The post-equalization channel modeling method based on a mixed density network adopted by the present invention fits the conditional probability distribution of the channel through a mixed density network, rather than learning a deterministic neural network model with fixed parameters, which is more conducive to simulating the random characteristics of the channel. The idea of ​​equalization first and then channel modeling is adopted, so that the mixed density network only needs to learn the residual distortion and noise of the channel after equalization, which simplifies the difficulty of channel modeling;

[0052] (3) The end-to-end four-dimensional constellation shaping method based on independent sideband modulation adopted by the present invention uses symmetric quadrant constraints to reduce the size of the four-dimensional geometric probability constellation search space, connects the transmitter constellation shaper and the receiver demapper through the equalized post-channel model to consider the full-link damage mechanism, and adaptively learns the globally optimal four-dimensional joint geometric probability constellation shaping scheme, which can further approach the Shannon limit. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0054] Figure 1 It is a schematic diagram of a framework diagram of an implementation method of an optical fiber terahertz communication system optimization method based on artificial intelligence provided by an embodiment of the present invention;

[0055] Figure 2 is a diagram of an equalizer model based on a pruned neural network in an embodiment of the present invention;

[0056] Figure 3 is a diagram of a post-equalization channel model based on a mixed density network in an embodiment of the present invention;

[0057] Figure 4 It is a schematic diagram of the implementation process of the end-to-end four-dimensional joint geometric probability constellation shaper in an embodiment of the present invention. DETAILED DESCRIPTION

[0058] To make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be described in detail and completely in conjunction with the drawings in the implementation of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings can be arranged and designed using different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the present application for protection, but merely represents selected embodiments of the present invention.

[0059] For the convenience of description, firstly, the relevant professional terms appearing in the embodiments of the present invention are explained as follows:

[0060] MZM (Mach-Zehnder Modulator): Mach-Zehnder Modulator

[0061] OFDM (Orthogonal Frequency Division Multiplexing): Orthogonal frequency division multiplexing;

[0062] ISB (Independent Sideband): Independent Sideband

[0063] MDN (Mixture Density Network): Mixture Density Network

[0064] ICI (Inter-Carrier Interference): Inter-carrier interference

[0065] GMI (Generalized Mutual Information): Generalized Mutual Information

[0066] As a possible implementation method, an artificial intelligence-based optical fiber terahertz communication system optimization method provided in an embodiment of the present invention is applicable to a terahertz optical fiber integrated communication system, and can also be applied to other communication systems. The method includes a training phase and a deployment phase. Figure 1 As shown, the specific implementation steps are:

[0067] 1. The training phase includes the following steps:

[0068] S1, the transmitting end randomly generates a transmission symbol sequence, and after passing through the full-link optical fiber terahertz channel, the receiving end obtains the receiving symbol sequence. The transmission symbol sequence and the receiving symbol sequence are used in the subsequent training process of the pruned neural network equalizer;

[0069] S1.1, the transmitter generates a random symbol sequence that obeys a uniform probability distribution, and obtains two independent data after serial-to-parallel conversion. The center frequency is up-converted to f after OFDM modulation and pulse shaping. s , and obtain two independent double-sideband OFDM signals; in this embodiment, the bit rate of a single sideband is set to 10Gbaud, f s =7.5GHz;

[0070] S1.2, two double-sideband signals provide driving input for two MZMs working in push-pull mode to modulate a working frequency of f c A continuous wavelength optical carrier is generated by a single-mode laser to obtain two double-sideband optical signals, and the left and right sidebands are extracted by two optical filters respectively, and an ISB optical signal is obtained by an optical power combiner; in this embodiment, f is set c =193.1THz;

[0071] S1.3, the generated ISB optical signal is sent into a single-mode optical fiber for transmission. At the receiving end of the optical fiber link, it is amplified by an erbium-doped fiber amplifier and then mixed with a signal with an operating frequency of f c+t The continuous wavelength optical carrier generated by the single-mode laser is combined. Based on the principle of optical heterodyne beat frequency, the combined signal is converted into a center frequency of f by UTC-PD. t ISB terahertz signal; in this embodiment, f is set c+t =193.4THz, f t =300GHz, that is, f c+t =f c +f t ;

[0072] S1.4, the generated ISB terahertz signal is radiated into free space by the terahertz antenna. At the receiving end of the terahertz link, the local RF source is mixed with the received terahertz signal after N times of frequency multiplication, and the ISB terahertz signal is converted into an ISB baseband signal. Then, two equal-power ISB baseband signals are obtained through a power divider, and the independent left and right sidebands are separated through two filters.

