A dual-branch transformer-based phase noise adaptive compensation method for photonic terahertz OFDM systems

Abstract

This invention discloses an adaptive phase noise compensation method for a photonic terahertz OFDM system based on a dual-branch Transformer, belonging to the field of terahertz communication. Specifically, the transmitter performs constellation mapping on the bitstream, generates complex symbols, performs OFDM encoding, maps them onto frequency domain subcarriers, and adjusts the signal amplitude to output the transmitted signal to the receiver. The receiver performs channel estimation on the terahertz OFDM signal to obtain the channel frequency response. Based on this, it performs OFDM decoding on the received signal to obtain a signal with preliminary phase correction. Then, it inputs the signal to a phase noise neural network decoder to extract phase noise features, which are then input into a dual-branch processing network for parallel processing. This yields the phase noise-compensated signal output from the Transformer phase estimation branch and the enhanced signal output from the CNN signal enhancement branch. Adaptive fusion processing is then performed based on confidence scores. Finally, demapping is performed to obtain the final result. This invention mitigates the impact of phase noise on system performance and improves bit error rate performance.

Description

Technical Field

[0001] This invention belongs to the field of terahertz communication, specifically relating to an adaptive compensation method for phase noise in a photonic terahertz OFDM system based on a dual-branch Transformer. Background Technology

[0002] With the rapid development of services such as enhanced mobile broadband (eMBB), ultra-high-definition video, and virtual reality (VR), wireless communication networks are facing enormous pressure from the explosive growth of data traffic. The limited spectrum resources provided by existing microwave and millimeter-wave bands can no longer meet the stringent requirements of future sixth-generation mobile communication (6G) for Tbps-level peak rates. Exploring and developing wireless communication technologies that utilize higher frequency bands has become an inevitable choice for the industry.

[0003] Terahertz communication technology, as a key support for future 6G wireless communication, has received widespread attention from academic and industrial communities both domestically and internationally in recent years. Terahertz waves typically refer to electromagnetic waves with frequencies in the range of 0.1-10THz. This frequency band has abundant spectrum resources, providing continuous bandwidth of tens of GHz, supporting ultra-high-speed data transmission at the Tbps level, and effectively alleviating the spectrum congestion problem in the current microwave frequency band.

[0004] In the terahertz band, the 275-400 GHz band has become the focus of current research due to its relatively low atmospheric attenuation and good technical feasibility, especially the 400 GHz frequency point, which provides a good balance between atmospheric attenuation and available bandwidth.

[0005] In terms of technical approach, photonic terahertz technology, which generates terahertz signals through optoelectronic integration, has become the mainstream technology. One such approach involves external modulation, where an optical modulator modulates the baseband signal onto an optical carrier wave, which is then beat-frequencyd to generate a terahertz signal using a single-row carrier photodetector. This method can produce high-quality terahertz waveforms, but it faces a significant challenge due to phase noise. Phase noise primarily originates from multiple factors, including laser linewidth, photoelectric conversion nonlinearity, and terahertz modulator distortion. Its intensity increases significantly with frequency, severely impacting the performance of communication systems.

[0006] Orthogonal Frequency Division Multiplexing (OFDM) technology is widely used in terahertz communication systems due to its high spectral efficiency and resistance to multipath fading. However, in the terahertz band, OFDM systems are extremely sensitive to phase noise, which not only generates common phase error (CPE) but also causes severe inter-carrier interference (ICI). Traditional pilot-based CPE correction methods are insufficient to effectively suppress the ICI effect, leading to a sharp deterioration in system performance. Especially in high-speed mobile scenarios, the time-varying characteristics of phase noise are more pronounced, further increasing the difficulty of compensation.

[0007] To address the phase noise problem in terahertz OFDM systems, existing processing methods mainly fall into two categories: traditional signal processing methods and machine learning methods.

[0008] Traditional methods, such as phase tracking loops or frequency domain equalization, suffer from poor adaptability and slow convergence. While neural network-based methods can learn the statistical characteristics of phase noise, most struggle to effectively capture the long-term dependencies of phase noise and lack mechanisms for evaluating the reliability of the estimation results. These limitations restrict the application of existing technologies in terahertz communication systems, necessitating the development of new phase noise compensation schemes to meet the performance requirements of future terahertz communication systems. Summary of the Invention

[0009] To overcome the limitations of existing phase noise compensation methods in photonic terahertz OFDM systems, this invention provides an adaptive phase noise compensation method for photonic terahertz OFDM systems based on a dual-branch Transformer. Through dual-branch collaborative operation and a confidence-driven adaptive fusion mechanism, high-precision estimation and real-time compensation of phase noise are achieved, effectively improving the reliability and transmission performance of terahertz communication systems.

[0010] The specific steps of the adaptive phase noise compensation method for the photonic terahertz OFDM system based on the dual-branch Transformer are as follows:

[0011] Step 1: At the transmitter of the photonic terahertz OFDM system, the input bit stream is mapped using a 16QAM constellation to generate a complex symbol sequence;

[0012] Step 2: Perform OFDM encoding on the complex symbol sequence to obtain the time-domain signal;

[0013] Specifically:

[0014] First, the pilot sequence length pilot_length is calculated based on the OFDM system parameters. Then, a pilot symbol sequence is randomly generated and inserted into the complex symbol sequence at fixed intervals ind.

