Non-terrestrial network ka band phase noise estimation method, apparatus and device

Abstract

The application relates to a non-ground network Ka-band phase noise estimation method, device and equipment. The method comprises the following steps: acquiring disturbed received signals after synchronization and matched filtering processing in a non-ground network Ka band, and extracting each received symbol data in a preset pilot band or decision feedback band; extracting real and imaginary parts of each received symbol data respectively to obtain a two-dimensional vector corresponding to a single received symbol; taking a plurality of continuous received symbols as an observation window, splicing the two-dimensional vectors corresponding to all the received symbols in the window in time sequence to construct a two-dimensional feature tensor; inputting the two-dimensional feature tensor into a pre-trained multi-layer perception hybrid network; and performing phase anti-rotation compensation on each received symbol in the observation window according to a phase noise estimation value to obtain a received symbol after phase noise elimination. The method can effectively improve the transmission performance and reliability of a high-order modulation communication system.

Description

Technical Field

[0001] This application relates to the field of signal processing technology, and in particular to a method, apparatus and device for estimating Ka-band phase noise in non-terrestrial networks. Background Technology

[0002] With the development of 5G non-terrestrial network technology, Ka-band satellite communication, with its abundant spectrum resources and high transmission rate, has been widely used in deep space telemetry and control, relay transmission, broadband multimedia services, and other scenarios. To further improve spectrum utilization efficiency, higher-order modulation methods such as 64QAM and 256QAM have also become mainstream technical choices. Regarding the phase noise problem in communication links, traditional technologies mainly fall into two categories: hardware compensation and digital domain compensation. Hardware compensation suppresses noise at its source through oscillator optimization schemes such as dielectric resonant cavities and sampling phase-locked loops (PLLs). Digital domain compensation, on the other hand, uses classic algorithms such as PLL phase tracking and decision feedback phase estimation in the receiver baseband to correct residual phase noise.

[0003] However, traditional hardware compensation schemes are limited by the material technology and circuit structure of Ka-band oscillators, and can only reduce the floor noise near the carrier offset. The phase noise level in the far carrier offset range of 100kHz to 1MHz is still difficult to meet the stringent requirements of high-order modulation, and hardware optimization will significantly increase device power consumption and chip area. The compensation accuracy of classical digital domain compensation methods is limited by the algorithm's ability to model the statistical characteristics of noise. Since the phase noise power spectral density has a power-law distribution and includes both low-frequency high-energy noise near the carrier frequency and high-frequency low-energy noise far from the carrier frequency, traditional linear filters cannot achieve synchronous and effective suppression over a wide frequency range. In recent years, AI-assisted compensation schemes have mostly adopted architectures such as convolutional networks and Transformers. Although they have a certain nonlinear fitting ability, they cannot simultaneously and effectively capture the time-frequency joint distribution of near-carrier frequency and far-carrier frequency noise. Moreover, the model parameters are large and the computational complexity is high, making it difficult to adapt to the deployment constraints of low power consumption, low latency, and limited computing power of spaceborne equipment. Summary of the Invention

[0004] Therefore, it is necessary to provide a method, apparatus, and device for estimating Ka-band phase noise in non-terrestrial networks to address the aforementioned technical problems.

[0005] A method for estimating Ka-band phase noise in non-terrestrial networks, the method comprising: Acquire the disturbed received signal in the Ka band of a non-terrestrial network after synchronization and matched filtering, and extract the data of each received symbol in the preset pilot band or decision feedback band; For each received symbol data, extract the real and imaginary components to obtain a two-dimensional vector corresponding to a single received symbol. Using multiple consecutive received symbols as an observation window, concatenate the two-dimensional vectors corresponding to all received symbols in the observation window in chronological order to construct a two-dimensional feature tensor. The two-dimensional feature tensor is input into a pre-trained multilayer perceptron hybrid network. The multilayer perceptron hybrid network includes alternating stacked token-mixing multilayer perceptron modules and channel-mixing multilayer perceptron modules. The token-mixing multilayer perceptron module performs global mixing operations on the two-dimensional feature tensor along the time dimension to extract the temporal correlation features of the phase noise between symbols. The channel-mixing multilayer perceptron module performs cross-channel mixing operations on the two-dimensional feature tensor along the channel dimension to extract the common structural features of the real and imaginary signals. The final fused features obtained by alternating multilayer mixing are then mapped to the estimated phase noise value corresponding to each received symbol within the observation window after residual connection and layer normalization stabilization processing. Based on the estimated phase noise value, phase inversion compensation is performed on each received symbol within the observation window to obtain the received symbol after phase noise elimination.