[0073] S1.5, two independent left and right sidebands are respectively subjected to center frequency down-conversion, pulse shaping and OFDM demodulation to obtain the received symbol sequence;

[0074] S2, the transmitted symbol sequence and the received symbol sequence are divided into training data and verification data for training an equalizer based on a pruned neural network to achieve mapping from received symbols to transmitted symbols;

[0075] In an embodiment of the present invention, the equalizer is implemented by a multi-layer perceptron, and the mean square error is used as the loss function to learn the mapping of the received OFDM frequency domain symbols to the transmitted OFDM frequency domain symbols. Considering the general situation, ICI is mainly concentrated on the adjacent subcarriers of the current subcarrier. Through the pruning technology, only the connection of the neurons corresponding to the input layer subcarriers adjacent to the corresponding neurons of the first hidden layer is retained in the first layer of the neural network, and the subsequent layers only perform point-by-point operations. The pruned neural network adopts a dual-branch heterogeneous form, one of which is a linear branch, which only considers linear distortion, that is, it does not contain an activation function, and the other is a nonlinear branch, which considers nonlinear distortion, that is, it contains an activation function. A self-attention mechanism is further introduced, and a weight of 0-1 is learned through input to weight the linear layer, so that the neural network can better learn the linear and nonlinear distortion itself. In this embodiment, the number of OFDM subcarriers is set to 1024, the number of hidden layers of the pruned neural network is 2, and the ICI length is 1. Since the number of subcarriers is large at this time, it is difficult to draw a diagram. Here, a simplified model is used to describe the structure of the pruned neural network. Assuming that the number of OFDM subcarriers is 3, the number of hidden layers of the pruned neural network is 2, and the ICI length is 1, the schematic diagram of the pruned neural network structure is as follows: Figure 2 As shown, the circles with numbers represent the packed neuron sets, the connection between them is fully connected, SiLU is selected as the activation function, Attn is the attention layer, which can be expressed as two weight parameters w calculated by input lin and w nlin , finally the pruned neural network of this example can be expressed as the following expression:

[0076]

[0077] The subscripts lin and nlin represent linear branches and nonlinear branches respectively, and the corresponding numbers are used to distinguish the network layers, that is, the layer numbers of the network layers. For example, lin3 represents the linear branch of the third layer, w represents the attention weight, and W p ,b represent the linear layer weight and bias respectively. SiLU activation function is selected. Note that due to the application of pruning technology, W p In fact, it is the original complete linear layer weight W and a mask matrix W m The result of the dot multiplication is obtained, and the mask matrix can be easily expressed by the number of subcarriers and the ICI length. For example, the first-layer mask matrix of the simplified pruned neural network model is as follows:

[0078]

[0079] S3, train an MDN-based equalized channel model through the transmitted symbol sequence and the equalized symbol sequence, which is used to subsequently connect the transmitter constellation shaper and the receiver demapper to realize end-to-end gradient direction propagation considering the full link damage mechanism;

[0080] In this embodiment, the channel model is implemented by MDN, the input is the transmitted OFDM frequency domain symbol, and the output is the probability distribution describing the channel condition after equalization The latter includes the residual linear and nonlinear distortion and noise of the equalized channel. MDN learns multiple Gaussian distributions and their weighted combinations by transmitting symbols to achieve the conditional probability distribution of the equalized channel. The fit is as follows:

[0081]

[0082] Among them, π k (x) is the mixing coefficient, The mean is μ k (x), the variance is Gaussian distribution. Note that π k (x) Must meet This can be achieved with Softmax, where the variance needs to be positive, and can be achieved with exponential activation outputs.

[0083] The essence of MDN output is probability distribution, so the negative log-likelihood function is selected as the loss function:

[0084]

[0085] Among them, ω is a trainable parameter in MDN.

