A digital post-distortion method based on time convolution network

Abstract

The present application belongs to the technical field of wireless communication signal processing, and particularly relates to a digital post-distortion method based on a timing convolution network, which is suitable for baseband signal recovery in a wireless communication system affected by power amplifiers and radio frequency link nonlinearity. The present application obtains an observed baseband signal, a reference baseband signal and optional working state information, and constructs a timing input containing in-phase / quadrature, amplitude, phase, difference, lag term and state description quantity. The present application extracts distortion memory features of different history spans through feature projection, distortion state perception modulation and multi-scale gated hollow causal convolution memory block, and adopts a residual inverse mapping readout module to separate linear straight-through and nonlinear correction. The present application combines in-phase / quadrature error, amplitude error, phase consistency error and timing difference error to optimize the model. The present application can improve amplitude and phase recovery accuracy, enhance state adaptability, and improve parallel training and engineering deployment efficiency.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication signal processing technology, specifically relating to a digital post-distortion method based on temporal convolutional networks. Background Technology

[0002] In wireless communication transmission systems, power amplifiers (PAs) typically need to operate in a high-efficiency range. As the input power increases, power amplifiers are prone to entering nonlinear regions, leading to in-band distortion, out-of-band spectral spread, and amplitude and phase distortion in the output signal. To compensate for the nonlinear distortion of the power amplifier, linearization methods such as digital pre-distortion (DPD) or digital post-distortion (DPoD) can be used.

[0003] Digital predistortion is typically deployed at the transmitter to pre-compensate the signal to be transmitted before the power amplifier. Digital postdistortion is typically deployed at the receiver or in the receiver-side processing link to recover the desired reference signal from observed signals that have been affected by nonlinearities. Compared to transmitter-side pre-compensation, digital postdistortion can directly utilize the end-to-end distortion results observed at the receiver, making it suitable for receiver-side compensation, offline modeling, link-level simulation, receiver enhancement, and some applications where it is not convenient to obtain the internal signals of the transmitter.

[0004] However, in digital post-distortion learning, the mapping to be learned is usually not a simple static nonlinear mapping. Real power amplifiers and their associated RF links may simultaneously exhibit memory effects, state-switching effects, broadband dynamic effects, and phase-dependent responses. That is, the recovery result at the current moment depends not only on the current observation sample but also on multiple historical time-stamped samples, power states, phase configurations, temperature states, control variables, or scene labels. The model input is typically the already distorted observed output signal, and the model output is the reference signal to be recovered. This inverse recovery task differs from the forward pre-compensation task of digital pre-distortion at the transmitter; its model needs to recover the original amplitude, phase, and timing dynamics from the nonlinearly affected received sequence under conditions of limited observation at the receiver, significant state changes, and prominent amplitude-phase coupling distortion.

[0005] Existing neural network-based modeling methods mostly employ recurrent structures such as recurrent neural networks, long short-term memory networks, or gated recurrent units. While recurrent structures can utilize historical information through hidden state propagation, their time-by-time recursive computation limits the parallelism of training and inference. Furthermore, during long-sequence training, they may still suffer from long gradient propagation paths, high training time, or high engineering deployment complexity. Temporal convolutional networks (TCNs) are a class of convolutional neural network structures suitable for temporal modeling. They ensure temporal causality through causal convolution, expand the temporal receptive field through dilated convolution, and improve the training stability of deep networks through residual connections. Compared to recurrent neural networks, TCNs do not rely on time-by-time recursion of hidden states and can extract multi-scale historical features in parallel within a fixed historical window. Therefore, they are suitable for tasks involving nonlinear memory effects in power amplifiers, dynamic state changes, and temporal signal recovery.

[0006] However, directly applying TCN to digital post-distortion still presents the following problems: First, TCN lacks explicit encoding for the amplitude, power, phase difference, and state changes of the distorted observed signal; second, the nonlinear memory depth varies under different power ranges, temperature conditions, or beam configurations, making it difficult for fixed-hole-rate convolutions to adaptively select the effective historical span; third, digital post-distortion inverse mapping typically includes a near-linear pass-through portion and a power-varying nonlinear correction portion, which the TCN readout layer struggles to separate; fourth, when trained solely on in-phase / quadrature errors, issues may arise where amplitude recovery is correct but phase deviation is large, or phase consistency is achieved but frequency domain dynamic recovery is insufficient.