[0006] A non-terrestrial network Ka-band phase noise estimation device, the device comprising: The signal acquisition module is used to acquire the disturbed received signal after synchronization and matched filtering in the Ka band of the non-terrestrial network, and extract the data of each received symbol in the preset pilot band or decision feedback band. The tensor construction module is used to extract the real and imaginary components of each received symbol data to obtain a two-dimensional vector corresponding to a single received symbol. Taking multiple consecutive received symbols as an observation window, the two-dimensional vectors corresponding to all received symbols in the observation window are concatenated in chronological order to construct a two-dimensional feature tensor. A noise estimation module is used to input the two-dimensional feature tensor into a pre-trained multilayer perceptron hybrid network. The multilayer perceptron hybrid network includes alternating stacked token-mixing multilayer perceptron modules and channel-mixing multilayer perceptron modules. The token-mixing multilayer perceptron module performs global mixing operations on the two-dimensional feature tensor along the time dimension to extract the temporal correlation features of the phase noise between symbols. The channel-mixing multilayer perceptron module performs cross-channel mixing operations on the two-dimensional feature tensor along the channel dimension to extract the common structural features of the real and imaginary signals. The fused features obtained by alternating multilayer mixing are then stabilized by residual connections and layer normalization to map the estimated phase noise value corresponding to each received symbol within the observation window. The result output module is used to perform phase inverse rotation compensation on each received symbol in the observation window based on the phase noise estimate, so as to obtain the received symbol after phase noise elimination.

[0007] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program performing the following steps: Acquire the disturbed received signal in the Ka band of a non-terrestrial network after synchronization and matched filtering, and extract the data of each received symbol in the preset pilot band or decision feedback band; For each received symbol data, extract the real and imaginary components to obtain a two-dimensional vector corresponding to a single received symbol. Using multiple consecutive received symbols as an observation window, concatenate the two-dimensional vectors corresponding to all received symbols in the observation window in chronological order to construct a two-dimensional feature tensor. The two-dimensional feature tensor is input into a pre-trained multilayer perceptron hybrid network. The multilayer perceptron hybrid network includes alternating stacked token-mixing multilayer perceptron modules and channel-mixing multilayer perceptron modules. The token-mixing multilayer perceptron module performs global mixing operations on the two-dimensional feature tensor along the time dimension to extract the temporal correlation features of the phase noise between symbols. The channel-mixing multilayer perceptron module performs cross-channel mixing operations on the two-dimensional feature tensor along the channel dimension to extract the common structural features of the real and imaginary signals. The final fused features obtained by alternating multilayer mixing are then mapped to the estimated phase noise value corresponding to each received symbol within the observation window after residual connection and layer normalization stabilization processing. Based on the estimated phase noise value, phase inversion compensation is performed on each received symbol within the observation window to obtain the received symbol after phase noise elimination.

[0008] The aforementioned non-terrestrial network Ka-band phase noise estimation method, apparatus, and device, by constructing a time-channel two-dimensional feature tensor using the real and imaginary parts of the received symbols, can completely preserve the continuous evolution trajectory of phase noise and the coupling characteristics of orthogonal channels, providing a raw input basis for noise estimation without pre-error propagation. By employing a network architecture that alternately stacks token-mixing and channel-mixing multilayer perceptrons, it globally extracts temporal correlation features along the time dimension and fuses common structural features of orthogonal components along the channel dimension. This allows it to simultaneously adapt to the slow-changing near-carrier frequency and fast-changing far-carrier frequency characteristics of power-law distributed phase noise, maintaining stable and consistent estimation accuracy across the entire offset frequency band. Through symbol-by-symbol phase inverse rotation compensation, it can directly correct signal constellation distortion caused by phase noise. The embodiments of this invention can achieve efficient phase noise suppression with low computational overhead in non-terrestrial network Ka-band scenarios, effectively improving the transmission performance and reliability of high-order modulation communication systems. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating a non-terrestrial network Ka-band phase noise estimation method in one embodiment; Figure 2 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0010] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0011] In one embodiment, such as Figure 1 As shown, a method for estimating Ka-band phase noise in non-terrestrial networks is provided, including the following steps: Step 102: Obtain the disturbed received signal after synchronization and matched filtering in the Ka band of the non-terrestrial network, and extract the data of each received symbol in the preset pilot band or decision feedback band.

[0012] Non-terrestrial networks refer to network architectures that achieve wide-area coverage based on non-terrestrial communication nodes such as satellites. The Ka band is a communication frequency band covering 20 / 30 GHz, widely used in satellite broadband and telemetry and control scenarios. Synchronization refers to the baseband processing operation at the receiver to complete carrier frequency calibration and symbol timing alignment. Matched filtering refers to the conventional operation at the receiver to maximize the signal-to-noise ratio of the received signal through filtering processing that matches the shaping filter at the transmitter. Pilot band refers to the interval of reference signals pre-agreed upon and known to both the transmitter and receiver within a communication frame. Decision feedback segment refers to the interval of the symbol sequence obtained after preliminary demodulation decision. Received symbol data refers to the baseband sampled data corresponding to a single modulation symbol carrying transmitted information.