[0086] In this embodiment, K=3 is set, the number of hidden layers is 1, and the first layer only considers ICI of length 1 through pruning technology, SiLU is selected as the activation function, the variance is guaranteed to be positive through exponential activation, and the sum of Gaussian distribution weights is guaranteed to be 1 through Softmax. Similarly, due to the large number of subcarriers, it is difficult to draw a picture to describe the specific structure of MDN. Here, a simplified model is still used. Assuming that the number of subcarriers is 3, K=2, and the ICI length is 1, the MDN structure at this time is as follows Figure 3 shown.

[0087] S4. The transmitter generates a four-dimensional constellation table through a geometric shaper, generates a corresponding probability distribution through a probability shaper, and in the training process of the four-dimensional joint geometric probability constellation shaper (referred to as the constellation shaper, the configured constellation shaper training algorithm framework includes a geometric shaper, a probability shaper, a symmetric quadrant constraint layer, a power normalization layer, a Softmax function layer and a sampler) and a demapper, the transmitter generates random data symbols of two independent sidebands, obtains received symbols after OFDM modulation, and obtains an estimate of the bit sequence after OFDM demodulation through a demapper; see Figure 4 , and its specific implementation steps include:

[0088] S4.1. The geometric shaper is implemented in the form of a multi-layer perceptron to complete the transformation from a sequential one-hot symbol table to an unnormalized constellation table in the first quadrant of a four-dimensional space. The mapping of Represents the bit sequence corresponding to the symbol;

[0089] S4.2, the probability shaper is also implemented in the form of a multilayer perceptron, completing the conversion of an arbitrarily specified constant to the corresponding A set of unnormalized log probabilities The mapping of

[0090] S4.3. Layer pairing by symmetry quadrant constraint and The mirror operation of , we get the complete four-dimensional unnormalized constellation table and the corresponding unnormalized log probability

[0091] S4.4, through the Softmax function layer The Softmax operation obtains the probability distribution of the four-dimensional constellation diagram The power normalization layer is based on probability distribution right Normalized to obtain a four-dimensional geometric shaping constellation table

[0092] S4.5. Probability distribution through a sampler based on the Gumbel-Softmax technique Sampling is performed to obtain a one-hot symbol sequence, which is then compared with the four-dimensional geometric shaping constellation table The four-dimensional frequency domain transmission symbol sequence x is obtained by multiplication, that is, the output of the constellation shaper;

[0093] S4.6, the four-dimensional frequency domain transmission symbol sequence x is trained and the equalized channel model with fixed parameters is used to obtain the four-dimensional frequency domain reception symbol sequence after equalization

[0094] S4.7, the demapper is implemented in the form of a multi-layer perceptron to complete the four-dimensional frequency domain reception symbol sequence To estimate the corresponding bit sequence The mapping of

[0095] S4.8. This end-to-end optimization framework takes maximizing GMI as the optimization goal. GMI is the achievable rate of the BICM system. Considering the relationship between the bit cross entropy loss function and GMI, the bit sequence is estimated The loss function for constructing end-to-end joint geometric probability constellation shaping by transmitting symbol distribution calculation is as follows:

[0096]

[0097] The first term in the above equation is equivalent to the bit cross entropy loss of the demapper output, represents the conditional probability distribution corresponding to the demapper, H w (X) represents the source entropy, which needs to be subtracted to avoid the probability shaping from learning the extreme one-hot form. The calculated loss function can be used to train the geometric shaper, probability shaper and demapper end-to-end;

[0098] 2. After the training phase, we will get the pruned neural network equalizer, geometric shaper, probability shaper and demapper with fixed parameters. The specific steps of the deployment phase are as follows:

[0099] S5. A geometric shaper with fixed parameters can be equivalently represented by a fixed constellation table, and a probability shaper with fixed parameters can be equivalently represented by a corresponding probability distribution. The transmission symbol sequence is first obtained according to the information bit sequence through the constellation table and the corresponding probability distribution, wherein the probability shaping can be achieved by a symbol-by-symbol matching or a block-by-block matching method. The former is like a lookup table and a dynamic programming method, and the latter is like CCDM; that is, the information bit sequence can be mapped into a four-dimensional frequency domain transmission symbol sequence x by the four-dimensional constellation table output by the trained four-dimensional joint geometric probability constellation shaper.

