Channel estimation method based on super-resolution optical UFMC system and receiving end

By performing least-squares estimation and super-resolution network reconstruction on the block pilot symbols of the UFMC system, the contradiction between high accuracy and low complexity in channel estimation in the UFMC system is resolved, improving the accuracy of channel reconstruction and receiver sensitivity. This method is suitable for optical communication scenarios with high-order modulation and dense subcarriers.

CN123123076APending Publication Date: 2026-11-17SUZHOU UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611261033.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-19
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing UFMC systems face a trade-off between high accuracy and low complexity in channel estimation. In particular, in the case of block pilots, traditional methods struggle to account for anisotropic structures, resulting in large channel reconstruction errors and difficulty in reducing the bit error rate.

Method used

A super-resolution-based channel estimation method is adopted. By performing least-squares estimation on block pilot symbols, a coarse channel frequency response is obtained. Then, a super-resolution network is used to reconstruct the channel in super-resolution. Frequency domain equalization and QAM demapping are performed in combination with the global channel frequency response, which is suitable for high-order modulation and dense subcarrier scenarios.

Benefits of technology

It achieves high-precision channel estimation, reduces computational complexity, improves receiver sensitivity, and is suitable for high-order modulation and dense subcarrier loading applications such as 64-QAM and 256-QAM, while reducing bit error rate.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123123076A_ABST
    Figure CN123123076A_ABST
Patent Text Reader

Abstract

This invention belongs to the field of optical communication technology and relates to a channel estimation method and receiver based on a super-resolution optical UFMC system. The method involves performing time-domain preprocessing and UFMC demodulation on the UFMC signal to obtain frequency-domain UFMC symbols, thereby extracting block pilot symbols arranged at multiple block pilot positions. Least-squares channel estimation is performed on each block pilot symbol to obtain a coarse channel frequency response corresponding to each block pilot symbol. The real and imaginary parts of all coarse channel frequency responses are stacked along the channel dimension to obtain a low-resolution channel matrix. The low-resolution channel matrix is ​​input into a super-resolution network for channel super-resolution reconstruction to obtain a high-resolution channel matrix. The real and imaginary channels of the high-resolution channel matrix are split to obtain a global channel frequency response. Based on the global channel frequency response, frequency-domain equalization compensation and orthogonal amplitude modulation demapping are performed on the effective data subcarriers within the frame to obtain the demodulated signal.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of optical communication technology, and in particular to a channel estimation method and receiver based on a super-resolution optical UFMC system. Background Technology

[0002] Intensity Modulation / Direct Detection (IM / DD) optical transmission is widely used in short-range interconnects, mobile fronthaul, and Passive Optical Network (PON) access scenarios due to its simple receiver structure and low cost. Multicarrier modulation, by dividing a bandwidth-constrained channel into multiple narrowband sub-signals, can achieve frequency domain equalization with moderate complexity and adapt to frequency-selective channels. Orthogonal Frequency Division Multiplexing (OFDM) is the most widely used multicarrier format, but its rectangular pulses lead to high out-of-band leakage, and the cyclic prefix (CP) overhead reduces spectral efficiency. Universal Filtered Multi-Carrier (UFMC) filters each sub-band separately, suppressing out-of-band radiation and improving spectral locality without using CP, making it more suitable for optical access links with bursty traffic and dynamic sub-band allocation.

[0003] However, the subband filtering that gives UFMC spectral advantages also significantly increases the receiver's dependence on the accuracy of Channel Frequency Response (CER) reconstruction. The aggregated channel response includes various effects such as bandwidth-limited devices, subband filter response, dispersion, subband superposition, and residual inter-symbol interference and inter-carrier interference (ISI / ICI). With the increase in the number of effective data subcarriers and the use of higher-order QAM modulation, these non-ideal effects have a particularly significant adverse impact on transmission performance. Channel estimation is a key technology for receiver reconstructing the channel frequency response, and the estimation results are directly used to compensate for various channel impairments. Due to the complex aggregated channel of UFMC, traditional channel estimation schemes struggle to achieve high-precision channel acquisition.

[0004] Pilot-assisted channel estimation (CFR) is a conventional core method for reconstructing channel frequency response and compensating for transmission impairments in multi-carrier optical transmission systems. Currently, the mainstream channel estimation methods applied to UFMC systems mainly include three categories: traditional least squares estimation, linear least mean square error estimation, and deep learning channel estimation. Traditional least squares estimation is computationally simple, but it directly introduces pilot noise and local distortion into CFR estimation. Linear least mean square error estimation utilizes channel correlation to suppress some noise, but it requires accurate channel statistics and noise variance that are difficult to reliably obtain in actual optical links; statistical mismatch can still cause significant performance degradation. Deep learning channel estimation typically first upsamples the sparse pilot CFR to a high-resolution size using fixed interpolation (such as bilinear interpolation), and then refines the residuals using a convolutional neural network on the high-resolution feature map. The method performs full convolution on high-resolution feature maps, resulting in high computational complexity. Furthermore, the fixed interpolation upsampling method is insufficient to fully recover the channel structure in the time dimension. In addition, when block pilots are used, since each pilot symbol covers all effective data subcarriers, the pilot observation is complete in the subcarrier dimension, but sparse only in the time dimension corresponding to the UFMC symbol. Therefore, the reconstruction problem has significant anisotropy in both the subcarrier and time dimensions. Unlike the two-dimensional interpolation problem where both frequency and time are missing, when the reconstruction is inaccurate due to this anisotropy, residual amplitude and phase errors will persist throughout the equalization process. This structural problem will simultaneously affect traditional least squares estimation, linear least mean square error estimation, and deep learning channel estimation, resulting in residual amplitude and phase errors in channel reconstruction, making it difficult for the bit error rate to decrease with the increase of received optical power.