[0015] The formula for pilot sequence length is:

[0016] pilot_length = (Ns × 2 / M_carr × N_pilot) × BitPerSym;

[0017] Where Ns is the total number of complex symbols transmitted, M_carr is the number of effective subcarriers, M_eff is the number of effective data subcarriers, N_pilot is the number of pilot symbols, and BitPerSym is the number of bits per symbol;

[0018] The formula for calculating the fixed interval is: ind = M_eff / (N_pilot + 1);

[0019] Then, the complex symbols and pilot symbols are mapped onto the frequency domain subcarriers, and zero carriers and guard bands are inserted at both ends of the frequency domain subcarriers;

[0020] Next, the frequency domain signal with inserted zero carrier and guard band is subjected to inverse fast Fourier transform (IFFT) to convert it into a time domain signal.

[0021] Finally, a cyclic prefix (CP) is added before the time-domain signal, and a training sequence prefix is ​​generated;

[0022] Step 3: Quantize the OFDM encoded time-domain signal, adjust the signal amplitude, and output the transmitted signal to the receiver of the OFDM system.

[0023] The quantization process includes steps such as normalization, calculation of the quantization interval, quantization codebook, and quantization mapping.

[0024] The calculation formula is:

[0025] output = frame × Amp;

[0026] Where frame is the quantized output signal; Amp is the system parameter for adjusting the transmit power;

[0027] Step 4: Normalize and quantize the terahertz OFDM signals with center frequencies in the range of 275-400 GHz, and perform channel estimation to obtain the channel frequency response (ChanEsti).

[0028] The calculation is as follows:

[0029] ChanEsti = Rx_TrainSeq_FFT / FFT_TrainSeq;

[0030] Rx_TrainSeq_FFT is the terahertz OFDM signal sequence after removing the cyclic prefix and performing an FFT transformation.

[0031] FFT_TrainSeq is the terahertz OFDM signal sequence after FFT transformation;

[0032] Step 5: Based on the channel frequency response ChanEsti, perform OFDM decoding on the received signal output to obtain the signal output_data after preliminary phase correction;

[0033] Decoding includes:

[0034] First, the cyclic prefix is ​​removed according to its position, and the received signal output is converted to the frequency domain by Fast Fourier Transform (FFT) to obtain the frequency domain signal data_FFT;

[0035] Then, frequency domain equalization is performed on the frequency domain signal data_FFT using the channel frequency response ChanEsti to obtain the frequency domain equalization result temp_data:

[0036] temp_data = data_FFT / ChanEsti,

[0037] Next, the pilot symbols and data symbols are separated according to the positions of the pilot symbols inserted into the complex symbol sequence;

[0038] Finally, using the frequency domain equalization result temp_data combined with the separated pilot symbols, the phase rotation is calculated and preliminary phase correction is performed:

[0039] output_data = temp_data / rotat

[0040] Where rotat is the phase rotation factor obtained by the phase rotation factor algorithm;

[0041] Step 6: Input the output_data signal after preliminary phase correction into the phase noise neural network decoder to perform phase noise feature extraction processing;

[0042] The feature extraction process includes: extracting amplitude features from the signal output_data and calculating the phase difference features between adjacent symbols, wherein the phase difference features are wrapped around to the [-π, π] interval to ensure phase continuity.

[0043] The phase difference characteristic is calculated as follows: , For the signal output_data in The phase value at any given moment; For the signal output_data in The phase value at time; where, , These are the real and imaginary parts of the signal output_data at time t, respectively.

[0044] The phase wrapping processing formula is as follows: .

[0045] Step 7: Input the extracted phase noise features into a dual-branch processing network for parallel processing;

[0046] The dual-branch processing network includes a Transformer phase estimation branch and a CNN signal enhancement branch;

[0047] The processing procedure for the Transformer phase estimation branch is as follows:

[0048] First, the amplitude and phase difference features of the extracted signal output_data are mapped to a high-dimensional space through a fully connected network and sine and cosine position codes are added.

[0049] Then, sequence modeling is performed using a 4-6 layer Transformer encoder to learn from historical information and predict the current moment. Phase noise estimate And its confidence score x.

[0050] Each layer of the Transformer encoder contains a multi-head self-attention mechanism and a feedforward neural network, and uses causal masks to ensure real-time processing.

[0051] Next, the phase noise estimate is... Applied to the input signal output_data, phase rotation correction is performed through complex multiplication to generate a pre-compensated signal. :

[0052]

[0053] in For the signal output_data.

[0054] Finally, this branch outputs the pre-compensated signal. And the confidence score x, which characterizes the reliability of this compensation.

[0055] The processing procedure of the CNN signal enhancement branch is as follows: the signal decoded by OFDM is processed by a multi-level convolutional neural network for feature extraction and enhancement, including convolutional layers, batch normalization layers and activation function layers.

[0056] Step 8: Adaptively fuse the phase noise-compensated signal output from the Transformer phase estimation branch and the enhanced signal output from the CNN signal enhancement branch based on confidence scores.

