Self-adaptive digital predistorter design method and system based on deep learning

Abstract

The invention discloses a self-adaptive digital predistorter design method and system based on deep learning, belongs to the field of satellite communication, and solves the technical problems that the compensation effect is greatly reduced and the actual linearization requirement cannot be met due to the fact that model parameters are fixed, the dynamic change is faced in actual application and the online self-adaptive capability is lacked in the prior art. Training the deep learning model based on an offline training module to obtain a weight parameter; the weight parameters are configured to a forward reasoning module, and the forward reasoning module is used for initial digital pre-distortion processing; performing difference evaluation on the PA output signal subjected to the initial digital pre-distortion processing and the ideal reference signal by using a loss calculation module to obtain a real-time loss value; and back propagation is carried out on the real-time loss value by using a back propagation module, and a weight parameter of the forward reasoning module is updated. The method is used for realizing real-time adaptive adjustment of the digital predistorter.

Description

Technical Field

[0001] This invention relates to the field of satellite communications, and more specifically to a design method and system for an adaptive digital predistorter based on deep learning. Background Technology

[0002] Satellite communication boasts numerous advantages, including wide coverage, long transmission distance, large communication capacity, stable transmission quality, flexible networking, and strong security, making it one of the most competitive communication methods today. The power amplifier (PA), a key component in satellite communication systems, is directly affected by its nonlinear distortion, which impacts communication quality. To improve overall system performance, linearizing the PA has become an important research and engineering application direction. Digital predistortion (DPD), as an effective linearization method, applies preprocessing distortion opposite to the PA's characteristics to the input signal before transmission to counteract the nonlinear effects at the output, thereby significantly reducing out-of-band leakage and adjacent channel interference in the output signal, and improving spectral efficiency and communication reliability.

[0003] Traditional DPD design methods often employ polynomial models, such as MP (memory polynomial) and GMP (generalized memory polynomial). These methods offer advantages such as high modeling efficiency and ease of implementation when dealing with static or weakly dynamic nonlinear distortions. However, polynomial models suffer from insufficient modeling capabilities and poor parameter generalization when facing complex modulation schemes, high-bandwidth signals, or highly dynamic changing scenarios, making it difficult to meet the increasingly stringent linearization requirements of modern communication systems.

[0004] In recent years, deep learning technology has received widespread attention in the field of digital predistortion due to its powerful nonlinear fitting capabilities and end-to-end modeling advantages. By constructing deep neural networks (such as convolutional neural networks and recurrent neural networks) to model the input-output relationship of PA, it is possible to more accurately approximate complex nonlinear characteristics and demonstrate superior performance compared to traditional methods in high-dynamic, multi-frequency, and complex modulation environments.

[0005] However, existing deep learning-based digital predistortion methods generally suffer from a key problem: model training is typically performed offline, and the trained network parameters remain fixed after deployment to embedded devices or FPGA platforms, lacking an adaptive adjustment mechanism for environmental changes. In practical applications, signals may dynamically change due to temperature variations, frequency band switching, device aging, or modulation method changes. The compensation effect of fixed-parameter models will be significantly reduced, and may even fail to meet practical linearization requirements. Therefore, designing a deep learning predistorter with online adaptive capabilities, enabling it to maintain good performance in variable environments, has become one of the key challenges in current engineering practice.

[0006] In the prior art, Chinese patent document CN119892571A discloses a "Design Method for Digital Predistorter Model Based on Neural Networks." This method combines the amplitude / phase and isolation / interaction characteristics of orthogonal / in-phase branching of communication data. Based on an LSTM model, it designs a digital predistorter model using ensemble vector decomposition and multi-expert decision-making. It comprehensively considers in-phase / orthogonal components and amplitude / phase information, and uses a vector splicing module to perform cross-layer splicing of the original distorted signal and the phase recovery signal, learning the isolation and interaction between different features. However, this technical solution uses an offline training method, which may dynamically change in practical applications due to temperature variations, operating frequency band switching, device aging, or modulation method changes. This can significantly reduce the compensation effect of the fixed-parameter model, and may even fail to meet actual linearization requirements.

[0007] In summary, existing technologies suffer from the technical problem that, due to fixed model parameters, they lack online adaptive capabilities when facing dynamic changes in practical applications, resulting in a significant reduction in compensation effectiveness and failing to meet actual linearization requirements. Summary of the Invention

[0008] This invention solves the technical problem that existing technologies, due to fixed model parameters, lack online adaptive capabilities when facing dynamic changes in practical applications, resulting in a significant reduction in compensation effect and an inability to meet actual linearization requirements.

[0009] The present invention discloses a deep learning-based adaptive digital predistorter design method. The adaptive digital predistorter includes an offline training module, a forward inference module, a loss calculation module, and a backpropagation module. The adaptive digital predistorter design method includes the following steps: Step 1: Train the deep learning model using the offline training module to obtain the weight parameters; Step 2: Configure the weight parameters in the forward inference module and use the forward inference module to perform initial digital predistortion processing; Step 3: Use the loss calculation module to evaluate the difference between the PA output signal after initial digital predistortion processing and the ideal reference signal to obtain the real-time loss value; Step 4: Calculate the weight gradient of the real-time loss value using the backpropagation module, backpropagate the weight gradient, update the weight parameters of the forward inference module, and realize the real-time adaptive adjustment of the digital predistorter.

[0010] Furthermore, in one embodiment of the present invention, the training of the deep learning model based on the offline training module in step 1 specifically includes: The offline training module uses a deep learning model that combines LSTM and FNN modules. The LSTM module is used for time series modeling, and the FNN module is used to capture static nonlinear features to obtain weight parameters.