[0013] It is understandable that this step can obtain a clean observation data source in the baseband stage before demodulation, providing a reliable input basis for subsequent phase noise estimation. This is beneficial for maintaining the stability of the estimation process under low signal-to-noise ratio and dynamic non-terrestrial network channels, and avoids introducing additional error propagation in the preprocessing stage.

[0014] Step 104: Extract the real and imaginary components of each received symbol data to obtain a two-dimensional vector corresponding to a single received symbol. Using multiple consecutive received symbols as an observation window, concatenate the two-dimensional vectors corresponding to all received symbols in the observation window in chronological order to construct a two-dimensional feature tensor.

[0015] The real and imaginary components are two orthogonal components of a complex baseband signal in an orthogonal coordinate system, together constituting a complete representation of the complex signal. A two-dimensional vector refers to a two-dimensional data structure composed of the real and imaginary components of a single symbol. An observation window refers to the length range of a continuous symbol sequence used for single-stage phase noise estimation. A two-dimensional feature tensor refers to a two-dimensional data structure arranged in a regular manner according to its dimensions, serving as input data for subsequent deep learning networks.

[0016] It is understandable that this step can completely preserve the continuity of phase evolution between symbols and the coupling relationship of the orthogonal communication path, providing the network with a learnable noise evolution trajectory. At the same time, it forms a regular input structure that is naturally adapted to subsequent hybrid networks, and can directly participate in the operation without additional dimensional transformation, which can improve the accuracy and computational efficiency of phase noise estimation.

[0017] Step 106: Input the two-dimensional feature tensor into a pre-trained multilayer perceptron hybrid network. The multilayer perceptron hybrid network includes alternating stacked token-mixing multilayer perceptron modules and channel-mixing multilayer perceptron modules. The token-mixing multilayer perceptron module performs global mixing operations on the two-dimensional feature tensor along the time dimension to extract the temporal correlation features of the phase noise between symbols. The channel-mixing multilayer perceptron module performs cross-channel mixing operations on the two-dimensional feature tensor along the channel dimension to extract the common structural features of the real and imaginary signals. The final fused feature obtained through alternating multilayer mixing is then stabilized by residual connections and layer normalization to map the estimated phase noise value corresponding to each received symbol within the observation window.

[0018] The Multilayer Perceptron Hybrid Network is a deep learning network architecture built upon the Multilayer Perceptron (MLP) and extracts information by alternately performing feature mixing across different dimensions. The token-mixing MLP module performs fully connected mapping along the time dimension to achieve global mixing of temporal features, while the channel-mixing MLP module performs fully connected mapping along the channel dimension to achieve cross-channel feature fusion. Global mixing refers to performing fully connected feature mapping without local receptive field limitations on all symbols within a window, while cross-channel mixing refers to performing joint feature mapping on two orthogonal components of the same symbol. Residual connection refers to a network structure that directly adds the input and output features of a module, mitigating gradient degradation in deep networks. Layer normalization refers to performing normalization calibration on feature data layer by layer, stabilizing the training and inference processes. The final fused feature refers to the comprehensive feature representation obtained after alternating multilayer mixing, and the phase noise estimate refers to the predicted amplitude of the phase perturbation for each symbol.

[0019] This step can be understood as simultaneously capturing the temporal evolution of phase noise and the coupling characteristics of the orthogonal transmission channel, adapting to the dual characteristics of power-law distributed phase noise—slow variation near the carrier frequency and rapid variation far from the carrier frequency—and maintaining consistent estimation accuracy across the entire offset frequency band. The purely fully connected network structure has low computational complexity, enabling real-time inference on computationally limited spaceborne platforms. Residual connections and layer normalization improve network stability, facilitating a balance between estimation accuracy and deployment feasibility.

[0020] Step 108: Based on the estimated phase noise value, perform phase inversion compensation on each received symbol in the observation window to obtain the received symbol after phase noise elimination.

[0021] Phase-reverse rotation compensation refers to applying a reverse phase rotation operation to the disturbed complex signal based on the estimated phase noise value in order to cancel the phase distortion caused by the phase noise.

[0022] It is understandable that this step can directly cancel the constellation diagram distortion and phase error caused by phase noise, reduce the bit error rate of the communication system, improve the transmission reliability in high-order modulation scenarios, and the compensation process only needs to perform phase rotation operation symbol by symbol, with low computational overhead, and can be seamlessly connected with the existing baseband receiving link.

[0023] In the aforementioned non-terrestrial network Ka-band phase noise estimation method, by constructing a time-channel two-dimensional feature tensor using the real and imaginary parts of the received symbols, the continuous evolution trajectory of phase noise and the coupling characteristics of orthogonal channels can be completely preserved, providing a raw input basis for noise estimation without pre-error propagation. By employing a network architecture that alternately stacks token-mixing and channel-mixing multilayer perceptrons, temporal correlation features are globally extracted along the time dimension, and common structural features of orthogonal components are fused along the channel dimension. This allows for simultaneous adaptation to the slow variation near the carrier frequency and the fast variation far from the carrier frequency of power-law distributed phase noise, maintaining stable and consistent estimation accuracy across the entire offset frequency band. Through symbol-by-symbol phase inverse rotation compensation, signal constellation distortion caused by phase noise can be directly corrected. This invention enables efficient phase noise suppression with low computational overhead in non-terrestrial network Ka-band scenarios, effectively improving the transmission performance and reliability of high-order modulation communication systems.

