End-to-end impairment suppression photon terahertz communication system based on channel modeling driving and geometric shaping

Abstract

The invention provides an end-to-end impairment suppression photon terahertz communication system based on channel modeling driving and geometric shaping, and belongs to the field of terahertz communication. Comprising a sending end and a receiving end; the transmitting end introduces learnable geometric shaping, generates a terahertz signal by using a photon beat frequency method, transmits the terahertz signal to the receiving end through a single-mode optical fiber, and the receiving end performs digital signal processing through learnable carrier phase recovery and a constellation mapping neural network to recover an original signal. A conditional denoising diffusion model is used for carrying out high-fidelity and strong-generalization-ability symbol-level differentiable channel modeling on an end-to-end channel of a photon terahertz communication system, and an auto-encoder neural network is embedded to carry out end-to-end overall training, so that geometric shaping and joint sensing, modeling and suppression on channel damage are realized; the performance of the photon terahertz communication system is optimized, and the quality and stability of signal transmission are remarkably improved.

Description

Technical Field

[0001] This invention relates to an end-to-end damage suppression photonic terahertz communication system based on channel modeling and geometry shaping, belonging to the field of terahertz communication. Background Technology

[0002] With the rapid development of information technology and the exponential growth of data demand, applications such as high-definition video streaming, the Internet of Things, smart cities, and immersive virtual experiences are constantly emerging. Existing 5G technology can no longer fully meet the future communication needs of high bandwidth and low latency.

[0003] Terahertz (THz) communication is considered globally as one of the core technologies of 6G communication and an important breakthrough for future communication technologies due to its extremely rich spectrum resources.

[0004] In photonic terahertz communication, there are various linear and nonlinear signal impairments. The signal impairments generated by different terahertz devices and channels are mutually coupled, nonlinear, and difficult to compensate accurately on their own. How to effectively compensate for linear and nonlinear impairments has become a key challenge that urgently needs to be solved.

[0005] Traditional damage compensation techniques based on cascaded signal processing modules typically require accurate damage models and parameter estimations, leading to increased complexity and performance bottlenecks. Neural network-based damage suppression techniques, however, have gained increasing attention due to their significant performance improvements and have demonstrated performance surpassing traditional digital signal processing to some extent.

[0006] A recent damage suppression technique based on end-to-end learning has attracted researchers' attention. This method can jointly sense and suppress various complex nonlinear damages in terahertz communication links. However, this method places higher demands on the accuracy and generalization ability of terahertz differentiable channel modeling.

[0007] Currently, channel modeling methods based on feedforward neural networks, long short-term memory networks, and conditional generative adversarial networks have been developed for terahertz differentiable channels. These methods can effectively fit the channel transfer function, capture the nonlinear characteristics and dynamic changes of complex channels, and have achieved results superior to traditional modeling methods in many studies. However, their generalization ability remains limited under conditions of different transmitted signal constellations, different transmitted powers, and different channel scenarios.

[0008] Therefore, there is an urgent need for a modeling method that can achieve both high accuracy and strong generalization ability, as well as an end-to-end learning neural network architecture that can accurately capture channel modeling features and transmission signal impairment features, to perform overall training at the transmitting and receiving ends, and to achieve joint perception, modeling and suppression of various complex nonlinear impairments in terahertz communication links, thereby improving the quality and stability of signal transmission. Summary of the Invention

[0009] To address the problem of multiple linear and nonlinear signal impairments being coupled together, nonlinear, and difficult to compensate accurately on their own in existing terahertz communication systems, this invention proposes an end-to-end impairment suppression photonic terahertz communication system based on channel modeling and geometric shaping. It utilizes a conditional denoising diffusion model to perform high-fidelity, highly generalizable symbol-level differentiable channel modeling of the end-to-end channel of the photonic terahertz communication system, and embeds an autoencoder neural network for end-to-end overall training. This enables geometric shaping and joint sensing, modeling, and suppression of channel impairments, thereby optimizing the performance of the photonic terahertz communication system and significantly improving the quality and stability of signal transmission.

[0010] The end-to-end damage suppression photonic terahertz communication system based on channel modeling and geometry shaping includes a transmitter and a receiver.

[0011] The transmitter introduces learnable geometric shaping and uses photonic beat frequency to generate terahertz signals, which are transmitted to the receiver via single-mode fiber. The receiver performs digital signal processing through a learnable carrier phase recovery and deconstellation mapping neural network to recover the original signal.

[0012] The transmitting end includes: two external cavity lasers ECL1 and ECL2, a TX-DSP module, an I / Q modulator, an optical coupler, a single-mode fiber, an erbium-doped fiber amplifier, and a single-row carrier photodetector.