[0007] Therefore, a digital post-distortion method based on temporal convolutional networks is needed. Building upon the advantages of TCN causal convolution, dilated convolution, and parallel computing, a dedicated module is set up to address the technical pain points of digital post-distortion: constructing a model that preserves in-phase / orthogonal, amplitude, phase, and differential information using complex domain input; adjusting the model's feature response according to changes in power, temperature, phase configuration, or dynamic scene using distortion state-aware modulation; adaptively selecting the historical memory scale using multi-scale gated dilated causal convolution; separating linear passthrough and nonlinear correction using residual inverse mapping readout; and improving the engineering applicability of the recovered signal using amplitude-phase coherence and temporal differential joint loss. Summary of the Invention

[0008] This invention proposes a digital postdistortion method based on temporal convolutional networks to address the problems of existing static modeling methods being unable to characterize the nonlinear memory effect of power amplifiers, the limited parallel efficiency of existing recursive temporal models, and the lack of complex domain consistency constraints and state-based modeling mechanisms in existing digital postdistortion learning schemes, thereby improving the signal recovery accuracy of digital postdistortion learning.

[0009] The technical solution adopted in this invention is:

[0010] A digital post-distortion method based on temporal convolutional networks is used for digital post-distortion learning. The method includes a feature projection layer, a distortion state-aware modulation module, a multi-scale gated dilated causal convolutional memory block, and a residual inverse mapping readout module, and includes the following steps:

[0011] S1. Obtain training samples, which include observed baseband signals, reference baseband signals, and optional operating state information, respectively represented as follows:

[0012] ,

[0013] ,

[0014] in, Represents the current sampling point. Indicates the observed baseband signal. Indicates the reference baseband signal. , , and These represent the corresponding in-phase and quadrature components, respectively;

[0015] S2. Construct a timing input sequence based on the observed baseband signal and operating state information. The timing input sequence includes in-phase components, quadrature components, complex domain features, and distortion state description quantities used to characterize the nonlinear strength, memory depth, and operating state changes of the power amplifier. The timing input sequence includes a sliding time window composed of feature vectors from the current time and historical time.

[0016] ,

[0017] in, For sequence length, For the first Complex domain eigenvectors of each sampling point;

[0018] S3. Input the time-series input sequence into the digital post-distortion temporal convolutional network; the feature projection layer maps the complex domain feature vector to the hidden feature space; the distortion state-aware modulation module performs channel modulation or gating on the hidden features according to the working state information and the distortion state descriptor; the multi-scale gated dilated causal convolutional memory block extracts historical memory features related to the digital post-distortion inverse mapping on multiple dilation rate branches, and adaptively fuses different memory scales according to the distortion state descriptor to obtain a temporal feature map;

[0019] S4. Obtain a global time series representation based on the time series feature map, and output the prediction result from the observed baseband signal to the reference baseband signal through the residual inverse mapping readout module, wherein the residual inverse mapping readout module includes a linear direct-through branch and a nonlinear residual correction branch;

[0020] S5. Construct a digital post-distortion learning loss function based on the prediction result and the reference baseband signal. The digital post-distortion learning loss function includes one or more of the following: in-phase / orthogonal error term, amplitude error term, phase consistency error term, temporal difference error term, and regularization term. Optimize the model parameters of the temporal convolutional network using the digital post-distortion learning loss function.

[0021] S6. Use the trained model for digital post-distortion learning or receiver-side baseband signal recovery. During the inference process, only the baseband signals observed at the current time and historical time are used to generate the reference baseband signal estimate at the current time.

[0022] Furthermore, in the training samples obtained by S1, the operating state information includes one or more of the following: power state, input power, output power, average power, peak-to-average power ratio, power range, phase configuration, beam control quantity, temperature state, bandwidth configuration, modulation method, scene label, and dynamic control quantity.

[0023] Furthermore, the complex domain features described in S2 include one or more of the following: in-phase component, quadrature component, amplitude, phase sine and cosine, amplitude power term, phase difference sine and cosine, amplitude difference, in-phase difference, quadrature difference, lag term, memory polynomial term, state feature, and angle sine and cosine feature; the memory polynomial term includes The characteristics of form, among which, Indicates the order of memory lag. It represents the order of nonlinearity.

[0024] Furthermore, the dedicated temporal convolutional network for digital post-distortion includes an input projection layer, multiple temporal convolutional blocks, and a readout layer. Specifically, the multiple temporal convolutional blocks include a distortion state-aware modulation module and a multi-scale gated dilated causal convolutional memory block, used to dynamically adjust the convolutional channel response and memory scale according to the amplitude, power, phase changes, and operating state of the observed baseband signal. The readout layer is specifically a residual inverse mapping readout module, used to output the predicted values ​​of the in-phase and quadrature components of the reference baseband signal.