[0024] In one embodiment, the two-dimensional feature tensor is constructed by concatenating the two-dimensional vectors corresponding to all received symbols in the observation window in chronological order. This includes: regularizing the two-dimensional vectors corresponding to each received symbol with the time dimension as the first dimension and the channel dimension as the second dimension to obtain the two-dimensional feature tensor.

[0025] In this embodiment, the dimensional arrangement of the two-dimensional feature tensor is perfectly compatible with the dual-domain hybrid logic of the multilayer perceptron hybrid network, eliminating the need for additional dimensional transformation and data reconstruction operations and avoiding redundant computational overhead. This reduces the computational load of the feature preprocessing stage, avoids information loss during dimensional transformation, and provides a regular input structure for subsequent alternating feature mixing of the time and channel dimensions, which is beneficial for improving network inference efficiency and feature extraction accuracy.

[0026] In one embodiment, the token-mixing multilayer perceptron module includes a first fully connected layer, a first activation layer, and a second fully connected layer cascaded in sequence. The global mixing operation along the time dimension on the two-dimensional feature tensor to extract the temporal correlation features of inter-symbol phase noise includes: transposing the input two-dimensional feature tensor to switch the time dimension to the channel dimension; sequentially inputting the transposed feature tensor into the first fully connected layer for linear mapping, the first activation layer for nonlinear transformation, and the second fully connected layer for quadratic linear mapping to obtain intermediate features; and performing a reverse dimensional transpose on the intermediate features to obtain the temporal correlation features of inter-symbol phase noise.

[0027] In this embodiment, the time dimension is switched to a channel dimension that can be processed by the fully connected layer through dimension transposition. A structure of two fully connected layers combined with an activation layer achieves global temporal mixing without local receptive field constraints, unaffected by the prior limitations of fixed window length filtering. It can completely capture the cumulative drift and temporal dependence of phase noise within the window, while adapting to the evolution characteristics of slow-varying components near the carrier frequency and high-frequency jitter components far from the carrier frequency. It can accurately fit the nonlinear temporal patterns of power-law distributed phase noise, effectively improving the extraction accuracy of time dimension features.

[0028] In one embodiment, the channel-mixing multilayer perceptron module includes a third fully connected layer, a second activation layer, and a fourth fully connected layer cascaded in sequence. The cross-channel mixing operation on the two-dimensional feature tensor along the channel dimension to extract the common structural features of the real and imaginary signals includes: sequentially inputting the input two-dimensional feature tensor into the third fully connected layer for linear mapping, the second activation layer for nonlinear transformation, and the fourth fully connected layer for quadratic linear mapping, to obtain the common structural features of the real and imaginary signals.

[0029] In this embodiment, a structure consisting of a double-layer fully connected layer and an activation layer is used to directly perform joint mapping on the real and imaginary components, completing feature interaction and fusion at the channel dimension without the need for manually setting orthogonal constraints. This approach can fully exploit the nonlinear coupling correlation between the two orthogonal signals under phase noise perturbation, accurately extract the common noise distortion patterns of the two signals, enhance the ability to capture common phase noise features, and improve the robustness of the estimation results.

[0030] In one embodiment, after each group of alternately arranged token-mixing multilayer perceptron modules and channel-mixing multilayer perceptron modules, a residual connection unit and a layer normalization unit are connected in series. The fused features obtained through alternating multilayer mixing are stabilized by residual connection and layer normalization, and then mapped to obtain the phase noise estimate corresponding to each received symbol in the observation window. This includes: after feature extraction processing by each group of token-mixing multilayer perceptron modules and channel-mixing multilayer perceptron modules, the input features and output features of the current group are input to the residual connection unit for element-wise addition, and the addition result is input to the layer normalization unit for normalization to obtain the output features of the current group. After processing by all the alternately stacked module groups in sequence, the final fused features are obtained. The final fused features are input to the fully connected output layer for linear mapping to obtain the phase noise estimate corresponding to each received symbol in the observation window.

[0031] In this embodiment, each module is followed by a residual connection and a layer normalization unit. Higher-order features are gradually extracted through iterative feature fusion, and finally, a fully connected output layer completes the linear mapping from the fused features to the phase noise estimate. This effectively alleviates the gradient degradation problem caused by deep network stacking, stabilizes the network's training convergence process and inference output, and can fuse temporal and channel-dimensional feature information layer by layer to form more representative joint features, ensuring the accuracy of symbol-by-symbol phase noise estimation.