[0013] An external cavity laser ECL1 generates a continuous optical carrier and inputs it to an I / Q modulator. The I / Q modulator is also connected to a TX-DSP module, which uses an encoder module of a learnable autoencoder neural network to generate a geometrically shaped constellation point set and maps the transmitted bit signal sequence onto the shaped constellation point set. Through digital signal processing, a GS-MQAM baseband signal is obtained, which is used to drive the I / Q modulator to modulate the input optical carrier. The modulated optical carrier is coupled to the continuous light wave output from the external cavity laser ECL2 through an optical coupler, transmitted through a single-mode fiber to an erbium-doped fiber amplifier for amplification, and finally input to a single-row carrier photodetector for beat frequency to obtain a terahertz signal.

[0014] The receiving end includes: a mixer, a local oscillator, and an RX-DSP module;

[0015] The terahertz signal is input into a mixer and mixed with the local oscillator signal generated by the local oscillator to achieve analog down-conversion of the signal and obtain an intermediate frequency signal. Then, it is input into the RX-DSP module, where it undergoes digital down-conversion, low-pass filtering, downsampling, symbol synchronization, power normalization, channel equalization, and demodulation (carrier phase recovery, geometric shaping demapping) by the decoder module of the learnable autoencoder neural network. Finally, the original data is obtained, and the bit error rate is calculated.

[0016] The encoder of the autoencoder neural network is, at its core, a learnable (M,2)-sized lookup table used to learn the constellation point I / Q values ​​for MQAM geometric shaping. The encoder's input is a binary bit sequence. After converting it into an M-ary index sequence, the corresponding I / Q constellation point is looked up in the lookup table and a symbol sequence is formed. Then, the power of the entire sequence is normalized, and the transmitted symbol sequence is output. .

[0017] The decoder of the autoencoder neural network consists of two core parts: a carrier phase recovery module and a constellation demapping module.

[0018] The carrier phase recovery module includes coarse phase compensation and fine phase compensation. The coarse phase compensation performs preliminary compensation on the initial received symbol sequence S0 received by the decoder to obtain the coarsely phase-compensated symbol S1. The fine phase compensation performs further compensation on the coarsely phase-compensated symbol sequence S1 to obtain the finely phase-compensated symbol S2. Each compensation includes one phase prediction and one phase recovery.

[0019] Phase prediction consists of a context information acquirer and a phase predictor. The context information acquirer is composed of a multi-layer bidirectional long short-term memory network, which acquires the context information contained in the symbol sequence (S0 or S1) of this compensation. The phase predictor is composed of a multi-layer perceptron, which collects the context information, adds it to the symbol sequence of this compensation and the geometrically shaped constellation point information generated by the encoder, and performs compression and refinement processing on it, mapping it to a rotation vector (cos(theta), sin(theta)).

[0020] The phase recovery section uses rotation vector information to perform rotation compensation on the symbol sequence of this compensation.

[0021] The constellation demapping module is responsible for performing geometric shaping and demapping on the symbol sequence S2 after it has been processed by the carrier phase recovery module, and then making a decision to obtain the demodulated bit sequence. This module consists of a multi-layer sensing mechanism.

[0022] The self-encoder neural network embedded channel modeling described in this invention is specifically as follows:

[0023] First, collect historical complex sequence numbers of symbols from the transmitting and receiving ends, and divide them into training and testing sets;

[0024] Then, the conditional denoising diffusion model is trained using the training set to obtain the denoising neural network weights of the channel model;

[0025] The conditional denoising diffusion model includes: a forward diffusion module, a conditional vector processing module, and a reverse denoising module.

[0026] The forward diffusion module is a non-neural network process, using the training set receiver symbol sequence. As input, a small amount of Gaussian noise is added at multiple time steps to obtain the final time step. Noisy data The conditional vector processing module will add noise time step information. , Global position information of each symbol in the entire sequence Training set sender symbol sequence The context information of each symbol before and after it Position encoding and merging are performed together, including the context information of the preceding and following symbols. Additional multi-head self-attention mechanism was implemented to learn the relationship between the varying degrees of influence of different positions on the output. The information merging process was learned using a multilayer perceptron to obtain a multi-source information conditional vector. ;

[0027] Inverse denoising module to add noise to data and multi-source information conditional vector As input, a denoising neural network is trained, and the condition vector is processed through a feature-level linear modulation (FiLM) mechanism in the intermediate hidden layer. The impact on the output is further amplified, making it possible to achieve the same effect. Under control, the sample output generated by the denoising neural network and Consistent.

[0028] Finally, the trained denoising neural network is embedded into the autoencoder neural network for end-to-end overall training.

[0029] The end-to-end overall training process is as follows:

[0030] First, the hyperparameters and the weights of the encoder and decoder are initialized. The optimizer used for training is AdamW, and the learning rate optimizer used is OneCycleLR.

[0031] Then, following the order of encoder-denoising neural network (channel modeling)-carrier phase recovery module-constellation demapping module, they are sequentially connected into a whole. Afterwards, pre-training, training, and post-training are performed in stages. Different modules actually participate in training at each stage, and the weights of modules not participating in training are frozen. For the final training result, a demodulated bit sequence is used. and the original bit sequence The bit error rate and generalized mutual information per bit are evaluated.