[0057] The adaptive fusion process includes: setting a confidence threshold θ = 0.7; when the confidence score is higher than or equal to the threshold,

[0058] Transformer branch weight Wt = 0.3x + 0.4, where x ∈ [0.7, 1], CNN branch weight Wc = 1 - Wt; when the confidence score is lower than the threshold, Transformer branch weight Wt = 0.2x + 0.1, where x ∈ [0, 0.7), CNN branch weight Wc = 1 - Wt;

[0059] The fusion formula is: Output signal = Wt × Transformer branch output + Wc × CNN branch output.

[0060] Step 9: Perform 16QAM demapping on the adaptively fused communication signal and calculate parameters such as bit error rate based on the original bit stream.

[0061] The advantages of this invention are:

[0062] 1. This invention discloses an adaptive compensation method for phase noise in a photonic terahertz OFDM system based on a dual-branch Transformer. The method employs a dual-branch deep learning network based on Transformer for phase noise suppression. The Transformer branch can effectively capture long-term dependencies in the signal sequence, improving the accuracy of phase noise estimation. The CNN branch focuses on overall signal enhancement. The adaptive fusion mechanism of the two branches can dynamically adjust the processing strategy according to the estimated confidence, improving processing accuracy while ensuring stability.

[0063] 2. The present invention provides an adaptive compensation method for phase noise in a photonic terahertz OFDM system based on a dual-branch Transformer. Through an adaptive fusion mechanism implemented by a confidence index, the optimal processing strategy can be automatically selected under different signal-to-noise ratios and phase noise intensities, thereby improving the robustness and adaptability of the system.

[0064] 3. The present invention provides an adaptive compensation method for phase noise in a photonic terahertz OFDM system based on a dual-branch Transformer, which ensures real-time processing capability through causal attention masking and meets the real-time requirements of terahertz communication.

[0065] 4. The present invention provides an adaptive phase noise compensation method for a photonic terahertz OFDM system based on a dual-branch Transformer. In communication environments with strong phase noise, compared with the traditional pilot-based phase correction method, it can significantly reduce the bit error rate and improve the reliability of the communication system.

[0066] 5. The present invention provides an adaptive phase noise compensation method for a photonic terahertz OFDM system based on a dual-branch Transformer, which is seamlessly integrated with the existing OFDM receiver process and can be deployed without changing the system architecture, thus having strong engineering practicality. Attached Figure Description

[0067] Figure 1 This is a flowchart of an adaptive phase noise compensation method for a photonic terahertz OFDM system based on a dual-branch Transformer according to the present invention.

[0068] Figure 2 This is a schematic diagram of the adaptive phase noise compensation method for photonic terahertz OFDM systems based on dual-branch Transformer of the present invention.

[0069] Figure 3 This is a structural diagram of the photonic terahertz OFDM system based on the dual-branch Transformer of this invention;

[0070] Figure 4 This is a diagram illustrating the specific working process of the present invention based on adaptive phase noise compensation using a dual-branch Transformer.

[0071] Figure 5 This is a spectrum of the optically coupled components according to an embodiment of the present invention.

[0072] Figure 6 This is the constellation diagram recovered by the receiver of this invention. Detailed Implementation

[0073] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that the following embodiments are only for explaining the present invention and are not intended to limit the scope of protection of the present invention. Specific implementations made by those skilled in the art based on an understanding of the technical solutions of the present invention should all be included within the scope of protection of the present invention.

[0074] To address the complex phase noise introduced by laser linewidth, photoelectric conversion nonlinearity, and terahertz channel fluctuations in terahertz OFDM communication systems, this invention provides a solution that can simultaneously achieve high-precision estimation and real-time compensation. By combining a dual-branch Transformer architecture with a confidence-driven adaptive fusion mechanism, it effectively suppresses inter-carrier interference and improves the system's robustness in fast time-varying channels. This approach aligns with the current development trend of terahertz communication technology and has significant value for improving the performance of terahertz systems.

[0075] This invention achieves high-precision phase noise extraction at the receiver through the Transformer phase estimation branch and outputs a confidence evaluation index; at the same time, it uses the CNN signal enhancement branch to recover features of the distorted signal; finally, through a confidence-driven adaptive fusion mechanism, it dynamically weights and integrates the dual-branch outputs, effectively overcoming the limitations of traditional compensation methods in complex terahertz channels, and significantly improving the system's bit error rate performance and transmission reliability.

[0076] The adaptive phase noise compensation method for the photonic terahertz OFDM system based on the dual-branch Transformer is as follows: Figure 1 and Figure 2 As shown, the specific steps are as follows:

[0077] Step 1: At the transmitter of the photonic terahertz OFDM system, the input bit stream is mapped using a 16QAM constellation to generate a complex symbol sequence;

[0078] Step 2: Perform OFDM encoding on the complex symbol sequence to obtain the time-domain signal;

[0079] Specifically:

[0080] First, the pilot sequence length pilot_length is calculated based on the OFDM system parameters. Then, a pilot symbol sequence of size (0, 16) is randomly generated and inserted into the complex symbol sequence at a fixed interval ind.

[0081] The formula for pilot sequence length is:

[0082] pilot_length = (Ns × 2 / M_carr × N_pilot) × BitPerSym;

[0083] Where Ns is the total number of complex symbols transmitted, M_carr is the number of effective subcarriers, M_eff is the number of effective data subcarriers, N_pilot is the number of pilot symbols, and BitPerSym is the number of bits per symbol;

[0084] The formula for calculating the fixed interval is: ind = M_eff / (N_pilot + 1);

[0085] Then, the complex symbols and pilot symbols are mapped onto the frequency domain subcarriers, and zero carriers and guard bands are inserted at both ends of the frequency domain subcarriers;

[0086] Next, the frequency domain signal with inserted zero carrier and guard band is subjected to inverse fast Fourier transform (IFFT) to convert it into a time domain signal.