[0025] Furthermore, the multi-scale gated dilated causal convolutional memory block includes at least two dilated causal convolutional branches, with different dilated causal convolutional branches having different dilation rates; for the first... The multi-scale fusion features and output representation of the layer-level temporal convolutional block are as follows:

[0026] ,

[0027] ,

[0028] ,

[0029] in, For temporal convolutional block layer indexes, For the first Multi-scale fusion features of layers For the first Layer output features, The number of branches in a dilated causal convolution. For branch index, Indicates the kernel length is void ratio causal convolution, The input features are after distortion state-aware modulation. For the first Input features of layer-time convolutional blocks, It is a distortion state description quantity composed of observation amplitude, power, phase difference, and operating state information. and For the first Trainable parameters of layer-gated networks For the first The gating weights corresponding to each memory scale Indicates in Normalize each branch dimension. Represents a non-linear activation function. This represents an identity mapping or a linear projection used for channel alignment.

[0030] Furthermore, the distortion state-aware modulation module generates channel scaling coefficients, channel bias coefficients, or gating coefficients based on the operating state information and the distortion state descriptor, and modulates the input or output features of the temporal convolutional block. The modulation methods include:

[0031] ,

[0032] ,

[0033] ,

[0034] ,

[0035] in, For work status information, This is a quantity describing the distorted state. This represents the concatenated vector of the two vectors. For the first Layer state embedding vector, It is a nonlinear transformation. , , , , and All of these are trainable parameters. This is the channel scaling factor. This is the channel offset coefficient.

[0036] Furthermore, in S4, the prediction result of the reference baseband signal is output through the residual inverse mapping readout module, specifically by adding the output of the linear direct-through branch to the output of the nonlinear residual correction branch:

[0037] ,

[0038] in, For the reference baseband signal prediction result, a two-dimensional real vector, To observe the baseband signal at the 1st A two-dimensional real vector of sampling points, For a fixed or trainable two-dimensional linear transformation matrix, The parameter is The nonlinear residual correction branch, For global timing representation, For work status information, The distortion state is described by the linear pass-through branch, which is used to maintain the linear mapping of the low distortion region, and the nonlinear residual correction branch is used to learn the nonlinear post-distortion correction amount caused by the power amplifier and RF link.

[0039] Furthermore, in S4, the method for obtaining the global temporal representation based on the temporal feature map is to fuse the feature map of the last time step with the state-aware attention pooling feature; the state-aware attention pooling feature is represented as follows:

[0040] ,

[0041] ,

[0042]

[0043] in, For the time series feature map in the 1st The characteristics of each moment Characteristics of the final moment For the first State perception attention weights at each moment Indicates in Normalize each time dimension. , , and These are the trainable parameters of the attention network. For attention pooling features, For global timing representation, This represents a splicing, linear mapping, or gated fusion function.

[0044] Furthermore, the digital post-distortion learning loss function in S5 is specifically as follows:

[0045] ,

[0046] ,

[0047] ,

[0048] ,

[0049] ,

[0050] in, For the total loss function, , , , and These are non-negative weighting coefficients. This is the in-phase / quadrature error term. For amplitude error term, For phase consistency error term, This is the timing difference error term. For model parameter regularization, The total number of training samples; , , , , representing the amplitude value of the predicted signal and the instantaneous phase angle of the reference signal, respectively; , These represent the dynamic changes of the predicted signal and the reference signal, respectively.

[0051] The beneficial effects of this invention are as follows:

[0052] 1) This invention uses a temporal convolutional network to model the received signal, which can extract the memory features required for nonlinear inverse modeling of the power amplifier using historical observation samples while satisfying temporal causality;

[0053] 2) This invention introduces a distortion state-aware modulation module, enabling the network to change the convolutional channel response according to power state, amplitude changes, temperature state, phase configuration, or dynamic scene.

[0054] 3) This invention introduces a multi-scale gated dilated causal convolutional memory block, which covers different historical spans through branches with different dilation rates, and uses the distortion state description quantity to select a memory scale that is more suitable for the current digital post-distortion sample;

[0055] 4) This invention introduces a residual inverse mapping structure, which combines a linear direct-path branch with a nonlinear residual correction branch, which is beneficial for the stable recovery of the low-distortion region and the fine correction of the high-power nonlinear region.

[0056] 5) This invention adopts a convolutional parallel modeling method, which has better parallel training capabilities and engineering deployment efficiency compared with time-by-time recursive time series models;

[0057] 6) This invention unifies the in-phase / quadrature components, amplitude, phase, difference, lag term and state features into a time-series input, which is beneficial to enhancing the model's ability to represent dynamic systems;

[0058] 7) This invention improves the amplitude-phase consistency and dynamic consistency of the recovered signal by jointly constraining the training process through in-phase / quadrature error, amplitude error, phase consistency error and timing difference error.