[0032] The training tasks for each stage are as follows:

[0033] 1) In Phase 1, the weights of the pre-trained encoder and carrier phase recovery module are frozen, and the weights of the denoising neural network and constellation demapping module are frozen, resulting in a short training cycle.

[0034] 2) Phase 2: Pre-train the constellation demapping module, freeze the weights of the encoder, denoising neural network and carrier phase recovery module, and the training rounds are short.

[0035] 3) In the third stage, the autoencoder is formally trained end-to-end, while only the denoising neural network is frozen, resulting in a long training cycle.

[0036] 4) Phase four: Post-training of carrier phase recovery module and constellation demapping module, freezing encoder and denoising neural network, with short training rounds.

[0037] Finally, the trained encoder is placed in the TX-DSP module to convert the transmitted bit sequence into a geometrically shaped transmitted symbol sequence; after processing by a denoising neural network, the received symbol sequence is obtained; the decoder is placed in the RX-DSP module to demodulate the received symbol sequence into the original transmitted bit sequence.

[0038] The advantages of this invention are:

[0039] 1. An end-to-end impairment-suppressed photonic terahertz communication system based on channel modeling and geometry shaping. Through a conditional denoising diffusion model, symbol-level differentiable modeling of the end-to-end channel of the photonic terahertz communication system is achieved, demonstrating high fidelity and strong generalization capability. Compared to traditional analytical mathematical channel models and other deep learning channel models, it exhibits significant advantages in sample generation quality, diversity, differentiability, training stability, and generalization ability.

[0040] 2. An end-to-end impairment suppression photonic terahertz communication system based on channel modeling and geometric shaping. By embedding a trained differentiable channel model into an autoencoder neural network for end-to-end overall training, geometric shaping and joint sensing, modeling and suppression of channel impairments are achieved, thereby optimizing the performance of the photonic terahertz communication system and significantly improving the quality and stability of signal transmission.

[0041] 3. An end-to-end damage suppression photonic terahertz communication system based on channel modeling and geometry shaping, which has superior system bit error rate performance and higher generalized mutual information compared with traditional photonic terahertz communication systems. Attached Figure Description

[0042] Figure 1 This is a structural diagram of an end-to-end damage suppression photonic terahertz communication system based on channel modeling and geometry shaping according to the present invention.

[0043] Figure 2 This is a map showing the acquisition location of the channel training dataset collected using traditional modulation and demodulation methods in this invention.

[0044] Figure 3 This is a structural diagram of the conditional denoising diffusion model for channel modeling of the photonic terahertz communication system in this invention;

[0045] Figure 4 Figure 1 shows the result of channel modeling for the conditional denoising diffusion model in this invention. Figure 2a is the training loss function variation curve, and Figure 3b is a constellation comparison diagram of the actual channel output and the modeled channel output.

[0046] Figure 5 This is a diagram of the end-to-end autoencoder neural network structure used for geometry shaping and channel impairment suppression in this invention;

[0047] Figure 6 This is a constellation diagram showing the experimental verification results of the overall scheme in this invention; Detailed Implementation

[0048] To facilitate understanding and implementation of the present invention by those skilled in the art, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Obviously, the described embodiments are merely some, not all, embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort should fall within the scope of protection of the present invention.

[0049] This invention proposes an end-to-end damage suppression photonic terahertz communication system based on channel modeling and geometric shaping, including a transmitter and a receiver. The transmitter introduces learnable geometric shaping and generates a terahertz signal using a photonic beat frequency method, which is transmitted to the receiver via a single-mode fiber. The receiver performs digital signal processing on the signal using a learnable carrier phase recovery and deconstellation mapping neural network to recover the original signal.

[0050] like Figure 1 As shown, the transmitting end includes: two external cavity lasers ECL1 and ECL2, a TX-DSP module, an I / Q modulator, an optical coupler, a single-mode fiber, an erbium-doped fiber amplifier, and a single-row carrier photodetector.

[0051] Two external cavity lasers, ECL1 and ECL2: the output of ECL1 is used as the input optical carrier of the I / Q modulator, and the output of ECL2 is used as the optical carrier for the beat frequency of the erbium-doped fiber amplifier (UTC-PD).

[0052] The TX-DSP module performs digital signal processing on the binary signal sequence to obtain the transmitted signal, which drives the I / Q modulator to modulate the optical carrier. It incorporates a learnable geometric shaping encoder neural network to convert the binary signal sequence into a transmitted symbol sequence.

[0053] I / Q modulator: Performs optical I / Q modulation controlled by the electrical signal output from the TX-DSP module.

[0054] Optical Coupler (OC): Used to couple the light wave output from the IQ modulator with the continuous light wave output from the ECL2.

[0055] 20km single-mode fiber (SSMF): used to transmit modulated optical signals.

[0056] Erbium-doped fiber amplifier (EDFA): Used to amplify signals generated by TX-DSP.