[0087] Finally, a cyclic prefix (CP) is added before the time-domain signal, and a training sequence prefix is ​​generated;

[0088] Step 3: Quantize the OFDM encoded time-domain signal, output the quantized signal frame, adjust the signal amplitude, and output the transmitted signal to the receiver of the OFDM system.

[0089] The quantization process includes steps such as normalization, calculation of the quantization interval, quantization codebook, and quantization mapping.

[0090] The calculation formula is:

[0091] output = frame × Amp;

[0092] Where frame is the quantized output signal; Amp is the system parameter for adjusting the transmit power;

[0093] Step 4: Normalize and quantize the terahertz OFDM signals with center frequencies in the range of 275-400 GHz, and perform channel estimation to obtain the channel frequency response (ChanEsti).

[0094] The calculation is as follows:

[0095] ChanEsti = Rx_TrainSeq_FFT / FFT_TrainSeq;

[0096] Rx_TrainSeq_FFT is the terahertz OFDM signal sequence after removing the cyclic prefix and performing an FFT transformation.

[0097] FFT_TrainSeq is the terahertz OFDM signal sequence after FFT transformation;

[0098] Terahertz OFDM signals with center frequencies in the range of 275-400 GHz are generated through photoelectric conversion and terahertz modulation; the preferred frequency is 400 GHz, at which the dual-branch Transformer architecture is optimized for the phase noise statistical characteristics unique to 400 GHz.

[0099] Step 5: Based on the channel frequency response ChanEsti, perform OFDM decoding on the received signal output to obtain the signal output_data after preliminary phase correction;

[0100] Decoding includes:

[0101] First, the cyclic prefix is ​​removed according to its position, and the received time-domain signal output is converted to the frequency domain by Fast Fourier Transform (FFT) to obtain the frequency domain signal data_FFT.

[0102] Then, frequency domain equalization is performed on the frequency domain signal data_FFT using the channel frequency response ChanEsti to obtain the frequency domain equalization result temp_data:

[0103] temp_data = data_FFT / ChanEsti,

[0104] Next, the pilot symbols and data symbols are separated according to the positions of the pilot symbols inserted into the complex symbol sequence;

[0105] Finally, using the frequency domain equalization result temp_data combined with the separated pilot symbols, the phase rotation is calculated and preliminary phase correction is performed:

[0106] output_data = temp_data / rotat

[0107] Where rotat is the phase rotation factor obtained by the phase rotation factor algorithm;

[0108] Step 6: Input the output_data signal after preliminary phase correction into the phase noise neural network decoder to perform phase noise feature extraction processing;

[0109] The feature extraction process includes: extracting amplitude features from the signal output_data and calculating the phase difference features between adjacent symbols, wherein the phase difference features are wrapped around to the [-π, π] interval to ensure phase continuity.

[0110] The phase difference characteristic is calculated as follows: , For the signal output_data in The phase value at any given moment; For the signal output_data in The phase value at time; where, , These are the real and imaginary parts of the signal output_data at time t, respectively.

[0111] The phase wrapping processing formula is as follows: .

[0112] Step 7: Input the extracted phase noise features into a dual-branch processing network for parallel processing;

[0113] The dual-branch processing network includes a Transformer phase estimation branch and a CNN signal enhancement branch;

[0114] The processing procedure for the Transformer phase estimation branch is as follows:

[0115] First, the amplitude and phase difference features of the extracted signal output_data are mapped to a high-dimensional space through a fully connected network and sine and cosine position codes are added.

[0116] Then, sequence modeling is performed using a 4-6 layer Transformer encoder to learn from historical information and predict the current moment. Phase noise estimate And its confidence score x.

[0117] Each layer of the Transformer encoder contains a multi-head self-attention mechanism and a feedforward neural network, and uses causal masks to ensure real-time processing.

[0118] Next, the phase noise estimate is... Applied to the input signal output_data, phase rotation correction is performed through complex multiplication to generate a pre-compensated signal. :

[0119]

[0120] in For the signal output_data.

[0121] Finally, this branch outputs the pre-compensated signal. The confidence score x, which characterizes the reliability of this compensation, provides a key decision-making basis for subsequent adaptive fusion.

[0122] The processing procedure of the CNN signal enhancement branch is as follows: the signal decoded by OFDM is processed by a multi-level convolutional neural network for feature extraction and enhancement, including convolutional layers, batch normalization layers and activation function layers.

[0123] Step 8: Adaptively fuse the phase noise-compensated signal output from the Transformer phase estimation branch and the enhanced signal output from the CNN signal enhancement branch based on confidence scores.

[0124] The adaptive fusion process includes: setting a confidence threshold θ = 0.7; when the confidence score is higher than or equal to the threshold,

[0125] Transformer branch weight Wt = 0.3x + 0.4, where x ∈ [0.7, 1], CNN branch weight Wc = 1 - Wt; when the confidence score is lower than the threshold, Transformer branch weight Wt = 0.2x + 0.1, where x ∈ [0, 0.7), CNN branch weight Wc = 1 - Wt;

[0126] The fusion formula is: Output signal = Wt × Transformer branch output + Wc × CNN branch output.