[0005] In summary, designing a channel estimation method that is suitable for block pilot UFMC, fits anisotropic structures, and combines high accuracy with low complexity is an urgent problem to be solved. Summary of the Invention

[0006] Therefore, the technical problem to be solved by this invention is to design a channel estimation method that is suitable for block pilot UFMC, fits anisotropic structure, and has both high accuracy and low complexity.

[0007] To address the aforementioned technical problems, this invention provides a channel estimation method based on a super-resolution optical UFMC system, applied to the receiver of a super-resolution optical UFMC system, comprising: Each received UFMC signal frame undergoes time-domain preprocessing and UFMC demodulation to obtain the entire frame's frequency-domain UFMC symbols. The symbols arranged in the frequency domain are then extracted from these symbols. Block pilot symbols at block pilot locations; where each block pilot symbol covers K valid data subcarriers, where K represents the total number of valid data subcarriers in the frame; Least-square channel estimation is performed on each block pilot symbol to obtain the coarse channel frequency response corresponding to each block pilot symbol; the real and imaginary parts of all coarse channel frequency responses are stacked along the channel dimension to obtain a low-resolution channel matrix. The low-resolution channel matrix is ​​input into the super-resolution network for channel super-resolution reconstruction to obtain the high-resolution channel matrix. The real and imaginary channels of the high-resolution channel matrix are split to obtain a channel covering the entire frame. The global channel frequency response of each frequency domain UFMC symbol; where... This represents the temporal upsampling factor of the anisotropic upsampling module in a super-resolution network. Based on the global channel frequency response, frequency domain equalization compensation and orthogonal amplitude modulation demapping are performed on the effective data subcarriers within the frame to obtain the demodulated frame signal.

[0008] Preferably, the time-domain preprocessing of each received UFMC signal frame includes performing third-order Volterra time-domain equalization on each received UFMC signal frame.

[0009] Preferably, the coarse channel frequency response corresponding to the nth block pilot symbol for: , in, This represents the diagonal matrix formed by the nth block of pilot symbols; This represents the nth block pilot symbol.

[0010] Preferably, the super-resolution network includes a head convolutional layer, multiple cascaded residual blocks, an anisotropic upsampling module, a reconstruction convolutional layer, a global skip connection branch, and a fusion module.

[0011] Preferably, the low-resolution channel matrix is ​​input into the super-resolution network for channel super-resolution reconstruction to obtain a high-resolution channel matrix, including: The low-resolution channel matrix is ​​input into the head convolutional layer, thereby optimizing the time dimension. The low-resolution feature map is convolved to output the initial feature map; The initial feature map is input into multiple concatenated residual blocks, each residual block having a time dimension length of [missing information]. The input feature map is convolved to output the target feature map; The target feature map is input into the anisotropic upsampling module, thereby reducing the time dimension length of the target feature map from... Upsampling to Output the upsampled feature map; The upsampled feature map is input into the reconstruction convolutional layer for convolution, and the reconstructed feature map is output. The low-resolution channel matrix is ​​input into the global skip connection branch, thereby performing linear interpolation of the low-resolution channel matrix along the time dimension and outputting a high-resolution observation matrix with the same dimension as the reconstructed feature map. The reconstructed feature map and the high-resolution observation matrix are input into the fusion module and fused element by element to output a high-resolution channel matrix.

[0012] Preferably, each residual block includes a first convolutional layer, a first normalization layer, an activation function, a second convolutional layer, and a second normalization layer connected in series; wherein, the input of the first convolutional layer and the output of the second normalization layer are added element-wise to obtain the output of the residual block.

[0013] Preferably, the head convolutional layer is a convolution used to map the real and imaginary input channels of the low-resolution channel matrix into C feature channels; The number of feature channels in each residual block is C; The anisotropic upsampling module includes a transposed convolution, where the number of input and output channels is C. The kernel size of the transposed convolution has dimensions of 1 in the subcarrier dimension and 1 in the time dimension. The step sizes in the subcarrier dimension and the time dimension are 1 and 1, respectively. Both the fill and output fill are 0; The reconstructed convolutional layer is a convolution used to map C feature channels to 2 output channels.

[0014] Preferably, before inputting the low-resolution channel matrix into the super-resolution network for channel super-resolution reconstruction to obtain the high-resolution channel matrix, the method further includes: Preset block pilot symbols are inserted at all frequency domain UFMC symbol positions in the training frames; Channel estimation is performed on all frequency domain UFMC symbol locations in the training frame to obtain the channel frequency response covering all frequency domain UFMC symbols in the entire frame, which serves as a high-resolution label; Extract from high-resolution tags The channel frequency response corresponding to each block pilot position is used to construct a low-resolution input matrix; The low-resolution input matrix is ​​input into the super-resolution network to be trained, and the output is a reconstructed high-resolution channel matrix. A loss function is constructed using the normalized mean square error between the reconstructed high-resolution channel matrix and the high-resolution label. This loss function is then used to iteratively train the super-resolution network until the value of the loss function converges, resulting in a well-trained super-resolution network.