[0100] S6, the transmitted symbol sequence obtains the corresponding received symbol sequence through the full-link optical fiber terahertz channel in step 1 of the same training phase, and obtains the equalized received symbol sequence through a pruned neural network equalizer with fixed parameters;

[0101] S7. The equalized received symbol sequence is passed through a demapper to obtain an estimate of a received information bit sequence.

[0102] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

[0103] The above are only some embodiments of the present invention. For those skilled in the art, several modifications and improvements can be made without departing from the creative concept of the present invention, which all belong to the protection scope of the present invention.

Claims

1. An optimization method for optical fiber terahertz communication system based on artificial intelligence, characterized in that: The following steps are involved: Training phase steps: Step a1, the transmitting end randomly generates a transmitting symbol sequence, and after passing through the full-link optical fiber terahertz channel, a receiving symbol sequence is obtained at the receiving end; Using the transmitted symbol sequence and the received symbol sequence as training data, an equalizer based on pruned neural network is trained; Step a2, training a post-equalization channel model based on a mixed density network based on the transmitted symbol sequence and the post-equalization symbol sequence; Among them, the equalized channel model is used to output the probability distribution describing the equalized channel conditions The composition parameters of The fitting of The symbol sequence after equalization is: the transmitted symbol sequence is input into the trained equalizer based on the pruned neural network, and the symbol sequence after equalization is obtained based on the output thereof; Step a3, simultaneously training a constellation shaper at the transmitting end and a demapper at the receiving end, wherein both the constellation shaper and the demapper are network models based on a neural network; The constellation shaper includes a geometric shaper and a probability shaper, which are used to generate a four-dimensional geometric and probability constellation table to shape the transmitting end bit sequence to obtain a four-dimensional frequency domain transmission symbol sequence x; The four-dimensional frequency domain transmission symbol sequence x is input into the trained equalized channel model to obtain the four-dimensional frequency domain reception symbol sequence Then the four-dimensional frequency domain received symbol sequence Input demapper to obtain the corresponding bit sequence estimate Deployment phase steps: Step b1, the trained constellation shaper generates a constellation table for shaping the information bit sequence of the transmitting end to obtain a four-dimensional frequency domain transmission symbol sequence x; Step b2, the four-dimensional frequency domain transmission symbol sequence x passes through the same full-link optical fiber terahertz channel as in step a1, and obtains a corresponding reception symbol sequence at the receiving end, and then sends the reception symbol sequence to a trained pruned neural network-based equalizer to obtain an equalized reception symbol sequence; Step b3: After equalization, the received symbol sequence is demapped to obtain an estimate of the received information bit sequence.

2. The method according to claim 1, characterized in that In step a1, the transmitting end randomly generates a transmission symbol sequence, and after passing through the full-link optical fiber terahertz channel, the receiving symbol sequence obtained at the receiving end includes: (a1.1) The transmitter generates a random symbol sequence that follows a uniform probability distribution. After serial-to-parallel conversion, two independent data channels are obtained. They are modulated by orthogonal frequency division multiplexing (OFDM) and pulse-shaped and then up-converted to a center frequency of f. s , and obtain two independent double-sideband OFDM signals; (a1.2) Two independent double-sideband OFDM signals provide driving input for two Mach-Zehnder modulators (MZMs) operating in push-pull mode to modulate the continuous wavelength optical carrier generated by the single-mode laser to obtain two double-sideband optical signals. The operating frequency of the single-mode laser is f c ; The left and right sidebands of the two double-sideband optical signals are extracted respectively through two optical filters, and the independent sideband ISB optical signal is obtained through an optical power combiner; (a1.3) The ISB optical signal is sent into a single-mode optical fiber for transmission. At the receiving end of the optical fiber link, it is amplified by an erbium-doped fiber amplifier and then mixed with a signal with an operating frequency of f. c+t The continuous wavelength optical carrier generated by the single-mode laser is combined, and the combined signal is converted into a central frequency f by a single-carrier photodiode. t ISB terahertz signal; (a1.4) The ISB terahertz signal is radiated into free space through the terahertz antenna. At the receiving end of the terahertz link, the local RF source is mixed with the received terahertz signal after multiple frequency multiplication to convert the ISB terahertz signal into an ISB baseband signal. Then, two equal-power ISB baseband signals are obtained through a power divider, and then the independent left and right sidebands are separated through two filters. (a1.5) The two independent left and right sidebands are respectively subjected to center frequency down-conversion, pulse shaping and OFDM demodulation to obtain the received symbol sequence.