[0057] Single-carrier photodetector (UTC-PD); used to beat optical signals to obtain terahertz frequency signals.

[0058] An external cavity laser ECL1 generates a continuous optical carrier and inputs it to an I / Q modulator. The I / Q modulator is also connected to a TX-DSP module, which uses an encoder module of a learnable autoencoder neural network to generate a geometrically shaped constellation point set and maps the transmitted bit signal sequence onto the shaped constellation point set. Through digital signal processing, a GS-MQAM baseband signal is obtained, which is used to drive the I / Q modulator to modulate the input optical carrier. The modulated optical carrier is coupled with the continuous light wave output from the external cavity laser ECL2 through an optical coupler, and then transmitted through a 20km single-mode fiber to an erbium-doped fiber amplifier (EDFA) for amplification. Finally, it is input to a single-row carrier photodetector (UTC-PD) for beat frequency to obtain a terahertz signal.

[0059] like Figure 1 As shown, the receiving end includes: a mixer, a local oscillator, and an RX-DSP module;

[0060] Mixer: Used to down-convert the received signal to obtain the baseband signal.

[0061] Local oscillator: used as input to the mixer.

[0062] The RX-DSP module performs digital signal processing on the received signal, including down-conversion, low-pass filtering, downsampling, symbol synchronization, power normalization, channel equalization, learnable autoencoder decoder neural network demodulation (carrier phase recovery, geometric shaping demapping), bit error rate calculation, and other operations.

[0063] The terahertz signal is input into a mixer and mixed with the local oscillator signal generated by the local oscillator to achieve analog down-conversion of the signal and obtain an intermediate frequency signal. Then it is input into the RX-DSP module, where it undergoes digital down-conversion, low-pass filtering, downsampling, symbol synchronization, power normalization, and demodulation (carrier phase recovery, geometric shaping demapping) by the decoder module of the learnable autoencoder neural network. Finally, the original data is obtained, and the bit error rate is calculated.

[0064] The self-encoder neural network embedded channel modeling described in this invention is specifically as follows:

[0065] Step 1: Build a photonic terahertz communication system, collect historical complex sequence of symbols from the transmitter and receiver, and divide the system into training and testing sets for training the channel model.

[0066] In the TX-DSP and RX-DSP modules, traditional non-neural network modulation and demodulation methods are used; complex sequences of transmitted symbols are acquired after constellation mapping at the transmitter and after symbol synchronization and channel equalization at the receiver, respectively. and receive symbolic complex sequence Collection location such as Figure 2As shown, each set of data has a length of 32768, and the data is preprocessed to obtain the training dataset required for the channel model, and the training set and test set are divided.

[0067] The preprocessing includes: separating the real and imaginary parts of the acquired complex sequences of transmitted and received symbols; aligning the transmitted symbols (input features) with their corresponding received symbols (output labels) to form correct matching pairs. .

[0068] Step 2: Construct a conditional denoising diffusion model for channel modeling of photonic terahertz communication systems.

[0069] The conditional denoising diffusion model is trained using the training set to obtain the denoising neural network weights of the channel model.

[0070] The conditional denoising diffusion model includes: a forward diffusion module, a conditional vector processing module, and a reverse denoising module.

[0071] like Figure 3 As shown, the forward diffusion process is fixed and requires no learning. Based on a pre-defined noise schedule, it progressively transforms arbitrarily complex data distributions into a simple, known prior distribution, typically a standard Gaussian distribution. The reverse denoising process, on the other hand, is a learning neural network that aims to learn from the prior distribution obtained in the forward diffusion process, progressively denoising to ultimately generate samples that conform to the original data distribution.

[0072] The specific construction process is as follows:

[0073] Step 2.1: Construct the forward diffusion process.

[0074] The forward diffusion module is a non-neural network process, using the training set receiver symbol sequence. As input, a small amount of Gaussian noise is added at multiple time steps to obtain the final time step. Noisy data ;

[0075] ( )

[0076] Each step moves towards the current data. Add a small amount of Gaussian noise (dimension and (Same) Its single-step transfer formula is:

[0077]

[0078] in It is a predefined noise scheduling, usually in The sequence gradually increases between them, ensuring that in The data becomes completely noisy after the step. This example uses reparameterization techniques to directly modify the original data. Get any time step Noisy data :

[0079]

[0080] Step 2.2: Process and merge the multi-source information condition vectors, which is divided into two stages: independent encoding and condition integration;

[0081] The conditional vector processing module adds noise time step information. , Global position information of each symbol in the entire sequence Training set sender symbol sequence The context information of each symbol before and after it Position encoding and merging are performed together, including the context information of the preceding and following symbols. Additional multi-head self-attention mechanism was implemented to learn the relationship between the varying degrees of influence of different positions on the output; the information merging process used a multilayer perceptron to learn the multi-source information conditional vector. ;

[0082] The operations in the independent coding phase include:

[0083] 1) Information on diffusion time steps and the global position information of the current symbol in the entire sequence (Time information) is embedded with positional encoding at dimensions of 256 and 128. Positional encoding uses sine and cosine functions to generate Fourier feature embeddings, mapping scalar discrete information to a high-dimensional continuous space, transforming it into a high-dimensional continuous feature representation that is easier for the model to learn.