[0015] Preferably, UFMC signals with different modulation schemes correspond to different super-resolution networks.

[0016] The present invention also provides a receiving end, comprising: The pilot extraction module performs time-domain preprocessing and UFMC demodulation on each frame of the received UFMC signal to obtain the full-frame frequency-domain UFMC symbols, and extracts the pilot symbols arranged in the frequency domain from the full-frame frequency-domain UFMC symbols. Block pilot symbols at block pilot locations; where each block pilot symbol covers K valid data subcarriers, where K represents the total number of valid data subcarriers in the frame; The low-resolution channel matrix construction module is used to perform least-squares channel estimation on each block pilot symbol to obtain the coarse channel frequency response corresponding to each block pilot symbol; the real and imaginary parts of all coarse channel frequency responses are stacked along the channel dimension to obtain the low-resolution channel matrix. The channel super-resolution reconstruction module is used to input the low-resolution channel matrix into the super-resolution network for channel super-resolution reconstruction to obtain a high-resolution channel matrix. The global channel frequency response acquisition module is used to split the real and imaginary channels of the high-resolution channel matrix to obtain a signal covering the entire frame. The global channel frequency response of each frequency domain UFMC symbol; where... This represents the temporal upsampling factor of the anisotropic upsampling module in a super-resolution network. The equalization and demapping module is used to perform frequency domain equalization compensation and orthogonal amplitude modulation demapping on the effective data subcarriers within the frame based on the global channel frequency response, so as to obtain the demodulated frame signal.

[0017] The channel estimation method based on a super-resolution optical UFMC system provided in this application has the following advantages: 1. By extracting block pilot symbols and performing least-squares estimation, a coarse channel frequency response is obtained. The real and imaginary parts are stacked along the channel dimension to obtain a low-resolution channel matrix, which is then fed into a super-resolution network for anisotropic channel reconstruction. The subcarrier dimension is maintained intact, and only the time dimension is upsampled. This matches the inherent structure of the block pilot with its intact subcarrier dimension and sparse time dimension, thus recovering the full-domain channel frequency response covering the entire frame to complete frequency domain equalization and QAM demapping. By fully utilizing the spatiotemporal distribution characteristics of the block pilot itself, the complete channel frequency response of the entire frame can be recovered. It is compatible with time-domain Volterra equalization and adaptable to high-order modulation such as 64-QAM and 256-QAM, as well as application scenarios with dense subcarrier loading. Compared with traditional channel estimation schemes such as LS and LMMSE, it effectively improves the receiving sensitivity. 2. The head convolutional layer and residual block of the super-resolution network are operated on small-sized, low-resolution feature maps. Only transposed convolution is used to achieve temporal upsampling, which significantly reduces the amount of multiplication operations under the condition of comparable parameter quantity. The pure feedforward parallel inference mode has a latency better than the RLS recursive class algorithm. At the same time, the complex channel is jointly reconstructed by real and imaginary dual channels. Combined with global skip connections, it takes into account the overall trend of the channel and the local amplitude and phase changes, further ensuring the reconstruction accuracy of channel amplitude and phase. While achieving high-precision channel estimation, the computational overhead is greatly reduced, which effectively solves the practical needs of high-precision, low-complexity channel estimation in block pilot optical UFMC systems. Attached Figure Description

[0018] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein: Figure 1 This application provides an end-to-end structural block diagram of the IM / DD optical UFMC system. Figure 2 A schematic diagram of the experimental setup for the 25-GBd IM / DD UFMC transmission system provided in this application embodiment; Figure 3 Flowchart of the channel estimation method based on super-resolution optical UFMC system provided in this application; Figure 4 A schematic diagram of the super-resolution network structure provided in this application; Figure 5 Bit error rate curves for a 64-QAM UFMC provided in this application embodiment with and without Volterra equalization under back-to-back conditions; Figure 6 Bit error rate curves of a 64-QAM UFMC transmitted over 30-km single-mode fiber with and without Volterra equalization, provided in the embodiments of this application; Figure 7 A comparison chart of subcarrier signal-to-noise ratios for different channel estimation methods provided in the embodiments of this application; Figure 8 A comparison diagram of the amplitude and phase response of the estimated channel for different channel estimation methods provided in the embodiments of this application; Figure 9 The receiving constellation diagram of 64-QAM under different channel estimation methods provided in the embodiments of this application; wherein, Figure 9 (a) in the diagram is the receiver constellation diagram for 64-QAM under the least squares channel estimation (LS) method. Figure 9 (b) in the diagram is the receiver constellation diagram for 64-QAM under the Linear Minimum Mean Square Error (LMMSE) channel estimation method. Figure 9(c) in the diagram represents the receiver constellation diagram for 64-QAM under the CNN-CE method for channel estimation based on convolutional neural networks. Figure 9 (d) in the diagram represents the receiver constellation diagram of 64-QAM under the proposed SR channel estimation method provided in this application. Figure 10 Bit error rate curves of the 256-QAM UFMC provided in the embodiments of this application with and without Volterra equalization under back-to-back conditions; Figure 11 Bit error rate curves of a 256-QAM UFMC provided in this application embodiment, after transmission through a 30-km single-mode fiber, with and without Volterra equalization; Figure 12 This application provides a receiver constellation diagram for 256-QAM under different channel estimation methods, as shown in the embodiments of this application; wherein, Figure 12 (a) in the diagram is the receiver constellation diagram of 256-QAM under the least squares channel estimation (LS) method. Figure 12 (b) in the diagram is the receiver constellation diagram for 256-QAM under the Linear Minimum Mean Square Error (LMMSE) channel estimation method. Figure 12 (c) in the diagram represents the receiver constellation diagram for 256-QAM under the CNN-CE method for channel estimation based on convolutional neural networks. Figure 12 In the diagram, (d) represents the receiver constellation diagram of 256-QAM under the proposed SR channel estimation method provided in this application. Detailed Implementation