[0127] Step 9: Perform 16QAM demapping on the adaptively fused communication signal and calculate parameters such as bit error rate based on the original bit stream.

[0128] During training, real phase noise data is used for supervised learning, and the network parameters are optimized through the backpropagation algorithm.

[0129] The photonic terahertz OFDM system based on a dual-branch Transformer described in this invention, such as Figure 3 As shown, the basic structure mainly consists of a laser, an optical modulator, an arbitrary waveform generator, an erbium-doped fiber amplifier, an optical coupler, an optical attenuator, a single-row carrier photodetector, a terahertz antenna, and a receiver.

[0130] First, an external cavity laser generates a light source signal. A baseband signal generated by an arbitrary waveform generator is loaded onto the light source signal generated by laser 1 by an I / Q modulator, and then coupled with the optical signal generated by laser 2 at an optical coupler. The coupled optical signal is transmitted through a standard single-mode fiber, amplified by an erbium-doped fiber amplifier, and then subjected to heterodyne beat frequency generation by a single-row carrier photodetector to generate a photonic terahertz signal, which is emitted by a terahertz antenna as the carrier for wireless transmission. The receiving end receives the terahertz signal through a receiving terahertz antenna and provides some power compensation. Subsequently, the received signal is down-converted to an intermediate frequency (IF) signal, then mixed with a local oscillator signal of the same frequency, and passed through a low-pass filter to recover the baseband signal. Finally, the baseband signal undergoes digital signal processing and demodulation to recover the original data.

[0131] The photonic terahertz OFDM system consists of a transmitter and a receiver. The terahertz signal at the transmitter is generated by optical heterodyne beat frequency. When two waves of similar frequencies are added together, a sum frequency and a difference frequency signal are generated. By adjusting these two frequencies, terahertz waves can be generated within a specific frequency range. The signal is then transmitted through a 10km single-mode fiber to the receiver, where a dual-branch Transformer phase noise adaptive compensation method is used to compensate for the received signal.

[0132] The sending system architecture includes the following core components:

[0133] Two independently operating external cavity lasers (ECL1 and ECL2) are used. The output of ECL1 serves as the input optical carrier of the IQ modulator, and its linewidth characteristics directly affect the phase noise performance of the system. ECL2 provides local oscillator light for subsequent beat frequency processes.

[0134] The TX-DSP module's processing flow includes standard OFDM modulation steps such as pilot insertion and cyclic prefix addition.

[0135] Two parallel-configured electrical amplifiers (EAs) employ a differential drive architecture to amplify the I-channel and Q-channel signals, respectively.

[0136] An optical coupler (OC) coherently couples the modulated light wave output from the IQ modulator with the continuous light wave output from the ECL2.

[0137] A 10km single-mode optical fiber is used to transmit optical signals to an optical fiber amplifier.

[0138] Erbium-doped fiber amplifiers (EDFAs) are used to amplify optical signals transmitted through optical fibers.

[0139] Single-row carrier photodetectors (UTC-PDs) can effectively achieve frequency conversion from the optical domain to the terahertz domain.

[0140] During system operation, the optical carrier output from ECL1 is injected into the IQ modulator, while the 16QAM-OFDM signal generated by the TX-DSP is amplified by the EA and drives the modulator to complete optical modulation. The modulated optical signal is coupled with the local oscillator light output from ECL2 in the OC, generating an optical frequency difference. After being transmitted through a 10KM optical fiber, the coupled optical signal is amplified by the EDFA and finally input to the UTC-PD for photoelectric conversion to generate a terahertz wave signal of 275-400GHz. In this embodiment, 400GHz is selected.

[0141] The receiver system consists of three core units:

[0142] The mixer is responsible for downconverting the received terahertz signal to baseband.

[0143] The local oscillator (LO) provides a stable and frequency-adjustable local oscillator signal for the mixing process;

[0144] The RX-DSP module employs a phase noise adaptive compensation algorithm based on a two-branch Transformer for digital signal processing. This algorithm includes a Transformer phase estimation branch and a CNN signal enhancement branch, and adaptive weighted fusion of the outputs from the two branches is achieved through confidence evaluation. All units work collaboratively to achieve accurate reconstruction of damaged signals.

[0145] The OFDM terahertz signal is input into a mixer and mixed with the local oscillator signal generated by the local oscillator to achieve down-conversion of the signal and obtain the base frequency OFDM terahertz signal. Then it is input into the RX-DSP module, where OFDM demodulation, phase noise estimation and compensation based on dual-branch Transformer, signal enhancement and adaptive fusion, demapping and other operations are performed in sequence to finally obtain the original data.

[0146] Example:

[0147] This embodiment mainly utilizes two software programs: the optical simulation software VPIphotonics Design Suite 11.1 and Python 3.10. The overall system structure is built on VPI, and digital signal processing at the transmitting and receiving ends is performed on Python, increasing the flexibility of digital signal processing and realizing the transmission, reception, and transmission of 16QAM-OFDM signals.