[0059] 8) This invention supports grouping or joint modeling according to operating status, power range, modulation method, bandwidth configuration, phase configuration, beam status or dynamic scene label, and has good engineering scalability;

[0060] 9) The reasoning process of this invention relies only on the observation samples at the current time and historical time, and can be used for digital post-distortion learning, nonlinear inverse modeling of power amplifiers, receiver compensation and other dynamic system modeling tasks. Attached Figure Description

[0061] Figure 1 A schematic diagram of a digital post-distortion method based on temporal convolutional networks is provided in this embodiment of the invention.

[0062] Figure 2 This is a schematic diagram of a temporal convolutional network structure provided in an embodiment of the present invention;

[0063] Figure 3 This is a schematic diagram of a temporal convolutional block structure provided in an embodiment of the present invention;

[0064] Figure 4 This is a schematic diagram of group state modeling and model parameter selection provided in an embodiment of the present invention. Detailed Implementation

[0065] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings and embodiments:

[0066] Example:

[0067] like Figure 1 As shown, this embodiment provides a digital post-distortion method based on temporal convolutional networks. It uses TCN as the basis for temporal memory extraction and adds modules such as feature projection, distortion state-aware modulation, multi-scale gated dilated causal convolutional memory, and residual inverse mapping readout to match the network structure with the inverse recovery characteristics of digital post-distortion. The method includes the following steps.

[0068] S101, Obtain training samples. Training samples include observed baseband signals. Reference baseband signal And optional work status information The observed baseband signal can be a signal obtained after passing through a power amplifier, RF link, channel, or receiver processing, while the reference baseband signal can be the original baseband signal to be recovered or an ideal reference signal after synchronization.

[0069] S102, Construct the time-series input sequence. The observed baseband signal is decomposed into in-phase and quadrature components, and further input features are constructed, including amplitude, phase sine / cosine, amplitude power term, phase difference sine / cosine, amplitude difference, lag term, memory polynomial term, state features, and distortion state descriptors. For each target time point, a sequence of length is formed in chronological order. A sliding window.

[0070] S103 is a post-distorted temporal convolutional network that takes a temporal input sequence as a digital input. The network first maps the input features to the hidden space through a feature projection layer, then adjusts the channel response according to the current state through a distortion state-aware modulation module, and then extracts historical memory features at different time spans through multi-scale gated dilated causal convolutional memory blocks, finally outputting a temporal feature map.

[0071] S104. Obtain the global temporal representation based on the temporal feature map. The fusion result of the last-moment features, state-aware attention pooling features, and working state information can be used as the global temporal representation so that the readout layer can simultaneously obtain the short-term features, long-term memory features, and state conditions required for recovery at the current moment.

[0072] S105 outputs the prediction result of the reference baseband signal through the residual inverse mapping readout module. The readout module adds the linear direct-through branch and the nonlinear residual correction branch, outputting a two-dimensional real vector. The predicted signal is reconstructed from this two-dimensional real vector. .

[0073] S106. Construct the digital post-distortion learning loss function and optimize the model parameters. The loss function may include in-phase / orthogonality error, amplitude error, phase consistency error, temporal difference error, and regularization term. Update the model parameters through backpropagation until the training termination condition is met.

[0074] S107, inference is performed using the trained model. In the inference phase, the observed baseband signal to be processed is constructed as the same time-series input sequence, and the corresponding normalization parameters, state modulation parameters, readout layer parameters or loss weight configurations are selected according to the current working state. The trained model outputs the digital post-distortion recovery result.

[0075] Based on the above process, this embodiment can integrate the historical dependency of the baseband signal, amplitude and phase information, distortion intensity information and optional operating state information into the modeling process.

[0076] In this embodiment, the observed baseband signal and the reference baseband signal are respectively represented as:

[0077] ,

[0078] ,

[0079] in, Represents the current sampling point. Indicates the observed baseband signal. Indicates the reference baseband signal. , , and These represent the corresponding in-phase and quadrature components, respectively. For the observed baseband signal, the following basic characteristics can be constructed:

[0080] ,

[0081] , ,

[0082] in, This indicates the amplitude of the observed baseband signal at the current moment. This represents the instantaneous phase angle of the observed baseband signal at the current moment. and Let represent the sine and cosine values ​​of the phase, respectively. Using sine and cosine representations of the phase can reduce the impact of phase jumps on model training. Furthermore, dynamic difference features can be constructed:

[0083] ,

[0084] , ,

[0085] , ,

[0086] in, , , , , These represent the first-order difference terms indicating the amplitude, in-phase / quadrature components, and sine and cosine values ​​of the observed baseband signal at the current moment, respectively. To enhance the representation of the nonlinear order and memory effect of the power amplifier, an amplitude power term can be constructed. and memorizing polynomial terms :

[0087] , ,

[0088] , , ,

[0089] in, Indicates the order of the magnitude power. Indicates the order of the maximum magnitude power. Indicates the order of memory lag. This represents the maximum memory lag order. Represents the nonlinear order. This represents the maximum nonlinear order. In one possible implementation, the _th_... The input feature vector at time n can be represented as:

[0090] ,

[0091] in, This is optional working status information. This is a descriptor of the distortion state. The descriptor of the distortion state can consist of the instantaneous amplitude of the observed signal, short-time average power, peak-to-average power ratio, phase difference statistics, temperature state, power range, or other variables that can reflect the nonlinearity and memory depth. For sequence length... The model input can be represented as:

[0092] ,

[0093] in, For sequence length, For the first The complex domain feature vector of each sampling point. In digital post-distortion learning, the model needs to recover the reference input signal from the observed output sequence. The basic inverse mapping can be represented as:

[0094] ,

[0095] in, The parameter is The digital post-distortion inverse mapping model, This represents the historical memory depth span introduced when considering RF link memory effects. When considering operational state information and distortion state descriptors, this mapping can be extended to:

[0096] ,

[0097] The above extensions enable the model to change the inverse mapping relationship according to different power states, different temperature states, or different dynamic scenarios, thus making it more suitable for engineering problems in digital post-distortion learning where "the same observation amplitude corresponds to different correction amounts under different states".

[0098] like Figure 2 As shown, a dedicated temporal convolutional network for digital post-distortion receives sequences. The network can generate and output temporal feature maps. It may include a feature projection layer, a distortion-state-aware modulation module, a multi-scale gated dilated causal convolutional memory block, a global temporal fusion unit, and a residual inverse mapping readout module. The feature projection layer is used to transform complex domain feature vectors... Mapping to the hidden feature space:

[0099] ,

[0100] in, For the first Projection features at each time point, and These are the trainable weights and biases of the feature projection layer, respectively. This is a non-linear activation function. This projection layer maps in-phase / orthogonal, amplitude-phase, differential, and state-related features to a unified dimension, facilitating subsequent processing by the TCN module. The distortion state-aware modulation module adjusts the modulation based on the operating state information. and distortion state description quantity Generate channel scaling and channel bias coefficients, and modulate the projected or convolutional features:

[0101] ,

[0102] , ,

[0103] ,

[0104] in, For work status information, This is a quantity describing the distorted state. For the first Layer state embedding vector, It is a nonlinear transformation. , , , , and All of these are trainable parameters. This is the channel scaling factor. This represents the channel offset coefficient. Through the above modulation method, when the input power, temperature, beam control parameters, or phase configuration change, the network can alter the response intensity of different channels; where, Used to amplify or suppress channels associated with the current distortion state. This is used to compensate for state-related offsets, thereby avoiding the use of the same set of fixed feature responses in all states by a general TCN. For the 1st The input feature vector at time step n, the output of the causal convolution is:

[0105] ,

[0106] in, Indicates the kernel length. Indicates the kernel position index. Indicates the convolution kernel number 1 Weight parameters for each position, This represents the input features corresponding to a historical moment. Indicates the current convolutional layer is at the th . The output characteristics at each time step. From the above equation, we can see that at time step... The output depends only on the input at the current and historical moments, and does not depend on the input at future moments. Therefore, it can meet the requirements of online processing or causal processing in digital post-distortion learning. To expand the receptive field, an expansion coefficient of can be used. The empty causal convolution, in which This represents the time interval between adjacent convolution sampling points:

[0107] .

[0108] In one implementation, the dilation coefficient of a multi-layer temporal convolutional block can increase exponentially, or it can be selected by a gating mechanism based on the state of the digitized distorted samples. The exponential increase can be expressed as:

[0109] .

[0110] When the kernel length is The number of layers is At that time, the theoretical receptive field of a network can be expressed as:

[0111] .

[0112] in, Indicates the theoretical receptive field, Indicates the first The dilation coefficient of a layer-time convolutional block. This indicates the total number of layers in the temporal convolutional block.