[0084] The purpose of positional encoding of the diffusion time step information is to distinguish different t values, as different t values ​​correspond to different noise levels. The denoising neural network needs to employ different denoising strategies for different t values. The purpose of positional encoding of the time information is to enable the model to better capture the relative time information between different symbols, which helps the denoising neural network learn the accumulation of phase and frequency shifts over time.

[0085] 2) Contextual information is extracted from the input features. First, the contextual input features of 15 symbols are concatenated. Then, sequence position encoding with an embedding dimension of 256 is performed to provide the relative and absolute position information of each element in the sequence. Finally, a module consisting of two Transformer Encoder layers is used to process the emitted symbol sequence with positional encoding. The multi-head attention mechanism (4 heads) of the Transformer Encoder layer is utilized to capture the contextual information and long-range dependencies within the sequence. Extracting contextual information from the input features helps the channel model learn and simulate features such as inter-symbol interference.

[0086] The operations in the condition integration phase include:

[0087] 1) The diffusion time step information, global location information, input features and their context information processed in the independent encoding stage are concatenated.

[0088] 2) The concatenated conditional information is further compressed and refined by passing it through a small, simple feedforward network consisting of two linear layers (with a hidden layer dimension of 256) and two LeakyReLU activation functions. The most relevant features are extracted to obtain the processed conditional vector. The dimension is 256.

[0089] Step 2.3: Construct the reverse denoising process.

[0090] The inverse denoising module obtains the noisy data from the forward diffusion process. and the processed multi-source information condition vector As input, a denoising neural network is trained, and the condition vector is processed through a feature-level linear modulation (FiLM) mechanism in the intermediate hidden layer. The impact on the output is further amplified, making it possible to achieve the same effect. Under control, the sample output generated by the denoising neural network and Consistent.

[0091] The denoising neural network uses a four-layer multilayer perceptron (MLP) to progressively denoise the data, ultimately generating samples that conform to the distribution of the original data. The goal of this process is to learn a parameterized Markov chain. It can approximate the true inverse conditional distribution (a Gaussian distribution). The mean and variance of this distribution depend on Training a neural network To predict in Added noise Based on the predicted noise, it can be deduced that... The estimated value, and then the estimated value. The mean and variance of the sample are calculated. This process is also called sampling, and the final formula for a single-step sampling is as follows:

[0092]

[0093] in, It is new random noise;

[0094] In the denoising neural network, the hidden layer dimension of the MLP is 256. A Feature-Level Linear Modulation (FiLM) mechanism is added to the two intermediate hidden layers to inject conditional information into the intermediate layers of the denoising process. The FiLM mechanism learns two affine transformation parameters (scaling...) and offset () is used to adjust the activation of neural network layers channel-wise, and its mathematical representation is as follows:

[0095]

[0096] in These are feature maps of conditional neural network layers. Represents a condition vector. This indicates element-wise multiplication.

[0097] In this way, FiLM allows the model to flexibly change the representation of internal features according to external conditions without modifying the main structure of the network to emphasize the correspondence between conditional information and generated data.

[0098] Step 3: Train the conditional denoising diffusion model to obtain the trained denoising neural network weights and evaluate the capabilities of the channel model;

[0099] The training process is as follows:

[0100] First, the hyperparameters and neural network weights are initialized, including the initial learning rate lr of 1e-4, batch size of 1024, training epochs of 1000, exponential moving average (EMA) parameter ema_decay of 0.999, diffusion steps of 1000, sampling steps of 50, and randomness parameter ddim_eta of 0.0.

[0101] During training, the following optimizers were used: AdamW, ReduceLROnPlateau (the learning rate optimizer showed no improvement in loss after 50 epochs, the learning rate decreased by 50%, and the minimum learning rate was 1e-6), and cosine wave optimizer. Noise scheduling.

[0102] Then, to receive symbols As the starting point of the forward diffusion process, the time step is obtained. Noisy data The denoised prediction data is then input into the denoising neural network along with the processed and merged multi-source information conditional vector to obtain the denoised original prediction data. .

[0103] Finally, normalized mean square error (NMSE) was used as an evaluation index to quantitatively analyze the accuracy of the conditional denoising diffusion model in modeling the photonic terahertz communication channel.

[0104] The formula for calculating the normalized mean square error is:

[0105]

[0106] After training is complete, the sampling process can be used to extract data from completely random pure Gaussian noise. Through neural networks Iteration Step by step, each step from Predict noise and generate This continues until an output label (receive symbol) is generated. The generated It is controlled by input characteristics (transmitted symbols). This process simulates the actual photonic terahertz communication channel's process from transmitting symbols to receiving symbols.