[0019] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0020] Please see Figure 1 , Figure 1 The diagram shows the end-to-end block diagram of the IM / DD optical UFMC system provided in this application. The transmitter maps the input bits to QAM symbols and allocates them to multiple UFMC subbands. Each subband is processed by an inverse fast Fourier transform (IFFT) and a dedicated subband filter, then superimposed to generate the generated UFMC waveform. The receiver first performs time-domain preprocessing on the received signal (including synchronization, waveform reconstruction, and optional three-section Volterra time-domain equalization), then zero-fills it and performs a 2N-point FFT transform to the frequency domain. UFMC demodulation obtains the received subcarrier symbols by extracting the even-indexed FFT output, which are used for pilot extraction and data equalization. The channel estimation module reconstructs the dense CFR from the super-resolution network and uses it for frequency-domain equalization, followed by QAM demapping.

[0021] Figure 2This is a schematic diagram of the experimental setup for the 25-GBd IM / DD UFMC transmission system provided in this application embodiment. The transmitter generates a UFMC signal using a DSP, employing a 512-point IFFT, 16 subbands, 16 subcarriers per sub-generation, and 256 effective data subcarriers. Both 64-QAM and 256-QAM modulation formats are considered. Each sub-generation is shaped by a Chebyshev filter with a length of 43 and a sidelobe attenuation of 43dB. The complex-valued UEMC waveform is converted to a real-valued signal through a concatenation operation and then uploaded to an arbitrary bandwidth of approximately 10GHz and a sampling rate of 50GSa / s. The waveform generator uses a continuous wave laser with a center wavelength of 1550.116nm. The modulated optical signal is measured back-to-back or transmitted through a 30-km standard single-mode fiber. The received optical power is adjusted by a VOA-EDFA-VOA link. The receiver is detected by a 10GHz photodetector and acquired by a 50GSa / s real-time oscilloscope. The offline receiving DSP performs inverse paralleling, synchronization, UFMC demodulation, 2N-point FFT, pilot extraction, channel estimation, frequency domain equalization, QAM demapping, and bit error rate calculation.

[0022] Please see Figure 3 , Figure 3 The diagram shows a flowchart of the channel estimation method based on a super-resolution optical UFMC system provided in this application. This method is applied to the receiver of a super-resolution optical UFMC system and includes: S10~S50.

[0023] S10: Perform time-domain preprocessing and UFMC demodulation on each received UFMC signal frame to obtain the full-frame frequency-domain UFMC symbols, and extract the symbols arranged in the frequency domain from the full-frame frequency-domain UFMC symbols. A block pilot symbol at each block pilot position; wherein each block pilot symbol covers K valid data subcarriers, where K represents the total number of valid data subcarriers in the frame.

[0024] Furthermore, the time-domain preprocessing of each received UFMC signal frame includes performing third-order Volterra time-domain equalization on each received UFMC signal frame.

[0025] For example, block pilot symbols Each pilot symbol covers all K=256 valid data subcarriers. Therefore, pilot observations are complete in the subcarrier dimension but sparse in the time dimension.

[0026] S20: Perform least-squares channel estimation on each block pilot symbol to obtain the coarse channel frequency response corresponding to each block pilot symbol; stack the real and imaginary parts of all coarse channel frequency responses along the channel dimension to obtain the low-resolution channel matrix. . Its dimensions are 2×256×8.

[0027] Specifically, the coarse channel frequency response corresponding to the nth block pilot symbol for: , in, This represents the diagonal matrix formed by the nth block of pilot symbols; This represents the nth block pilot symbol.

[0028] S30: Input the low-resolution channel matrix into the super-resolution network to perform channel super-resolution reconstruction and obtain the high-resolution channel matrix.

[0029] Furthermore, such as Figure 4 The diagram shown is a schematic of the super-resolution network structure provided in this application. The super-resolution network includes a head convolutional layer, multiple cascaded residual blocks, an anisotropic upsampling module, a reconstruction convolutional layer, a global skip connection branch, and a fusion module.

[0030] Based on the above network structure, S30 includes: S300~S305.

[0031] S300: Input the low-resolution channel matrix into the head convolutional layer, thereby optimizing the time dimension. The low-resolution feature map is convolved to output the initial feature map.