[0113] like Figure 3 As shown, this embodiment further extends the single dilated causal convolution to a multi-scale gated dilated causal convolution memory block. This memory block includes... Each dilated causal convolutional branch corresponds to a different dilation rate. or kernel length It is used to cover short-term memory, medium-term memory, and long-term memory. The multi-scale fusion features and output representation of the temporal convolutional block are as follows:

[0114] ,

[0115] ,

[0116] ,

[0117] in, For temporal convolutional block layer indexes, For the first Multi-scale fusion features of layers For the first Layer output features, The number of branches in a dilated causal convolution. For branch index, Indicates the kernel length is void ratio causal convolution, The input features are after distortion state-aware modulation. For the first Input features of layer-time convolutional blocks, This is a quantity describing the distorted state. and For the first Trainable parameters of layer-gated networks, For the first The gating weights corresponding to each memory scale Indicates in Normalize each branch dimension. Represents a non-linear activation function. This represents an identity mapping or a linear projection used for channel alignment. When the observed signal is in a low-power or weakly nonlinear state, the model can rely more on short-term or linear memory branches; when the observed signal is in a high-power, significantly temperature-rising, or strongly nonlinear state, the model can increase the weight of long-memory branches to characterize the dynamic memory effects in the power amplifier and RF link.

[0118] In one implementation, each temporal convolutional block may include dilated causal convolutions, normalized layers, nonlinear activation functions, dropout layers, and residual connections. The nonlinear activation function can be ReLU, GELU, SiLU, or other activation functions. The normalized layer can be batch normalization, layer normalization, weight normalization, or conditional normalization combined with state modulation. Dropout layers are used to improve the model's generalization ability. To alleviate the gradient vanishing, convergence difficulties, and performance degradation problems in deep convolutional networks, this embodiment introduces residual connections into the temporal convolutional network. This reflects residual connections; when a multi-scale gated dilated causal convolutional memory block degenerates into a single convolutional branch... It can be represented as ,in Indicates the first Layered convolution transformation.

[0119] In one implementation, the output feature map of the last layer of the temporal convolutional network is:

[0120] .

[0121] The characteristics of the final moment are:

[0122] .

[0123] State-aware attention pooling features are:

[0124] ,

[0125] .

[0126] By fusing the last-moment features, attention pooling features, working state information, and distortion state descriptions, a global temporal representation can be obtained:

[0127] ,

[0128] in, For the time series feature map in the 1st The characteristics of each moment Characteristics of the final moment For the first State perception attention weights at each moment Indicates in Normalize each time dimension. , , and These are the trainable parameters for the attention network. For attention pooling features, For global timing representation, This represents a concatenation, linear mapping, or gated fusion function. Last-moment features provide causal recovery information for the current sampling point, attention-pooling features aggregate historical memory contributions, and working state information and distortion state descriptors condition the recovery process.

[0129] In this embodiment, the readout layer is a residual inverse mapping readout module, rather than a regular fully connected readout layer. The readout module includes a linear pass-through branch and a nonlinear residual correction branch, and its output is:

[0130] ,

[0131] in, For the reference baseband signal prediction result, a two-dimensional real vector, To observe the baseband signal at the 1st A two-dimensional real vector of sampling points, It can be an identity matrix, a fixed linear calibration matrix, or a trainable two-dimensional linear matrix. The parameter is The nonlinear residual correction branch, For global timing representation, For work status information, These are the distortion state descriptors. The linear direct-through branch represents the dominant linear relationship between the observed and reference signals, while the nonlinear residual correction branch learns the amplitude-phase nonlinearity, memory error, and state-related corrections introduced by the power amplifier and RF link. Thus, the model does not require a deep nonlinear network to forcibly fit an approximate identity mapping in the low-distortion region, while providing fine-grained corrections through the residual branch in the high-distortion region.

[0132] In one implementation, the training objective is used to constrain the error between the predicted signal and the reference signal. The predicted signal is represented as:

[0133] ,

[0134] The in-phase / quadrature error term is:

[0135] ,

[0136] The amplitude error term is:

[0137] ,

[0138] The phase consistency error term is:

[0139] ,

[0140] Compared to direct phase difference squared error, cosine-based consistency error reduces the impact of phase period jumps on training, making it more suitable for baseband digital post-distortion recovery. The timing difference error term is:

[0141] ,

[0142] Timing difference error is used to constrain the dynamic changes of the recovered signal, which is beneficial for improving the spectral and transient recovery deviations caused by memory effects in broadband signals. The total loss function is:

[0143] ,

[0144] in, For the total loss function, , , , and These are non-negative weighting coefficients. For model parameter regularization, The total number of training samples; , , , These represent the amplitude values ​​and instantaneous phase angles of the predicted signal and the reference signal, respectively. , These represent the dynamic changes in the predicted signal and the reference signal, respectively. During training, stochastic gradient descent, Adam, AdamW, or other optimization algorithms can be used to update the model parameters. Optionally, gradient pruning, learning rate scheduling, early stopping strategies, weight decay, and data augmentation can also be used to improve training stability.