[0107] To analyze the channel model's capabilities, traditional 256QAM transmit and receive signals were used as the channel training dataset for training the channel model. A standard 16QAM signal was then input into the trained channel model. The training process included loss function curves and a constellation comparison diagram of the actual and modeled channel outputs under the input standard 16QAM signal. Figure 4 As shown.

[0108] Step 4: Using the trained denoising neural network, construct an end-to-end autoencoder neural network for geometry shaping and channel impairment suppression.

[0109] An autoencoder neural network consists of two parts: an encoder and a decoder. Figure 5 As shown, the encoder is responsible for converting the transmitted bit sequence into a geometrically shaped transmitted symbol sequence, which, after processing by the channel model, yields the received symbol sequence. The decoder is responsible for demodulating the received symbol sequence back into the original transmitted bit sequence. During this process, the encoder and decoder are trained together to achieve joint sensing, modeling, and suppression of channel impairments, thereby optimizing the performance of the photonic terahertz communication system.

[0110] The specific construction process is as follows:

[0111] Step 4.1: Construct the encoder. The core of the encoder is a learnable (M,2)-sized lookup table used to learn the constellation point I / Q values ​​for MQAM geometry shaping. The encoder's input is the transmitted binary bit sequence. After converting it into an M-ary index sequence, the corresponding I / Q constellation points are found in the learnable lookup table to form a sequence. Then, the power of the entire sequence is normalized to obtain the output transmitted symbol sequence. .

[0112] Step 4.2: Construct the decoder; the core of the decoder consists of two parts: a carrier phase recovery module and a constellation demapping module.

[0113] The carrier phase recovery module includes two compensation processes: coarse phase compensation and fine phase compensation. These two compensation processes work on similar principles, but the signals and network scales they process are slightly different. Each process includes one phase prediction and one phase recovery.

[0114] The coarse phase compensation performs preliminary compensation on the initial received symbol sequence S0 received by the decoder to obtain the coarsely phase-compensated symbol S1. The fine phase compensation performs further compensation on the coarsely phase-compensated symbol sequence S1 to obtain the finely phase-compensated symbol S2. Compared with the fine phase compensation network, the coarse phase compensation network has a larger compensation range and a larger network scale.

[0115] For coarse-tuned phase compensation, the context information acquirer Bi-LSTM has a hidden layer dimension of 64 and a layer count of 2; the phase predictor MLP consists of 3 linear layers (with hidden layer dimensions of 64, 32, and 2 respectively) and 2 LeakyReLU activation functions; for fine-tuned phase compensation, the context information acquirer Bi-LSTM has a hidden layer dimension of 32 and a layer count of 1; the phase predictor MLP consists of 3 linear layers (with hidden layer dimensions of 32, 16, and 2 respectively) and 2 LeakyReLU activation functions.

[0116] The phase prediction section of the carrier phase recovery module consists of two parts: a bidirectional long short-term memory (Bi-LSTM) context information acquirer and a feedforward neural network phase predictor (a multilayer perceptron MLP). The context information acquirer obtains the context information contained in the symbol sequence (S0 or S1) of the current compensation. The phase predictor collects this context information, adds it to the symbol sequence of the current compensation, and the constellation point set of the encoder's geometric shaping, compresses and refines this information together, and finally maps it to a rotation vector (cos(theta), sin(theta)).

[0117] The phase recovery section of the carrier phase recovery module performs rotation compensation on the symbol sequence of this compensation using the obtained rotation vector information.

[0118] The constellation demapping module is a simple multilayer perceptron consisting of three linear layers (with hidden layer dimensions of 64, 32, and 4 respectively). It consists of two LeakyReLU activation functions, responsible for geometrically shaping and demapping the symbol sequence S2 after it has been processed by the carrier phase recovery module, and making a decision to obtain the demodulated bit sequence. This module consists of a multi-layer sensing mechanism.

[0119] Step 5: Train the end-to-end autoencoder neural network and evaluate it.

[0120] Autoencoder neural networks are self-supervised neural networks that do not require a pre-prepared training dataset. They simply generate random binary bit sequences and then connect them sequentially in the order of encoder - denoising neural network (channel modeling) - carrier phase recovery module - constellation demapping module. Afterwards, they undergo phased pre-training, training, and post-training, with different modules participating in training at each stage. Modules not participating in training have their weights frozen. The final training result is obtained using the demodulated bit sequence. and the original bit sequence The bit error rate and generalized mutual information per bit are evaluated, and the loss function is calculated by comparing the obtained bit sequence with the original bit sequence, which can then be used for training.

[0121] End-to-end autoencoder neural networks have many modules, making simultaneous and rapid training difficult. Therefore, a staged training approach is adopted for pre-training and post-training. The training tasks for each stage are as follows:

[0122] 1. Phase 1: Pre-train the carrier phase recovery module of the encoder and decoder, freeze the weights of the constellation demapping module of the channel model and decoder, and train 1000 rounds.