[0148] The specific working process of the adaptive phase noise compensation based on the dual-branch Transformer is as follows: Figure 4 As shown, the specific process is as follows:

[0149] Step 1: Signal preprocessing and feature extraction;

[0150] Receive an input signal containing phase noise, wherein the input signal is a communication signal in complex form; extract amplitude features and phase difference features from the input signal, wherein the phase difference features are obtained by calculating the phase change of adjacent time steps and performing a wrap-around processing; specifically:

[0151] Let Yk[n] be the received signal of the nth symbol on the kth subcarrier after OFDM demodulation.

[0152] First, convert it into a time-domain sequence (which can be done through IFFT or by directly processing the symbol sequence of each subcarrier) to obtain the complex signal y[n] = x[n] * exp(j * φ[n]) + w[n] at each time step;

[0153] Where x[n] is the transmitted symbol, φ[n] is the phase noise to be estimated, and w[n] is additive white Gaussian noise.

[0154] Preprocessing steps include:

[0155] Amplitude feature extraction: A[n] = |y[n]|;

[0156] Instantaneous phase extraction: θ[n] = arg(y[n]) = arctan(Im(y[n]) / Re(y[n]));

[0157] Phase difference feature extraction and wrapping: Calculate the phase difference between adjacent symbols and wrap them to the [-π, π] interval to prevent phase jumps.

[0158] The amplitude A[n] and the wrapped phase difference Δθw[n] are combined into a feature vector F[n] = [A[n], Δθw[n]], forming a feature sequence {F[1], F[2], ..., F[L]}, which serves as the input to the Transformer branch.

[0159] Step 2: The Transformer phase estimation branch and the CNN signal enhancement branch are processed in parallel;

[0160] The amplitude and phase difference features are input into the Transformer phase estimation branch, and sequence modeling is performed through a multi-layer Transformer encoder to output the phase noise estimate and the corresponding confidence score.

[0161] The input signal is simultaneously fed into the CNN signal enhancement branch, and local features are extracted through the convolutional neural network to output the signal enhancement result.

[0162] The Transformer phase estimation branch includes: mapping 2D input features to a high-dimensional feature space through a fully connected network; and using sine and cosine functions for position encoding.

[0163] PE(pos,2i)=sin(pos / 10000^(2i / d_model))

[0164] PE(pos,2i+1)=cos(pos / 10000^(2i / d_model))

[0165] A 4-6 layer Transformer encoder is used, with each layer containing: a multi-head self-attention mechanism with 4-8 heads; a feedforward neural network with hidden layer dimensions of 128-256; layer normalization and residual connections; and causal attention masks to ensure real-time processing capabilities.

[0166] Specifically:

[0167] First, a fully connected layer maps the 2D feature F[n] to a high-dimensional space. Then, a sinusoidal positional code is added to each position in the sequence to inject the sequence's order information.

[0168] The sequence is fed into an N-layer Transformer encoder, each layer of which contains a multi-head self-attention mechanism and a feedforward neural network. Causal masks are used to ensure that when estimating the phase noise at time n, only information from time 1 to time n is seen, thus meeting the requirements for real-time processing.

[0169] After Transformer encoding, a feature sequence H[n] rich in contextual information is obtained.

[0170] The phase noise value of H[n] is estimated by passing it through a fully connected layer and the Tanh activation function, and then scaled to the range of [-π, π].

[0171] The confidence level of the estimate is generated by passing H[n] through another fully connected layer and a Sigmoid activation function.

[0172] CNN signal enhancement branch:

[0173] The real and imaginary parts of the original complex signal y[n] are input into the CNN branch.

[0174] This branch consists of multiple one-dimensional convolutional layers, batch normalization layers, and ReLU activation functions. It is used to extract local features of the signal and enhance its quality, outputting the enhanced signal features.

[0175] Step 3: Adaptive Fusion and Final Output

[0176] The fusion weights are dynamically calculated based on the confidence score. When the confidence score is higher than a preset threshold, the output of the Transformer branch is favored; when the confidence score is lower than the preset threshold, the output of the CNN branch is favored. The outputs of the two branches are then weighted and fused according to the fusion weights to generate a phase-compensated output signal. Specifically:

[0177] First, the phase noise estimated by the Transformer branch is used to initially compensate the original signal. Then, the fusion weights are dynamically calculated based on the confidence level x.

[0178] First, set a confidence threshold θ (e.g., θ = 0.7). If x >= θ (high confidence), then the phase compensation result is given more confidence: Wt = 0.3 * x + 0.4; if x <= θ (low confidence), then the CNN enhancement result is given more confidence: Wt = 0.2 * x + 0.1. Finally, the phase-compensated signal and the CNN-enhanced signal are weighted and fused to obtain the final compensated signal.

[0179] The compensated signal is sent to subsequent modules such as the demapper to ultimately recover the original data bit stream.

[0180] This example is based on the aforementioned phase noise adaptive compensation device, and includes:

[0181] A signal input interface for receiving complex signals containing phase noise;

[0182] The feature extraction module, connected to the signal input interface, is used to extract amplitude features and phase difference features;

[0183] The dual-branch processing module includes:

[0184] The Transformer phase estimation submodule includes a position coding unit and a multi-layer Transformer encoder;

[0185] The CNN signal enhancement submodule contains a multi-level convolutional neural network.

[0186] The adaptive fusion module dynamically adjusts the branch fusion weights based on the confidence score.

[0187] The signal output interface is used to output the phase-compensated signal.