[0145] like Figure 4 As shown, in one implementation, training samples can be divided into multiple groups of modeling samples. Different groups of modeling samples can correspond to different power ranges, modulation schemes, bandwidth configurations, phase configurations, beam states, or dynamic scene labels. For example, training samples can be divided according to the input power range. Group:

[0146] ,

[0147] For the Group modeling samples can maintain corresponding stateful modeling information:

[0148]

[0149] in, Represents the training sample set, Indicates the first Group modeling samples, Indicates the number of modeling sample groups. Indicates the first Group state-based modeling information, and These represent the normalized mean and standard deviation of the feature, respectively. Indicates the loss weight configuration. and Indicates optional state modulation parameters. This represents optional group-specific readout layer parameters or residual correction parameters. During training or inference, first state indication information can be obtained. This first state indication information indicates the first modeling sample group to which the current sample to be modeled belongs. Based on the first modeling sample group, the corresponding feature normalization parameters, state modulation parameters, loss weights, readout layer parameters, convolutional block parameters, or model parameters can be determined, and training or inference can be completed based on the determined parameters. In this way, the model can maintain shared temporal convolutional coding capabilities while selecting corresponding state parameters for different power ranges, different modulation methods, different temperature states, or different dynamic scenarios, thereby improving the accuracy of nonlinear inverse modeling under different operating states. After training, the model performance can be evaluated on the test dataset. The normalized mean square error is:

[0150] .

[0151] The error vector magnitude is:

[0152] .

[0153] Where NMSE is the normalized mean square error and EVM is the error vector magnitude. Other indicators such as mean absolute error of amplitude, mean absolute error of phase, complex correlation coefficient, improvement in adjacent channel leakage ratio, or spectral recovery error can also be used for evaluation as needed.

[0154] Although the present invention has been described in conjunction with specific embodiments, those skilled in the art can make modifications, substitutions, combinations, or variations to the above embodiments without departing from the spirit and scope of the invention. This specification and accompanying drawings are for illustrative purposes only, and the scope of protection of the present invention should be determined by the claims.

Claims

1. A digital post-distortion method based on temporal convolutional networks, characterized in that, A temporal convolutional network is constructed, comprising a feature projection layer, a distortion state-aware modulation module, a multi-scale gated dilated causal convolutional memory block, and a residual inverse mapping readout module. The temporal convolutional network is trained, and digital post-distortion recovery is performed using the trained network. Specifically, the following steps are included: S1. Obtain training samples, which include observed baseband signals, reference baseband signals, and optional operating state information, respectively represented as follows: ， ， in, Represents the current sampling point. Indicates the observed baseband signal. Indicates the reference baseband signal. , , and These represent the corresponding in-phase and quadrature components, respectively; S2. Construct a timing input sequence based on the observed baseband signal and operating state information. The timing input sequence includes in-phase components, quadrature components, complex domain features, and distortion state description quantities used to characterize the nonlinear strength, memory depth, and operating state changes of the power amplifier. The timing input sequence includes a sliding time window composed of feature vectors from the current time and historical time. ， in, For sequence length, For the first Complex domain eigenvectors of each sampling point; S3. Input the temporal input sequence into the temporal convolutional network; the feature projection layer maps the complex domain feature vector to the hidden feature space; the distortion state-aware modulation module performs channel modulation or gating on the hidden features according to the working state information and the distortion state descriptor; the multi-scale gated dilated causal convolutional memory block extracts historical memory features related to the digital post-distortion inverse mapping on multiple dilation rate branches, and adaptively fuses different memory scales according to the distortion state descriptor to obtain a temporal feature map; S4. Obtain a global time series representation based on the time series feature map, and output the prediction result from the observed baseband signal to the reference baseband signal through the residual inverse mapping readout module, wherein the residual inverse mapping readout module includes a linear direct-through branch and a nonlinear residual correction branch; S5. Construct a digital post-distortion learning loss function based on the prediction result and the reference baseband signal. The digital post-distortion learning loss function includes one or more of the following: in-phase / orthogonal error term, amplitude error term, phase consistency error term, temporal difference error term, and regularization term. Optimize the model parameters of the temporal convolutional network using the digital post-distortion learning loss function. S6. Use the trained model for digital post-distortion learning or receiver-side baseband signal recovery. During the inference process, only the baseband signals observed at the current time and historical time are used to generate the reference baseband signal estimate at the current time.

2. The digital post-distortion method based on temporal convolutional networks according to claim 1, characterized in that, In the training samples obtained by S1, the working state information includes one or more of the following: power state, input power, output power, average power, peak-to-average power ratio, power range, phase configuration, beam control quantity, temperature state, bandwidth configuration, modulation method, scene label, and dynamic control quantity.