[0032] In one embodiment, the step of training the multilayer perceptron hybrid network includes: acquiring a training sample set, wherein each training sample in the training sample set includes a two-dimensional feature tensor of the disturbed received signal and a corresponding true phase noise label; the true phase noise is a phase noise sequence that conforms to the power-law distribution characteristics of the Ka band; and iteratively training the multilayer perceptron hybrid network using the training sample set and the Huber loss function until the iteration stops when the iteration stopping condition is met, thereby obtaining the trained multilayer perceptron hybrid network.

[0033] In this embodiment, the training samples are constructed based on power-law distributed phase noise generated from measured parameters of a Ka-band transceiver oscillator. Additive white Gaussian noise with different signal-to-noise ratios is superimposed to simulate actual disturbed signals, and the corresponding real phase noise sequences are used as supervision labels. This allows the network to fully learn the inherent distribution characteristics of phase noise in Ka-band scenarios, adapt to channel environments of non-terrestrial networks, and improve the model's generalization ability and estimation accuracy in practical application scenarios.

[0034] In one embodiment, the method further includes: performing weight quantization and structural pruning on the trained multilayer perceptron hybrid network to obtain a lightweight inference model; and deploying the lightweight inference model to the receiver baseband processing unit.

[0035] In this embodiment, the model size and computational load are reduced by compressing parameter bit width through weight quantization and removing redundant channels through structural pruning. A split deployment mode of offline training on the ground and online inference on space is adopted. This can significantly reduce the network's storage footprint and computational overhead, allowing it to be directly deployed on spaceborne or ground-embedded baseband processing units with limited computing and storage resources, meeting the real-time processing requirements of non-terrestrial network receiving links.

[0036] In one embodiment, the Huber loss function employs a squared penalty term for estimation errors less than a preset threshold and a linear penalty term for estimation errors greater than a preset threshold, wherein the preset threshold is set based on the noise statistics at high bias frequencies.

[0037] In this embodiment, a threshold is set based on the statistical characteristics of noise at high bias frequencies. Differential penalties are applied to estimation errors of different magnitudes, using this threshold as a boundary. Small errors retain the convergence characteristics of squared penalties, while large errors switch to linear penalties. This effectively suppresses the interference of noise outliers at high bias frequencies on model training, avoids model fitting biased towards extreme error samples, improves the phase noise estimation accuracy in the high bias frequency range, and optimizes estimation stability across the entire frequency band.

[0038] In one specific embodiment, this invention proposes a Ka-band phase noise estimation and cancellation method based on the MLP-Mixer architecture. The core idea is to utilize the feature mixing capability of the MLP-Mixer network to map the time-frequency domain features of the received signal into symbol-by-symbol estimates of the phase noise, and then restore the constellation quality of the transmitted signal through online compensation. This method addresses the problems of existing technologies in 5G NTN Ka-band scenarios, such as network structure mismatch with power-law distributed phase noise, inability to simultaneously handle near-carrier frequency and far-carrier frequency noise, insufficient estimation accuracy, high algorithm complexity, and difficulties in satellite-based deployment. It provides an MLP-Mixer-based phase noise estimation scheme to achieve high-precision, low-latency, and lightweight online phase noise compensation, thereby improving system transmission reliability. The specific implementation steps are as follows: Step 1: Received signal preprocessing and feature extraction.

[0039] In the receiver baseband, samples are extracted from the disturbed received signal after synchronization and matched filtering in a predefined pilot band or decision feedback band. For each sample, the real and imaginary parts are taken to form a two-dimensional vector; in the time domain, a feature tensor of (K, 2) dimensions is constructed using K consecutive symbols as the observation window (K is the window length, typically 8-32).

[0040] Step 2: Forward inference using the MLP-Mixer network.

[0041] The feature tensor constructed in step one is input into the pre-trained MLP-Mixer network. This network consists of alternating stacked token-mixing MLP and channel-mixing MLP modules: the token-mixing MLP performs MLP operations along the time axis (between K symbols) to extract the temporal correlation features of the phase noise between symbols; the channel-mixing MLP performs MLP operations along the channel axis (real part / imaginary part) to extract the common structural features of the phase noise of the I and Q channels.

[0042] Through residual connections and layer normalization stabilization network training and inference processes, the extracted temporal correlation features and I / Q channel common structural features are fused dimension-wise to form a joint feature representation. The network completes nonlinear mapping and symbol-by-symbol prediction based on this joint feature representation, and the final output is the phase noise estimate of each symbol in the corresponding input window.

[0043] The MLP-Mixer network was trained offline using training data containing power-law distributed phase noise generated through simulation before deployment, and the training employed the Huber loss function.

[0044] The Huber threshold is calibrated based on the statistical characteristics of high-bias-frequency noise. Phase noise is characterized in logarithmic dB form.