[0188] At the transmitting end, laser 1 generates a continuous light wave CW1 with a wavelength of 1553.933 nm, an output power of 11.76 dBm, and a linewidth of 100 kHz. Laser 2 generates a continuous light wave CW2 with a wavelength of 1551.119 nm, an output power of 9.54 dBm, and a linewidth of 100 kHz. Figure 5 The image shows the spectra of the coupled two optical waves. The left side shows the output spectrum of the I / Q modulator, and the right side shows the output spectrum of CW2. The frequency difference between the two is 400 GHz.

[0189] The TX-DSP section uses a co-simulation module to jointly debug VPI and Python. The baseband signal generated by Python is sent to the VPI system and transmitted through the channel to the receiving end. At the receiving end, the received signal is directly mixed with the local oscillator signal of the same frequency, down-converted to baseband, and then sent to Python for digital signal processing.

[0190] The specific steps of digital signal processing in the TX-DSP module are as follows:

[0191] Step 1: Perform constellation mapping on the input bit stream to generate a sequence of complex symbol numbers.

[0192] Step 2: Generate pilot symbol sequence based on system parameters.

[0193] Step 3: OFDM encoding of the data symbol sequence: First, map the data symbols and pilot symbols onto frequency domain subcarriers to construct frequency domain OFDM symbols. Then, perform inverse fast Fourier transform (IFFT) on the frequency domain signal to convert it into a time domain signal. Finally, add a cyclic prefix (CP) to the time domain signal to obtain the OFDM encoded signal.

[0194] Step 4: Quantize the OFDM encoded signal, adjust the signal amplitude, and then output the transmitted signal.

[0195] The baseband signal generated by the TX-DSP module is sent to the system and split into I and Q paths. It is amplified by two parallel amplifiers with a gain of 25dB, then modulated onto the terahertz light wave generated by CW1 in the IQ modulator and coupled to CW2. The coupled signal passes through a 10km single-mode fiber and is amplified by an erbium-doped fiber amplifier. Then, it is beat-frequencyd by a UTC-PD to obtain a terahertz wave signal with a bandwidth of 275-400GHz. This signal is then fed into the RX-DSP and mixed with a 400GHz local oscillator signal, down-converted to the baseband frequency. Subsequent steps include downsampling, obtaining the preamble training sequence, channel estimation, synchronization, cyclic prefix removal, fast Fourier transform, and coarse compensation for amplitude and phase distortion to complete OFDM demodulation. The OFDM decoded signal is then sent to a phase noise adaptive compensation module for noise depth compensation.

[0196] The specific method of the phase noise adaptive compensation module is as follows:

[0197] The OFDM decoded signal was divided into training and test sets and fed into a two-branch neural network, with 80% of the data used as the training set and 20% as the test set. The model was trained for 300 epochs. The Adam algorithm was used as the optimizer.

[0198] The parameters are configured as follows: learning rate 0.001, first-order moment decay factor 0.9, and second-order moment decay factor 0.999.

[0199] The activation functions used are ReLU, TanH, and Sigmoid, which are used for feature extraction, phase output, and confidence evaluation, respectively.

[0200] The loss function is designed as a multi-task combination: the main loss function uses mean squared error loss (MSELoss) to calculate the signal reconstruction error, and the auxiliary loss function also uses MSELoss to calculate the phase estimation error. The total loss is weighted and summed with a weight ratio of 1:0.5. During training, Dropout (dropout rate of 0.1) and gradient clipping (threshold of 1.0) are used to prevent overfitting.

[0201] The Transformer branch uses a multi-head self-attention mechanism to extract temporal features of phase noise and outputs a phase estimate and confidence score; the CNN branch uses a convolutional neural network for signal enhancement. Finally, the dual-branch outputs are dynamically weighted and fused based on the confidence score (with a confidence threshold set to 0.7) to complete phase noise compensation.

[0202] Finally, the deep-compensated signal is demodulated using 16-QAM to recover the original bit sequence.

[0203] The results indicate: Figure 6As shown, the constellation diagram of the signal recovered by the receiver under the conditions of OSNR of 30dB and fiber length of 10KM is presented. (a) is the constellation diagram without dual-branch Transformer phase noise compensation, and (b) is the constellation diagram after demodulation with dual-branch Transformer phase noise compensation. The comparison shows that the dual-branch Transformer phase noise compensation technique significantly optimizes the demodulation performance. Without compensation, noise causes symbol spread, while after compensation, the point set converges, and the model successfully estimates and compensates for the phase rotation. The bit error rate decreases from 4.0674e-4 to 5.9737e-5, demonstrating the effectiveness of the algorithm in suppressing phase noise.