[0032] Specifically, the header convolutional layer is used to map the real and imaginary input channels of the low-resolution channel matrix into C feature channels. For example, the header convolutional layer is a 3×3 convolution, which maps the two input channels (real and imaginary) into C=64 feature channels with padding of 1, and the output feature map size is 64×256×8 (channels × subcarriers × time).

[0033] S301: Input the initial feature map into a multi-layer concatenated residual block, where each residual block has a time dimension length of... The input feature map is convolved to output the target feature map.

[0034] Specifically, each residual block includes a first convolutional layer, a first normalization layer, an activation function, a second convolutional layer, and a second normalization layer connected in series; wherein, the input of the first convolutional layer and the output of the second normalization layer are added element-wise to obtain the output of the residual block of that layer.

[0035] For example, 6 residual blocks are concatenated, and each residual block consists of: 3×3 convolution (64→64, padding 1), batch normalization, PReLU activation, 3×3 convolution (64→64, padding 1), batch normalization, and the input of the residual block is added to the above output element by element through residual connection; all residual blocks are operated on a low-resolution feature map with a time dimension of 8, and the feature map size is maintained at 64×256×8.

[0036] S302: Input the target feature map into the anisotropic upsampling module, thereby reducing the time dimension length of the target feature map from... Upsampling to Output upsampled feature map.

[0037] Specifically, the anisotropic upsampling module includes a transposed convolution, where the number of input and output channels is C, and the kernel size of the transposed convolution is 1 in the subcarrier dimension and 1 in the time dimension. The step sizes in the subcarrier dimension and the time dimension are 1 and 1, respectively. Both the padding and output padding are 0.

[0038] For example, the transposed convolution has 64 input and 64 output channels, with kernel sizes of 1 and 8 in the subcarrier and time dimensions, respectively, strides of 1 and 8, and padding of 0 in both the input and output. Thus, the time dimension changes from 8 to (8... 1) × 8 + 8 = 64, the subcarrier dimension remains unchanged at 256, and the output size is 64 × 256 × 64.

[0039] S303: Input the upsampled feature map into the reconstruction convolutional layer for convolution, and output the reconstructed feature map.

[0040] Specifically, the reconstructed convolutional layer is a convolution used to map C feature channels to two output channels.

[0041] For example, the reconstructed convolutional layer is a 3×3 convolution (64→2, padding 1), which outputs the real and imaginary parts of a high-resolution CFR with dimensions of 2×256×64.

[0042] S304: Input the low-resolution channel matrix into the global skip connection branch, thereby performing linear interpolation of the low-resolution channel matrix along the time dimension and outputting a high-resolution observation matrix with the same dimension as the reconstructed feature map.

[0043] S305: Input the reconstructed feature map and the high-resolution observation matrix into the fusion module, perform element-wise addition and fusion, and output a high-resolution channel matrix.

[0044] For example, a global skip connection will The 2×256×64 observations obtained by linear interpolation (8→64) along the time dimension are added element-wise to the output of the reconstructed convolutional layer to obtain the final high-resolution channel matrix. In this context, global hop connections preserve the dominant channel trend, while residual branches refine local amplitude and phase variations. In a specific example, the input network is pre-processed with... Standardize the data based on the channel-wise mean and standard deviation of the training set statistics.

[0045] In some embodiments, before inputting the low-resolution channel matrix into the super-resolution network for channel super-resolution reconstruction to obtain the high-resolution channel matrix, the method further includes steps 1 to 5.

[0046] Step 1: Insert preset block pilot symbols at all frequency domain UFMC symbol positions in the training frame.

[0047] Step 2: Perform channel estimation on all frequency domain UFMC symbol locations in the training frame to obtain the channel frequency response covering all frequency domain UFMC symbols in the entire frame, which serves as a high-resolution label.

[0048] Step 3: Extract from high-resolution tags The channel frequency response corresponding to each block pilot position is used to construct a low-resolution input matrix.

[0049] Step 4: Input the low-resolution input matrix into the super-resolution network to be trained, and output the reconstructed high-resolution channel matrix.

[0050] Step 5: Construct a loss function using the normalized mean square error between the reconstructed high-resolution channel matrix and the high-resolution label, and then iteratively train the super-resolution network to be trained until the value of the loss function converges, thus obtaining the trained super-resolution network.

[0051] Specifically, different modulation schemes of UFMC signals correspond to different super-resolution networks. That is, when the modulation scheme changes, the corresponding super-resolution network is retrained based on the signal under that modulation scheme.

[0052] For example, during the training phase, the transmitter inserts training frames with known symbols at all 64 UEMC symbol positions, and the receiver performs LS channel estimation at all 64 symbol positions to obtain a complete CFR covering all 64 symbol positions as a high-resolution label. (Dimensions 2×256×64); paired with low-resolution input Then, the CFR (dimension 2×256×8) at the Np=8 block pilot positions in the label is taken. It should be noted that the above-mentioned known symbol training frames are only used for offline acquisition and training of labels; during the deployment phase, the transmitted frame only contains pilots at these 8 block pilot positions, and the receiver constructs the HLR from these 8 positions and reconstructs the CFR of all 64 symbols through the network. Therefore, compared with the full pilot frame, no additional pilot overhead is required.