[0045] Let the phase noise power be (linear dimensionless); let the true power be... Estimated power Absolute error in the linear domain The corresponding dB domain residuals are:

[0046] The high-bias frequency outliers follow a heavy-tailed non-Gaussian distribution. Let the outlier anomaly error in the power domain be of magnitude [value missing]. 1) Classic MSE optimization: Squared penalty weighted by wild value. Optimization parameters are influenced by outliers, and the power domain average residual is... ,in 1) The ratio of the power domain average residual to the true power when using mean squared error loss optimization; 2) Huber loss optimization: outliers beyond the threshold are replaced with linear penalties, and the outlier weights are reduced from quadratic to linear terms, increasing the power domain average residual. ,in This is the ratio of the power domain average residual to the true power when using Huber loss optimization.

[0047] Substituting the two types of residuals into the dB domain residual calculation relationship, we can obtain the dB domain residual corresponding to MSE as follows: The dB domain residual corresponding to Huber is The difference between the two is the amount of compensation accuracy improvement. .

[0048] Calculations based on noise characteristics across different frequency offset ranges show that at the lower limit of the high offset frequency (100kHz), outliers are few, with α≈0.5 and β≈0.2, resulting in an accuracy improvement of approximately 2dB. At the upper limit of the high offset frequency (1MHz), adjacent instrument noise and outliers are dense, with α≈1.6 and β≈0.3, resulting in an accuracy improvement of approximately 3dB. In the extremely high offset near-noise-floor region (frequency offset greater than 1MHz), spurious emissions are dense, with α≈2.2 and β≈0.35, resulting in an accuracy improvement of approximately 4.8dB. Therefore, in the high offset frequency range of 100kHz to 1MHz and above, the accuracy improvement of Huber loss relative to the classic MSE method naturally falls within the 2–5dB range.

[0049] The training process is as follows: Pure baseband symbols are generated according to the 5G NR NTN standard waveform, with modulation methods supporting QPSK, 16QAM, 64QAM, and 256QAM; based on the measured parameters of the Ka-band transceiver oscillator, power-law spectral phase noise containing both flicker and thermal noise is generated, covering low-frequency components near the carrier frequency and high-frequency components far from the carrier frequency; the phase noise is superimposed on the pure signal, and additive white Gaussian noise (AWGN) with different signal-to-noise ratios is superimposed to construct disturbed received signal samples; the superimposed real phase noise sequence is used as a label and paired one-to-one with the I / Q feature tensor of the disturbed signal to form supervised training sample pairs; after training, the model is weighted and pruned to transform it into a lightweight inference model, which is then deployed to the receiver baseband processing unit.

[0050] Step 3: Online phase noise compensation.

[0051] For each of the K received symbols within an observation window, the original symbols and the phase noise estimates output by the network in step two are subjected to phase inversion compensation to obtain clean symbols after phase noise removal, which are then used for subsequent demodulation and decoding.

[0052] The (K, 2)-dimensional feature tensor constructed in step one above directly uses the original I / Q components, fully preserving the phase continuity and I / Q coupling relationship between symbols. This provides the network with a learnable noise evolution trajectory, a prerequisite for high-precision estimation of power-law spectral noise. This tensor has a regular two-dimensional structure in both the time and channel dimensions, and can be directly processed alternately by the token mixing and channel mixing modules without dimension transformation, information loss, or computational redundancy, making it the optimal input form for the MLP-Mixer architecture. Furthermore, feature extraction is performed only after synchronization and matched filtering and before demodulation, without relying on subsequent modules such as carrier recovery, phase estimation, and decision-making, thus avoiding error propagation. This allows for stable acquisition of clean observation samples even in low signal-to-noise ratio, dynamic NTN channels.

[0053] In step two above, the token mixing MLP performs a fully connected mapping along the time axis, enabling global temporal mixing of consecutive symbols within the window. This naturally matches the temporal characteristics of the Ka-band phase noise, which features short-term strong correlation, slow drift, and high-frequency jitter. It is not limited to local convolutional windows and can fully learn the cumulative drift and temporal dependence of phase noise among K consecutive symbols, accurately capturing the temporal correlation of slow-varying components near the carrier frequency. The pure MLP structure has no convolution or local attention bias, which can adaptively fit the nonlinear temporal evolution of power-law spectral phase noise. It fits the real temporal distribution better than fixed-window filtering, and the time dimension of the input tensor is completely aligned with the token mixing dimension, without dimension transformation or information loss, allowing for direct and efficient extraction of temporal correlation features. Building upon this foundation, the channel-mixing MLP performs fully connected feature mixing along the I / Q channel dimensions. This allows for direct joint modeling and cross-channel information interaction of the real and imaginary signals under the same symbol. It naturally adapts to the strong coupling characteristics of phase noise in orthogonal channels, where the I and Q components of the same symbol are perturbed by the same phase noise. Channel mixing can directly learn the common perturbation patterns of the two signals, accurately capturing the coupling distortion caused by phase noise. Through fully connected mapping along the channel dimensions, it automatically mines the nonlinear correlation between I and Q signals without requiring manual orthogonal constraints. This preserves noise coupling information more completely than single-channel independent processing, and the channel dimensions of the input tensor are perfectly aligned with the channel mixing dimension, resulting in no redundant transformations or loss of coupling information. This allows for efficient extraction of orthogonal channel coupling features. Combined with residual connections and layer normalization structures, it stabilizes the network's training and inference processes. It weights and fuses the features extracted from the two dimensions dimension-wise to form a joint feature representation, ensuring the accuracy of the final symbol-by-symbol phase noise prediction.