3. The method according to claim 1, characterized in that The pruned neural network based equalizer includes multiple layers of perception layers, and the number of neurons in the input layer of the pruned neural network equalizer is twice the number of subcarriers of the input transmission symbol sequence.

4. The method according to claim 3, characterized in that The equalizer based on pruned neural network adopts a dual-branch heterogeneous form, one of which is a linear branch including several layers of linear layers, and the other is a nonlinear branch including several layers of linear layers with nonlinear activation functions, and the number of layers of the two branches is the same; Among them, each neuron of the first hidden layer of the linear branch and the nonlinear branch is connected only with the neuron corresponding to the subcarrier of the input layer adjacent to the subcarrier represented by the current neuron; from the second layer onwards, the neurons of the adjacent layers of the linear branch and the nonlinear branch are connected only with the neurons representing the same subset of subcarriers; The input data of the pruned neural network based equalizer is sent to the self-attention layer to learn the self-attention weight of the linear branch, and the output result of each subcarrier output by the linear branch is weighted, and then added to the output result of the corresponding subcarrier output by the nonlinear branch to obtain the output of the pruned neural network based equalizer.

5. The method according to claim 4, characterized in that The weight of each network layer of the linear branch and the nonlinear branch during operation is the weight W of the network layer and the corresponding mask matrix W m The dot product of Among them, the mask matrix W of the first layer of the linear branch and the nonlinear branch is m Based on the number of subcarriers N sub and inter-subcarrier interference ICI length N ICI Determine, 0 means the subcarrier at the current position is not interfered, 1 means the subcarrier at the current position is interfered, and the number of elements 1 in each row is: 2×N ICI +2; The linear branch and the nonlinear branch start from the second layer, and the mask matrix W during operation is m is a block diagonal matrix.

6. The method according to any one of claims 1 to 5, characterized in that: The pruned neural network based equalizer uses mean square error as the loss function during training.

7. The method according to claim 1, characterized in that Each neuron in the first hidden layer of the equalized channel model is connected only to the neurons corresponding to the subcarriers of the input layer adjacent to the subcarrier represented by the current neuron, and starting from the second layer, the neurons in adjacent layers are connected only to the neurons representing the subcarriers of the same subset.

8. The method according to claim 1, characterized in that The equalized channel model selects the negative log-likelihood function as its loss function during training.

9. The method according to claim 1, characterized in that In step a3, when the constellation shaper of the transmitter and the demapper of the receiver are trained simultaneously, the training algorithm framework of the configured constellation shaper includes a geometric shaper, a probability shaper, a symmetric quadrant constraint layer, a power normalization layer, a Softmax function layer and a sampler; The geometry reshaper is used to transform the sequential one-hot symbol table into an unnormalized constellation table in the first quadrant of the four-dimensional space. The mapping of Represents the bit sequence corresponding to the symbol; The probability shaper is used to convert an arbitrarily specified constant to the corresponding A set of unnormalized log probabilities The mapping of The symmetric quadrant constraint layer is used to and Perform a mirror operation to obtain a complete four-dimensional unnormalized constellation table and the corresponding unnormalized log probability The Softmax function layer is used to output the symmetric quadrant constraint layer. Perform the Softmax operation to obtain the probability distribution of the four-dimensional constellation diagram The power normalization layer is based on probability distribution Output of the symmetric quadrant constraint layer Perform power normalization operation to obtain a four-dimensional geometric shaping constellation table Sampler by probability distribution The one-hot symbol sequence is obtained by adopting it, and then combined with the four-dimensional geometric shaping constellation table Multiply to obtain a four-dimensional frequency domain transmission symbol sequence x; When the preset training convergence conditions are met, based on the currently obtained four-dimensional geometric shaping constellation table and the unnormalized log probability Get the trained constellation shaper.

10. The method according to claim 1 or 9, characterized in that: In step a3, when the constellation shaper at the transmitting end and the demapper at the receiving end are trained simultaneously, maximizing the generalized mutual information is used as the optimization goal of the training.

Images