[0053] In a specific example, the training loss function is the normalized mean square error (NMSE) between the reconstructed high-resolution CFR and the high-resolution label, and its expression is: ,in, For batch size, Let Frobenius norm be denoted, and the superscript (i) denotes the i-th sample in the batch. This normalization makes the loss insensitive to different channel amplitudes, thus making it applicable to CFR reconstruction under different received optical power conditions.

[0054] In a specific example, the number of trainable parameters for the super-resolution network is approximately 0.594 × 10⁶. For the 64-QAM dataset, the training, validation, and test sets contain 2112, 264, and 264 pairs of samples, respectively. The 256-QAM dataset uses the same partitioning and trains a separate network with the same structure. The network is trained for 500 epochs using the Adam optimizer, with a batch size of Nb = 32 and an initial learning rate of 10⁻³. The network weights are fixed during the testing and deployment phases.

[0055] S40: Decompose the real and imaginary channels of the high-resolution channel matrix to obtain a matrix covering the entire frame. The global channel frequency response of each frequency domain UFMC symbol; where... This represents the time-dimensional upsampling factor of the anisotropic upsampling module in a super-resolution network.

[0056] S50: Based on the global channel frequency response, frequency domain equalization compensation and orthogonal amplitude modulation demapping are performed on the effective data subcarriers within the frame to obtain the demodulated frame signal.

[0057] For each subcarrier k and symbol position n, by Two channels reconstruct complex CFR: Based on this, single-tap frequency domain equalization is performed on all data subcarriers, and QAM demapping and bit error rate assessment are completed.

[0058] Based on the channel estimation method for super-resolution optical UFMC systems provided in the above embodiments, this application also provides a receiver, including: The pilot extraction module performs time-domain preprocessing and UFMC demodulation on each frame of the received UFMC signal to obtain the full-frame frequency-domain UFMC symbols, and extracts the pilot symbols arranged in the frequency domain from the full-frame frequency-domain UFMC symbols. Block pilot symbols at block pilot locations; where each block pilot symbol covers K valid data subcarriers, where K represents the total number of valid data subcarriers in the frame; The low-resolution channel matrix construction module is used to perform least-squares channel estimation on each block pilot symbol to obtain the coarse channel frequency response corresponding to each block pilot symbol; the real and imaginary parts of all coarse channel frequency responses are stacked along the channel dimension to obtain the low-resolution channel matrix. The channel super-resolution reconstruction module is used to input the low-resolution channel matrix into the super-resolution network for channel super-resolution reconstruction to obtain a high-resolution channel matrix. The global channel frequency response acquisition module is used to split the real and imaginary channels of the high-resolution channel matrix to obtain a signal covering the entire frame. The global channel frequency response of each frequency domain UFMC symbol; where... This represents the temporal upsampling factor of the anisotropic upsampling module in a super-resolution network. The equalization and demapping module is used to perform frequency domain equalization compensation and orthogonal amplitude modulation demapping on the effective data subcarriers within the frame based on the global channel frequency response, so as to obtain the demodulated frame signal.

[0059] To verify the effectiveness of the above method, this application also conducted experimental tests on the method provided in the above embodiments: like Figure 5 The figure shows the bit error rate curves of the 64-QAM UFMC provided in this application embodiment under back-to-back conditions with and without Volterra equalization; as shown Figure 6 The figure shows the bit error rate curves of the 64-QAM UFMC provided in this application after transmission over 30-km single-mode fiber with and without Volterra equalization. Under 64-QAM, the method provided in this application achieves the lowest bit error rate among all schemes under both back-to-back and 30-km fiber transmission conditions, and its scheme combined with third-order Volterra equalization (VE) has the best performance. The hard-decision forward error correction threshold is 3.8 × 10⁻⁶. -3 For reference, under back-to-back conditions, super-resolution combined with VE improves receiver sensitivity by approximately 2.9 dB compared to LS combined with VE; this application can surpass this threshold even without using VE, and outperforms LS estimation combined with VE.

[0060] like Figure 7 The figure shown is a comparison of the subcarrier signal-to-noise ratios of different channel estimation methods provided in the embodiments of this application; as shown Figure 8 The figure shown is a comparison of the amplitude and phase response of the estimated channel for different channel estimation methods provided in the embodiments of this application; as shown... Figure 9 The diagram shown is a receiver constellation diagram of 64-QAM under different channel estimation methods according to an embodiment of this application. Figure 9 (a) in the diagram is the receiver constellation diagram for 64-QAM under the least squares channel estimation (LS) method. Figure 9 (b) in the diagram is the receiver constellation diagram for 64-QAM under the Linear Minimum Mean Square Error (LMMSE) channel estimation method. Figure 9 (c) in the diagram represents the receiver constellation diagram for 64-QAM under the CNN-CE method for channel estimation based on convolutional neural networks. Figure 9In the diagram (d), the receiving constellation diagram of 64-QAM under the proposed SR method for channel estimation provided in this application is shown. The subcarrier signal-to-noise ratio, channel amplitude and phase response, and receiving constellation diagram respectively demonstrate that this application can improve the signal-to-noise ratio of weak subcarriers such as subband edges, reconstruct the channel amplitude and phase more reliably, and obtain a more compact constellation.