[0054] It is understandable that the MLP-Mixer's dual-layer feature mixing mechanism can simultaneously capture the time, frequency correlation, and I / Q channel coupling characteristics of phase noise. It can effectively estimate phase noise at both near and far carrier frequencies under power-law spectral distribution. Combined with the Huber loss function's suppression of outliers at high bias frequencies, it achieves a 2–5 dB improvement in compensation accuracy at high bias frequencies. Furthermore, the MLP-Mixer consists only of fully connected layers, without convolution or attention mechanisms. The computational cost per inference is on the order of O(K²) (K is the window length, typically ≤32), allowing it to run directly on embedded processors in space or ground stations in real time. It can also adaptively track changes in the statistical characteristics of phase noise under different oscillators and environments online. Moreover, it only requires adding a lightweight network inference module to the digital baseband receiver link, without affecting the RF front-end, synchronization, or decoding modules, resulting in minimal modifications to existing systems.

[0055] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0056] In one embodiment, a non-terrestrial network Ka-band phase noise estimation apparatus is provided, comprising: The signal acquisition module is used to acquire the disturbed received signal after synchronization and matched filtering in the Ka band of the non-terrestrial network, and extract the data of each received symbol in the preset pilot band or decision feedback band. The tensor construction module is used to extract the real and imaginary components of each received symbol data to obtain a two-dimensional vector corresponding to a single received symbol. Taking multiple consecutive received symbols as an observation window, the two-dimensional vectors corresponding to all received symbols in the observation window are concatenated in chronological order to construct a two-dimensional feature tensor. A noise estimation module is used to input the two-dimensional feature tensor into a pre-trained multilayer perceptron hybrid network. The multilayer perceptron hybrid network includes alternating stacked token-mixing multilayer perceptron modules and channel-mixing multilayer perceptron modules. The token-mixing multilayer perceptron module performs global mixing operations on the two-dimensional feature tensor along the time dimension to extract the temporal correlation features of the phase noise between symbols. The channel-mixing multilayer perceptron module performs cross-channel mixing operations on the two-dimensional feature tensor along the channel dimension to extract the common structural features of the real and imaginary signals. The fused features obtained by alternating multilayer mixing are then stabilized by residual connections and layer normalization to map the estimated phase noise value corresponding to each received symbol within the observation window. The result output module is used to perform phase inverse rotation compensation on each received symbol in the observation window based on the phase noise estimate, so as to obtain the received symbol after phase noise elimination.

[0057] Specific limitations regarding the Ka-band phase noise estimation device for non-terrestrial networks can be found in the limitations of the Ka-band phase noise estimation method for non-terrestrial networks described above, and will not be repeated here. Each module in the aforementioned Ka-band phase noise estimation device for non-terrestrial networks can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in hardware or independent of the processor in a computer device, or stored in software in the memory of a computer device, so that the processor can call and execute the corresponding operations of each module.

[0058] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 2 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When executed by the processor, the computer program implements a Ka-band phase noise estimation method for non-terrestrial networks. The display screen can be an LCD screen or an e-ink display screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0059] Those skilled in the art will understand that Figure 2The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0060] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the method described above.

[0061] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0062] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for estimating Ka-band phase noise in non-terrestrial networks, characterized in that, The method includes: Acquire the disturbed received signal in the Ka band of a non-terrestrial network after synchronization and matched filtering, and extract the data of each received symbol in the preset pilot band or decision feedback band; For each received symbol data, extract the real and imaginary components to obtain a two-dimensional vector corresponding to a single received symbol. Using multiple consecutive received symbols as an observation window, concatenate the two-dimensional vectors corresponding to all received symbols in the observation window in chronological order to construct a two-dimensional feature tensor. The two-dimensional feature tensor is input into a pre-trained multilayer perceptron hybrid network. The multilayer perceptron hybrid network includes alternating stacked token-mixing multilayer perceptron modules and channel-mixing multilayer perceptron modules. The token-mixing multilayer perceptron module performs global mixing operations on the two-dimensional feature tensor along the time dimension to extract the temporal correlation features of the phase noise between symbols. The channel-mixing multilayer perceptron module performs cross-channel mixing operations on the two-dimensional feature tensor along the channel dimension to extract the common structural features of the real and imaginary signals. The final fused features obtained by alternating multilayer mixing are then mapped to the estimated phase noise value corresponding to each received symbol within the observation window after residual connection and layer normalization stabilization processing. Based on the phase noise estimate, phase inversion compensation is performed on each received symbol in the observation window to obtain the received symbol after phase noise elimination. The steps for training the multilayer perceptron hybrid network include: A training sample set is obtained, wherein each training sample in the training sample set includes a two-dimensional feature tensor of the disturbed received signal and a corresponding true phase noise label; the true phase noise is a phase noise sequence that conforms to the power-law distribution characteristics of the Ka band. The multilayer perceptron hybrid network is iteratively trained using the training sample set and the Huber loss function until the iteration stops when the stopping condition is met, thus obtaining the trained multilayer perceptron hybrid network.