[0123] 2. Phase Two: Pre-train the constellation demapping module of the decoder, freeze the weights of the encoder, channel model and the carrier phase recovery module of the decoder, and train 2000 rounds.

[0124] 3. Phase Three: Perform formal training on the end-to-end autoencoder as a whole, freeze the channel model, and conduct 12,000 training rounds.

[0125] 4. Phase Four: Post-training the carrier phase recovery module and constellation demapping module of the decoder, freezing the encoder and channel model, training 2000 rounds.

[0126] The training process is as follows:

[0127] First, the hyperparameters and the weights of the encoder and decoder are initialized, including an initial learning rate (lr) of 5e-4, a batch size of 8192, and the number of training epochs for each stage.

[0128] Training uses the AdamW optimizer and the OneCycleLR learning rate optimizer (the learning rate is increased for the first 30% and decreased for the next 70%, with the minimum learning rate reduced to 1% of the maximum value).

[0129] The autoencoder neural network is trained sequentially in stages. The final training result is then processed using a demodulated bit sequence. and the original bit sequence The bit error rate and generalized mutual information per bit are evaluated.

[0130] The formula for bit error rate:

[0131]

[0132] Formula for generalized mutual information per bit:

[0133]

[0134] Step 6: Place the trained encoder module in the TX-DSP module to convert the transmitted bit sequence into a geometrically shaped transmitted symbol sequence; after processing by the denoising neural network, the received symbol sequence is obtained; place the decoder in the RX-DSP module to demodulate the received symbol sequence into the original transmitted bit sequence.

[0135] Instead of traditional modulation and demodulation schemes, it can achieve geometric shaping and joint sensing, modeling and suppression of channel impairments, thereby optimizing the performance of photonic terahertz communication systems and significantly improving the quality and stability of signal transmission.

[0136] Experimental verification was conducted, taking 32QAM as an example, performing geometric shaping and channel impairment suppression to obtain the geometrically shaped constellation diagram, the equalized received signal constellation diagram, and the phase-recovered constellation diagram, as shown below. Figure 6 As shown, a bit error rate of 1.7e-3 and a performance of 0.9876 were achieved. The system's bit error rate performance is improved by 6.49 dB compared to the standard 32QAM.

[0137] The results indicate that:

[0138] This invention achieves high-fidelity, highly generalizable symbol-level differentiable modeling of end-to-end channels in photonic terahertz communication systems through a conditional denoising diffusion model. Compared to traditional analytical mathematical channel models and other deep learning channel models, it significantly improves sample generation quality, diversity, differentiability, training stability, and generalization ability.

[0139] This invention achieves end-to-end training of a trained differentiable channel model by embedding it into an autoencoder neural network. This enables geometric shaping and joint sensing, modeling, and suppression of channel impairments, thereby optimizing the performance of the photonic terahertz communication system. It significantly improves the quality and stability of signal transmission, resulting in superior system bit error rate performance and higher generalized mutual information.

Claims

1. An end-to-end impairment-suppressed photonic terahertz communication system based on channel modeling and geometry shaping, comprising a transmitter and a receiver; characterized in that, The transmitter includes a TX-DSP module with an encoder, which converts the transmitted bit sequence into a learnable geometric shape, generates a terahertz signal using photonic beat frequency, and transmits it to the receiver via single-mode fiber. The receiver includes an RX-DSP module containing a decoder, which performs digital signal processing on the received signal through a learnable carrier phase recovery and deconstellation mapping neural network to recover the original signal. The encoder and decoder together form an autoencoder neural network. The core of the encoder is a learnable (M,2)-sized lookup table used to learn the constellation point I / Q values ​​for MQAM geometric shaping. The encoder's input is a binary bit sequence. After converting it into an M-ary index sequence, the corresponding I / Q constellation point is looked up in the lookup table and a symbol sequence is formed. Then, the power of the entire sequence is normalized, and the transmitted symbol sequence is output. ; The core of the decoder consists of two parts: a carrier phase recovery module and a constellation demapping module; The carrier phase recovery module includes coarse phase compensation and fine phase compensation. The coarse phase compensation performs preliminary compensation on the initial received symbol sequence S0 received by the decoder to obtain the coarsely phase-compensated symbol S1. The fine phase compensation performs further compensation on the coarsely phase-compensated symbol sequence S1 to obtain the finely phase-compensated symbol S2. Each compensation includes one phase prediction and one phase recovery. Phase prediction consists of a context information acquirer and a phase predictor. The context information acquirer is composed of a multi-layer bidirectional long short-term memory network, which acquires the context information contained in the symbol sequence S0 or S1 of this compensation. The phase predictor is composed of a multi-layer perceptron, which collects the context information, and together with the symbol sequence of this compensation and the geometrically shaped constellation point information generated by the encoder, it performs compression and refinement processing and maps it to a rotation vector (cos(theta), sin(theta)). The phase recovery section uses rotation vector information to perform rotation compensation on the symbol sequence of this compensation. The constellation demapping module is responsible for performing geometric shaping and demapping on the symbol sequence S2 after it has been processed by the carrier phase recovery module, and then making a decision to obtain the demodulated bit sequence. This module consists of a multi-layer sensing mechanism.