[0061] like Figure 10 The figure shows the bit error rate curves of the 256-QAM UFMC provided in this application embodiment under back-to-back conditions with and without Volterra equalization; as shown Figure 11 The figure shows the bit error rate curves of the 256-QAM UFMC provided in this application embodiment after transmission through 30-km single-mode fiber, with and without Volterra equalization; as shown Figure 12 The diagram shown is a receiver constellation diagram of 256-QAM under different channel estimation methods according to an embodiment of this application. Figure 12 (a) in the diagram is the receiver constellation diagram of 256-QAM under the least squares channel estimation (LS) method. Figure 12 (b) in the diagram is the receiver constellation diagram for 256-QAM under the Linear Minimum Mean Square Error (LMMSE) channel estimation method. Figure 12 (c) in the diagram represents the receiver constellation diagram for 256-QAM under the CNN-CE method for channel estimation based on convolutional neural networks. Figure 12 In the diagram, (d) represents the receiver constellation diagram for 256-QAM under the proposed SR channel estimation method provided in this application. Under 256-QAM, this application also achieves the lowest bit error rate; under 30-km transmission conditions, only the super-resolution combined with the VE scheme in this application surpasses the soft-decision forward error correction threshold of 2.7 × 10⁻⁶. -2 In terms of computational complexity, the multiplication operation per frame in this application is approximately 1.23 × 10⁹, which is lower than the approximately 9.66 × 10⁹ for convolutional networks that first interpolate to high resolution and then refine. 9 Furthermore, its computation is purely feedforward and can be parallelized.

[0062] In summary, this application models the channel estimation of block pilot UFMC as an anisotropic low-resolution to high-resolution CFR reconstruction problem. The super-resolution network maintains the integrity of the subcarrier dimension and upsamples only along the time dimension, which fits the inherent structure of block pilots and improves the accuracy of CFR reconstruction. The head convolutional layer and residual block of the super-resolution network are operated on the low-resolution feature map with a small time dimension. The upsampling in the time dimension is completed by a single transposed convolution. Compared with networks that first interpolate the CFR to a high-resolution size and then perform full convolution refinement on the high-resolution feature map, this reduces the amount of multiplication operations under the condition of comparable parameter quantity. Its computation is pure feedforward and parallelizable, and the inference latency is lower than that of the recursive least squares (RLS) third-order Volterra equalization updated sample by sample. The complex CFR is then reconstructed by jointly using the real and imaginary parts, and the dominant channel trend is preserved through global skip connections, while the local amplitude and phase changes are refined by the residual branch, making the reconstruction more reliable in both amplitude and phase. Furthermore, this application is compatible with time-domain Volterra equalization and can be used for high-order modulations such as 64-QAM and 256-QAM, as well as dense subcarrier loading conditions with 256 effective data subcarriers. It supports 64-QAM back-to-back transmission and a hard-decision forward error correction threshold of 3.8 × 10⁻⁶. - For reference, this application, combining third-order Volterra equalization, improves receiver sensitivity by approximately 2.9 dB relative to least squares estimation combined with third-order Volterra equalization, and improves it by approximately 1.1 dB relative to linear minimum mean square error estimation combined with third-order Volterra equalization; in 256-QAM transmission over 30-km single-mode fiber, with a soft-decision forward error correction threshold of 2.7 × 10⁻⁶. - Under the condition of ², among the compared schemes, only this application, combined with third-order Volterra equalization, surpasses the threshold; and the super-resolution network of this application requires approximately 1.23 × 10⁻⁶ units per frame. 9 This real multiplication is approximately equivalent to a convolutional network that first interpolates the channel frequency response to a high-resolution scale and then refines it (approximately 9.66 × 10⁻⁶). 9 One-eighth of (times).

[0063] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0064] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0065] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0066] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0067] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A channel estimation method based on a super-resolution optical UFMC system, characterized in that, The receiver used in super-resolution optical UFMC systems includes: Each received UFMC signal frame undergoes time-domain preprocessing and UFMC demodulation to obtain the entire frame's frequency-domain UFMC symbols. The symbols arranged in the frequency domain are then extracted from these symbols. Block pilot symbols at block pilot locations; where each block pilot symbol covers K valid data subcarriers, where K represents the total number of valid data subcarriers in the frame; Least-square channel estimation is performed on each block pilot symbol to obtain the coarse channel frequency response corresponding to each block pilot symbol; the real and imaginary parts of all coarse channel frequency responses are stacked along the channel dimension to obtain a low-resolution channel matrix. The low-resolution channel matrix is ​​input into the super-resolution network for channel super-resolution reconstruction to obtain the high-resolution channel matrix. The real and imaginary channels of the high-resolution channel matrix are split to obtain a channel covering the entire frame. The global channel frequency response of each frequency domain UFMC symbol; where... This represents the temporal upsampling factor of the anisotropic upsampling module in a super-resolution network. Based on the global channel frequency response, frequency domain equalization compensation and orthogonal amplitude modulation demapping are performed on the effective data subcarriers within the frame to obtain the demodulated frame signal.

2. The channel estimation method based on a super-resolution optical UFMC system according to claim 1, characterized in that, The time-domain preprocessing of each received UFMC signal frame includes performing third-order Volterra time-domain equalization on each received UFMC signal frame.