Claims

1. A method for adaptive phase noise compensation in a photonic terahertz OFDM system based on a dual-branch Transformer, characterized in that, The specific steps are as follows: Step 1: At the transmitter of the photonic terahertz OFDM system, the input bit stream is mapped using a 16QAM constellation to generate a complex symbol sequence; Step 2: Perform OFDM encoding on the complex symbol sequence to obtain the time-domain signal; Step 3: Quantize the OFDM encoded time-domain signal, adjust the signal amplitude, and output the transmitted signal to the receiver of the OFDM system. Step 4: Normalize and quantize the terahertz OFDM signals with center frequencies in the range of 275-400 GHz, and perform channel estimation to obtain the channel frequency response (ChanEsti). Step 5: Based on the channel frequency response ChanEsti, perform OFDM decoding on the received signal output to obtain the signal output_data after preliminary phase correction; Step 6: Input the output_data signal after preliminary phase correction into the phase noise neural network decoder to perform phase noise feature extraction processing; The feature extraction process includes: extracting amplitude features from the signal output_data and calculating the phase difference features between adjacent symbols, wherein the phase difference features are wrapped around to the [-π, π] interval to ensure phase continuity. Step 7: Input the extracted phase noise features into a dual-branch processing network for parallel processing; The dual-branch processing network includes a Transformer phase estimation branch and a CNN signal enhancement branch; The processing procedure for the Transformer phase estimation branch is as follows: First, the amplitude and phase difference features of the extracted signal output_data are mapped to a high-dimensional space through a fully connected network and sine and cosine position codes are added. Then, sequence modeling is performed using a 4-6 layer Transformer encoder to learn from historical information and predict the current moment. Phase noise estimate and its confidence score x; Each layer of the Transformer encoder contains a multi-head self-attention mechanism and a feedforward neural network, and uses causal masks to ensure real-time processing. Next, the phase noise estimate is... Applied to the input signal output_data, phase rotation correction is performed through complex multiplication to generate a pre-compensated signal. : ； in For the signal output_data; Finally, this branch outputs the pre-compensated signal. And the confidence score x, which characterizes the reliability of this compensation; The processing steps of the CNN signal enhancement branch are as follows: the signal decoded from OFDM is processed by a multi-level convolutional neural network for feature extraction and enhancement, including convolutional layers, batch normalization layers and activation function layers; Step 8: Adaptively fuse the phase noise-compensated signal output from the Transformer phase estimation branch and the enhanced signal output from the CNN signal enhancement branch based on confidence scores. The adaptive fusion process includes: setting a confidence threshold θ = 0.7; When the confidence score is higher than or equal to the threshold, the Transformer branch weight W t =0.3x+0.4 where x∈[0.7,1], CNN branch weights W c = 1 - W t ; When the confidence score is below the threshold, the Transformer branch weight W t = 0.2x + 0.1, where x∈[0,0.7). CNN branch weights W c = 1 - W t ; The fusion formula is: Output signal = W t × Transformer branch output + W c × CNN branch output; Step 9: Perform 16QAM demapping on the adaptively fused communication signal and calculate the bit error rate based on the original bit stream.

2. The method according to claim 1, characterized in that, Step two specifically involves: First, the pilot sequence length pilot_length is calculated based on the OFDM system parameters. Then, a pilot symbol sequence is randomly generated and inserted into the complex symbol sequence at fixed intervals ind. The formula for pilot sequence length is: pilot_length = (Ns × 2 / M_carr × N_pilot) × BitPerSym; Where Ns is the total number of complex symbols transmitted, M_carr is the number of effective subcarriers, M_eff is the number of effective data subcarriers, N_pilot is the number of pilot symbols, and BitPerSym is the number of bits per symbol; The formula for calculating the fixed interval is: ind = M_eff / (N_pilot + 1); Then, the complex symbols and pilot symbols are mapped onto the frequency domain subcarriers, and zero carriers and guard bands are inserted at both ends of the frequency domain subcarriers; Next, the frequency domain signal with inserted zero carrier and guard band is subjected to inverse fast Fourier transform (IFFT) to convert it into a time domain signal; Finally, a cyclic prefix (CP) is added before the time-domain signal to generate a training sequence prefix.

3. The method according to claim 1, characterized in that, In step three, the quantization process includes normalization, calculation of quantization interval, quantization codebook, and quantization mapping; The formula for calculating the transmitted signal is: output = frame × Amp; Where frame is the output signal after quantization; Amp is the system parameter for adjusting the transmit power.

4. The method according to claim 1, characterized in that, In step four, the channel frequency response ChanEsti is calculated as follows: ChanEsti = Rx_TrainSeq_FFT / FFT_TrainSeq; Rx_TrainSeq_FFT is the terahertz OFDM signal sequence after removing the cyclic prefix and performing an FFT transformation. FFT_TrainSeq is the terahertz OFDM signal sequence after FFT transformation.

5. The method according to claim 1, characterized in that, In step five, decoding includes: First, the cyclic prefix is ​​removed according to its position, and the received signal output is converted to the frequency domain by a fast Fourier transform to obtain the frequency domain signal data_FFT; Then, frequency domain equalization is performed on the frequency domain signal data_FFT using the channel frequency response ChanEsti to obtain the frequency domain equalization result temp_data: temp_data = data_FFT / ChanEsti, Next, the pilot symbols and data symbols are separated according to the positions of the pilot symbols inserted into the complex symbol sequence; Finally, using the frequency domain equalization result temp_data combined with the separated pilot symbols, the phase rotation is calculated and preliminary phase correction is performed: output_data = temp_data / rotat Where rotat is the phase rotation factor obtained through the phase rotation factor algorithm.

6. The method according to claim 1, characterized in that, In step six, the phase difference characteristic is calculated as follows: ； For the signal output_data in The phase value at any given moment; For the signal output_data in The phase value at time; where, , These are the real and imaginary parts of the signal output_data at time t, respectively. The phase wrapping processing formula is as follows: .

Images