2. The method according to claim 1, characterized in that, The two-dimensional vectors corresponding to all received symbols within the observation window are concatenated in chronological order to construct a two-dimensional feature tensor, which includes: Based on the two-dimensional vectors corresponding to each received symbol, the vectors are regularized with the time dimension as the first dimension and the channel dimension as the second dimension to obtain a two-dimensional feature tensor.

3. The method according to claim 1, characterized in that, The token-hybrid multilayer perceptron module includes a first fully connected layer, a first activation layer, and a second fully connected layer cascaded in sequence. Performing a global mixing operation on the two-dimensional feature tensor along the time dimension to extract the temporal correlation features of inter-symbol phase noise includes: The input two-dimensional feature tensor is transposed to switch the time dimension to the channel dimension; The transposed feature tensor is sequentially input into the first fully connected layer for linear mapping, the first activation layer for nonlinear transformation, and the second fully connected layer for quadratic linear mapping to obtain intermediate features. The intermediate features are reversed in the opposite dimension to obtain the temporal correlation features of inter-symbol phase noise.

4. The method according to claim 1, characterized in that, The channel hybrid multilayer sensor module includes a third fully connected layer, a second activation layer, and a fourth fully connected layer cascaded in sequence. Performing cross-channel mixing operations on the two-dimensional feature tensor along the channel dimension to extract common structural features of the real and imaginary signals includes: The input two-dimensional feature tensor is sequentially input into the third fully connected layer for linear mapping, the second activation layer for nonlinear transformation, and the fourth fully connected layer for quadratic linear mapping, to obtain the common structural features of the real and imaginary signals.

5. The method according to claim 1, characterized in that, After each group of alternating token-mixed multilayer perceptron modules and channel-mixed multilayer perceptron modules, a residual connection unit and a layer normalization unit are connected in series. The fused features obtained through multi-layer alternating mixing are stabilized by residual connection and layer normalization, and then mapped to obtain the phase noise estimate for each received symbol within the observation window, including: After each set of features is extracted by the token hybrid multilayer perceptron module and the channel hybrid multilayer perceptron module, the input features and output features of the current group are input to the residual connection unit for element-wise addition. The addition result is then input to the layer normalization unit for normalization to obtain the output features of the current group. After processing all the alternately stacked module groups in sequence, the final fused feature is obtained; The final fused features are input into the fully connected output layer to perform a linear mapping operation, thereby obtaining the phase noise estimate for each received symbol within the observation window.

6. The method according to claim 1, characterized in that, The method further includes: Weight quantization and structural pruning are performed on the trained multilayer perceptron hybrid network to obtain a lightweight inference model. The lightweight inference model is deployed to the receiver baseband processing unit.

7. The method according to claim 1, characterized in that, The Huber loss function employs a squared penalty term for estimation errors less than a preset threshold and a linear penalty term for estimation errors greater than a preset threshold. The preset threshold is set based on the noise statistical characteristics at high bias frequencies.

8. A non-terrestrial network Ka-band phase noise estimation device applied to the method described in any one of claims 1-7, characterized in that, The device includes: The signal acquisition module is used to acquire the disturbed received signal after synchronization and matched filtering in the Ka band of the non-terrestrial network, and extract the data of each received symbol in the preset pilot band or decision feedback band. The tensor construction module is used to extract the real and imaginary components of each received symbol data to obtain a two-dimensional vector corresponding to a single received symbol. Taking multiple consecutive received symbols as an observation window, the two-dimensional vectors corresponding to all received symbols in the observation window are concatenated in chronological order to construct a two-dimensional feature tensor. A noise estimation module is used to input the two-dimensional feature tensor into a pre-trained multilayer perceptron hybrid network. The multilayer perceptron hybrid network includes alternating stacked token-mixing multilayer perceptron modules and channel-mixing multilayer perceptron modules. The token-mixing multilayer perceptron module performs global mixing operations on the two-dimensional feature tensor along the time dimension to extract the temporal correlation features of the phase noise between symbols. The channel-mixing multilayer perceptron module performs cross-channel mixing operations on the two-dimensional feature tensor along the channel dimension to extract the common structural features of the real and imaginary signals. The fused features obtained by alternating multilayer mixing are then stabilized by residual connections and layer normalization to map the estimated phase noise value corresponding to each received symbol within the observation window. The result output module is used to perform phase inverse rotation compensation on each received symbol in the observation window based on the phase noise estimate, so as to obtain the received symbol after phase noise elimination.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

Images