3. The digital post-distortion method based on temporal convolutional networks according to claim 2, characterized in that, The complex domain features described in S2 are one or more of the following: in-phase component, quadrature component, amplitude, phase sine and cosine, amplitude power term, phase difference sine and cosine, amplitude difference, in-phase difference, quadrature difference, lag term, memory polynomial term, state feature, and angle sine and cosine feature; the memory polynomial term includes The characteristics of form, among which, Indicates the order of memory lag. It represents the order of nonlinearity.

4. The digital post-distortion method based on temporal convolutional networks according to claim 3, characterized in that, The temporal convolutional network includes an input projection layer, multiple temporal convolutional blocks, and a readout layer. The multiple temporal convolutional blocks specifically include a distortion state-aware modulation module and a multi-scale gated dilated causal convolutional memory block, used to dynamically adjust the convolutional channel response and memory scale according to the amplitude, power, phase changes, and operating state of the observed baseband signal. The readout layer is specifically a residual inverse mapping readout module, used to output the predicted values ​​of the in-phase and quadrature components of the reference baseband signal.

5. A digital post-distortion method based on temporal convolutional networks according to claim 4, characterized in that, The multi-scale gated dilated causal convolutional memory block includes at least two dilated causal convolutional branches, with different dilated causal convolutional branches having different dilation rates; for the first... The multi-scale fusion features and output representation of the layer-level temporal convolutional block are as follows: ， ， ， in, For temporal convolutional block layer indexes, For the first Multi-scale fusion features of layers For the first Layer output features, The number of branches in a dilated causal convolution. For branch index, Indicates the kernel length is void ratio causal convolution, The input features are after distortion state-aware modulation. For the first Input features of layer-time convolutional blocks, It is a distortion state description quantity composed of observation amplitude, power, phase difference, and operating state information. and For the first Trainable parameters of layer-gated networks, For the first The gating weights corresponding to each memory scale Indicates in Normalize each branch dimension. Represents a non-linear activation function. This represents an identity mapping or a linear projection used for channel alignment.

6. The digital post-distortion method based on temporal convolutional networks according to claim 5, characterized in that, The distortion state-aware modulation module generates channel scaling coefficients, channel bias coefficients, or gating coefficients based on the operating state information and the distortion state descriptor, and modulates the input or output features of the temporal convolutional block. The modulation methods include: ， ， ， ， in, For work status information, This is a quantity describing the distorted state. This represents the concatenated vector of the two vectors. For the first Layer state embedding vector, It is a nonlinear transformation. , , , , and All of these are trainable parameters. This is the channel scaling factor. This is the channel offset coefficient.

7. A digital post-distortion method based on temporal convolutional networks according to claim 6, characterized in that, In S4, the prediction result of the reference baseband signal is output through the residual inverse mapping readout module, specifically by adding the output of the linear direct-through branch to the output of the nonlinear residual correction branch: ， in, For the reference baseband signal prediction result, a two-dimensional real vector, To observe the baseband signal at the 1st A two-dimensional real vector of sampling points, For a fixed or trainable two-dimensional linear transformation matrix, The parameter is The nonlinear residual correction branch, For global timing representation, For work status information, The distortion state is described by the linear pass-through branch, which is used to maintain the linear mapping of the low distortion region, and the nonlinear residual correction branch is used to learn the nonlinear post-distortion correction amount caused by the power amplifier and RF link.

8. A digital post-distortion method based on temporal convolutional networks according to claim 7, characterized in that, The method in S4 for obtaining the global temporal representation based on the temporal feature map is to fuse the feature map of the last time step with the state-aware attention pooling feature; the state-aware attention pooling feature is represented as follows: ， ， ， in, For the time series feature map in the 1st The characteristics of each moment Characteristics of the final moment For the first State perception attention weights at each moment Indicates in Normalize each time dimension. , , and These are the trainable parameters for the attention network. For attention pooling features, For global timing representation, This represents a splicing, linear mapping, or gated fusion function.

9. A digital post-distortion method based on temporal convolutional networks according to claim 8, characterized in that, The digital post-distortion learning loss function in S5 is as follows: ， ， ， ， ， in, For the total loss function, , , , and These are non-negative weighting coefficients. This is the in-phase / quadrature error term. For amplitude error term, This is the phase consistency error term. This is the timing difference error term. For model parameter regularization, The total number of training samples; , , , These represent the amplitude values ​​and instantaneous phase angles of the predicted signal and the reference signal, respectively. , These represent the dynamic changes of the predicted signal and the reference signal, respectively.

Images