2. The method as described in claim 1, characterized in that, The transmitting end also includes: two external cavity lasers ECL1 and ECL2, a TX-DSP module, an I / Q modulator, an optical coupler, a single-mode fiber, an erbium-doped fiber amplifier, and a single-row carrier photodetector. An external cavity laser ECL1 generates a continuous optical carrier and inputs it to an I / Q modulator. The I / Q modulator is also connected to a TX-DSP module, which uses a learnable autoencoder neural network to generate a geometrically shaped constellation point set and maps the transmitted bit signal sequence onto the shaped constellation point set. Through digital signal processing, a GS-MQAM baseband signal is obtained, which drives the I / Q modulator to modulate the input optical carrier. The modulated optical carrier is coupled to the continuous light wave output from the external cavity laser ECL2 through an optical coupler, transmitted through a single-mode fiber to an erbium-doped fiber amplifier for amplification, and finally input to a single-row carrier photodetector for beat frequency to obtain a terahertz signal.

3. The method as described in claim 1, characterized in that, The receiving end also includes: a mixer and a local oscillator; The terahertz signal is input into a mixer and mixed with the local oscillator signal generated by the local oscillator to achieve analog down-conversion of the signal and obtain an intermediate frequency signal. Then, it is input into the RX-DSP module, where digital down-conversion, low-pass filtering, downsampling, symbol synchronization, power normalization, channel equalization, and carrier phase recovery and geometric shaping demapping demodulation operations are performed sequentially by a decoder of a learnable autoencoder neural network. Finally, the original data is obtained, and the bit error rate is calculated.

4. The method as described in claim 1, characterized in that, The autoencoder neural network is embedded in channel modeling, and the specific process is as follows: First, collect historical complex sequence numbers of symbols from the transmitting and receiving ends, and divide them into training and testing sets; Then, the conditional denoising diffusion model is trained using the training set to obtain the denoising neural network weights of the channel model; The conditional denoising diffusion model includes: a forward diffusion module, a conditional vector processing module, and an inverse denoising module; The forward diffusion module is a non-neural network process, using the training set receiver symbol sequence. As input, a small amount of Gaussian noise is added at multiple time steps to obtain the final time step. Noisy data ; The conditional vector processing module adds noise time step information. , Global position information of each symbol in the entire sequence Training set sender symbol sequence The context information of each symbol before and after it Position encoding and merging are performed together, including the context information of the preceding and following symbols. An additional multi-head self-attention mechanism was implemented to learn the relationship between the different influences of different positions on the output; The information merging process uses a multilayer perceptron for learning to obtain a multi-source information conditional vector. ; Inverse denoising module to add noise to data and multi-source information conditional vector As input, a denoising neural network is trained, and the condition vector is processed through a feature-level linear modulation (FiLM) mechanism in the intermediate hidden layer. The impact on the output is further amplified, making it possible to achieve the same effect. Under control, the sample output generated by the denoising neural network and Consistent; Finally, the trained denoising neural network is embedded into the autoencoder neural network for end-to-end overall training.

5. The method as described in claim 4, characterized in that, The end-to-end overall training process is as follows: First, the hyperparameters and the weights of the encoder and decoder are initialized. The optimizer used for training is AdamW, and the learning rate optimizer is OneCycleLR. Then, following the order of encoder-denoising neural network-carrier phase recovery module-constellation demapping module, they are sequentially connected into a whole. Pre-training, training, and post-training are then performed in stages, with different modules participating in training at each stage. Modules not participating in training have their weights frozen. For the final training result, a demodulated bit sequence is used. and the original bit sequence The bit error rate and generalized mutual information per bit are evaluated. Finally, the trained encoder is placed in the TX-DSP module to convert the transmitted bit sequence into a geometrically shaped transmitted symbol sequence; after processing by a denoising neural network, the received symbol sequence is obtained; the decoder is placed in the RX-DSP module to demodulate the received symbol sequence into the original transmitted bit sequence.

6. The method as described in claim 5, characterized in that, The training tasks for each stage are as follows: 1) In Phase 1, the weights of the pre-trained encoder and carrier phase recovery module are frozen, and the weights of the denoising neural network and constellation demapping module are frozen, resulting in a short training cycle; 2) Phase 2: Pre-train the constellation demapping module, freeze the weights of the encoder, denoising neural network and carrier phase recovery module, and shorten the training rounds; 3) In the third stage, the entire autoencoder is formally trained end-to-end, while only the denoising neural network is frozen, resulting in a long training cycle. 4) Phase four: Post-training of carrier phase recovery module and constellation demapping module, freezing encoder and denoising neural network, with short training rounds.

Images