3. The channel estimation method based on a super-resolution optical UFMC system according to claim 1, characterized in that, The coarse channel frequency response corresponding to the nth block pilot symbol for: , in, This represents the diagonal matrix formed by the nth block of pilot symbols; This represents the nth block pilot symbol.

4. The channel estimation method based on a super-resolution optical UFMC system according to claim 1, characterized in that, The super-resolution network includes a head convolutional layer, multiple cascaded residual blocks, anisotropic upsampling modules, reconstruction convolutional layers, global skip connection branches, and a fusion module.

5. The channel estimation method based on a super-resolution optical UFMC system according to claim 4, characterized in that, The low-resolution channel matrix is ​​input into the super-resolution network for channel super-resolution reconstruction to obtain a high-resolution channel matrix, including: The low-resolution channel matrix is ​​input into the head convolutional layer, thereby optimizing the time dimension. The low-resolution feature map is convolved to output the initial feature map; The initial feature map is input into multiple concatenated residual blocks, each residual block having a time dimension length of [missing information]. The input feature map is convolved to output the target feature map; The target feature map is input into the anisotropic upsampling module, thereby reducing the time dimension length of the target feature map from... Upsampling to Output the upsampled feature map; The upsampled feature map is input into the reconstruction convolutional layer for convolution, and the reconstructed feature map is output. The low-resolution channel matrix is ​​input into the global skip connection branch, thereby performing linear interpolation of the low-resolution channel matrix along the time dimension and outputting a high-resolution observation matrix with the same dimension as the reconstructed feature map. The reconstructed feature map and the high-resolution observation matrix are input into the fusion module and fused element by element to output a high-resolution channel matrix.

6. The channel estimation method based on a super-resolution optical UFMC system according to claim 5, characterized in that, Each residual block consists of a first convolutional layer, a first normalization layer, an activation function, a second convolutional layer, and a second normalization layer, all connected in series. The input of the first convolutional layer and the output of the second normalization layer are added element-wise to obtain the output of the residual block.

7. The channel estimation method based on a super-resolution optical UFMC system according to claim 5, characterized in that, The head convolutional layer is used to map the real and imaginary input channels of the low-resolution channel matrix into C feature channels; The number of feature channels in each residual block is C; The anisotropic upsampling module includes a transposed convolution, where the number of input and output channels is C. The kernel size of the transposed convolution has dimensions of 1 in the subcarrier dimension and 1 in the time dimension. The step sizes in the subcarrier dimension and the time dimension are 1 and 1, respectively. Both the fill and output fill are 0; The reconstructed convolutional layer is a convolution used to map C feature channels to 2 output channels.

8. The channel estimation method based on a super-resolution optical UFMC system according to claim 5, characterized in that, Before inputting the low-resolution channel matrix into the super-resolution network for channel super-resolution reconstruction and obtaining the high-resolution channel matrix, the following steps are also included: Preset block pilot symbols are inserted at all frequency domain UFMC symbol positions in the training frames; Channel estimation is performed on all frequency domain UFMC symbol locations in the training frame to obtain the channel frequency response covering all frequency domain UFMC symbols in the entire frame, which serves as a high-resolution label; Extract from high-resolution tags The channel frequency response corresponding to each block pilot position is used to construct a low-resolution input matrix; The low-resolution input matrix is ​​input into the super-resolution network to be trained, and the output is a reconstructed high-resolution channel matrix. A loss function is constructed using the normalized mean square error between the reconstructed high-resolution channel matrix and the high-resolution label. This loss function is then used to iteratively train the super-resolution network until the value of the loss function converges, resulting in a well-trained super-resolution network.

9. The channel estimation method based on a super-resolution optical UFMC system according to claim 8, characterized in that, Different modulation schemes of UFMC signals correspond to different super-resolution networks.

10. A receiving end, characterized in that, include: The pilot extraction module performs time-domain preprocessing and UFMC demodulation on each frame of the received UFMC signal to obtain the full-frame frequency-domain UFMC symbols, and extracts the pilot symbols arranged in the frequency domain from the full-frame frequency-domain UFMC symbols. Block pilot symbols at block pilot locations; where each block pilot symbol covers K valid data subcarriers, where K represents the total number of valid data subcarriers in the frame; The low-resolution channel matrix construction module is used to perform least-squares channel estimation on each block pilot symbol to obtain the coarse channel frequency response corresponding to each block pilot symbol; the real and imaginary parts of all coarse channel frequency responses are stacked along the channel dimension to obtain the low-resolution channel matrix. The channel super-resolution reconstruction module is used to input the low-resolution channel matrix into the super-resolution network for channel super-resolution reconstruction to obtain a high-resolution channel matrix. The global channel frequency response acquisition module is used to split the real and imaginary channels of the high-resolution channel matrix to obtain a signal covering the entire frame. The global channel frequency response of each frequency domain UFMC symbol; where... This represents the temporal upsampling factor of the anisotropic upsampling module in a super-resolution network. The equalization and demapping module is used to perform frequency domain equalization compensation and orthogonal amplitude modulation demapping on the effective data subcarriers within the frame based on the global channel frequency response, so as to obtain the demodulated frame signal.