[0011] Furthermore, in one embodiment of the present invention, the forward inference module in step 2 includes an LSTM module, which is expanded into multiple cascaded gated unit computation modules according to time steps. Each gated unit computation module shares control logic and schedules historical state information through on-chip block memory to maintain the temporal continuity of the LSTM layer.

[0012] Furthermore, in one embodiment of the present invention, the forward inference module further includes an FNN module. The FNN module is a parallel structure in which multiple digital signal processing units perform weight matrix multiplication operations in parallel to obtain a digital predistortion signal. The digital predistortion signal is then processed by a PA to obtain a PA output signal after digital predistortion processing.

[0013] Furthermore, in one embodiment of the present invention, step 4, which involves using a backpropagation module to calculate the weight gradient of the real-time loss value and performing backpropagation on the weight gradient, specifically involves: Error initialization is performed based on real-time loss values. Gradients are backpropagated to the FNN module. Historical data from the LSTM module in the forward propagation module are cached in the on-chip block memory. Gradients are backpropagated to the LSTM module to obtain weight parameters for updating the forward inference module.

[0014] Furthermore, in one embodiment of the present invention, the error initialization based on the real-time loss value specifically includes: ; in, The output layer error of the FNN module. For loss function, The PA output signal after digital predistortion processing. This is a real label.

[0015] Furthermore, in one embodiment of the present invention, the backpropagation of gradients for the FNN module specifically includes: ; in, For the hidden layer error of the FNN module, The output layer weights of the FNN module, The gradients of the output layer error and the hidden layer error of the FNN module are respectively: (The weighted sum of the inputs to the hidden layer of the FNN module is used.) ; ; in, These are the hidden layer weights of the FNN module. This refers to the hidden layer state of the FNN module. This represents the final hidden state of the LSTM module.

[0016] Furthermore, in one embodiment of the present invention, the backpropagation of gradients from the LSTM module specifically includes: Obtain the error terms of multiple gated variables of the LSTM module, as follows: ; ; ; ; ; in, For the current time step, In hidden state, This is the error term for the current time step. For the output gate error term, This is the unit state error term. For the input gate error term, Forget gate error term, For candidate unit error terms, for The derivative of It is the hyperbolic tangent function. This represents the memory state at the current time step. For output gate, The candidate cell state is... For input gate, This refers to the memory state of the previous time step. For the Gate of Oblivion For the forgetting gate of the next time step, This represents the error term for a single time step within a given unit state. Based on the error terms of multiple gated variables in the LSTM module, their corresponding gradients are obtained as follows: ; ; ; ; in, For the output gate weights, For the input gate weights, For the weight of the forget gate, For candidate memory cell weights, This is the input for all time points in the current sequence.

[0017] Furthermore, in one embodiment of the present invention, the backpropagation module employs a stochastic gradient descent optimizer. After receiving B sample data, it performs gradient averaging and updates the weight parameters, where B is the set number of sample data received.

[0018] The present invention discloses a deep learning-based adaptive digital predistorter design system. The system, constructed based on the aforementioned method, includes the following modules: The offline training module is used to train the deep learning model and obtain the weight parameters. The forward inference module configures the weight parameters and uses the forward inference module to perform initial digital predistortion processing. The loss calculation module evaluates the difference between the PA output signal after initial digital predistortion processing and the ideal reference signal to obtain the real-time loss value. The backpropagation module calculates the weight gradient of the real-time loss value, performs backpropagation on the weight gradient, updates the weight parameters of the forward inference module, and realizes the real-time adaptive adjustment of the digital predistorter.

[0019] This invention solves the technical problem of existing technologies where, due to fixed model parameters, the lack of online adaptive capability in the face of dynamic changes in practical applications leads to a significant reduction in compensation effectiveness and an inability to meet practical linearization requirements. Specific beneficial effects include: 1. This invention proposes a deep learning-based adaptive digital predistorter design method, which implements a complete backpropagation process on an FPGA platform. Through deep learning modeling, a real-time backpropagation algorithm is integrated, including core steps such as offline training, forward inference, loss calculation, and backpropagation. The loss calculation module can receive the PA output signal after real digital predistortion processing in real time, and form a loss function by comparing it with the expected output signal. This function drives the backpropagation module to adjust parameters and achieve online closed-loop training. This invention ensures that the model parameters are continuously optimized according to the dynamic changes of the hardware operating status, and has continuous learning and adaptive capabilities. 2. This invention proposes a deep learning-based adaptive digital predistorter design method. It introduces the SGD (stochastic gradient descent) optimization mechanism into the DPD model and adopts batch gradient accumulation technology to trigger a parameter update after a certain amount of data accumulation. This can reduce the hardware burden while ensuring model convergence. It realizes the joint execution of model inference and parameter update process on FPGA platform. By using the power amplifier output signal as feedback information to dynamically adjust the model parameters, the predistortion model can quickly adapt to environmental changes, thereby significantly improving the compensation accuracy and engineering feasibility of DPD system in variable application scenarios. Compared to traditional static DPD models, this invention has the ability of "online self-adaptation and continuous optimization", which significantly improves the system's robustness to temperature drift, frequency offset and nonlinear drift, and has good system versatility and engineering application value. Attached Figure Description

[0020] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is the power spectral density evaluation diagram of the offline-trained deep learning digital predistorter described in Implementation Method 2; Figure 2 This is a flowchart of the LSTM module calculation process described in Implementation Method 3; Figure 3 This is the inference and closed-loop training architecture diagram described in Implementation Method 4; Figure 4 This is a loss change graph during the closed-loop training process described in Implementation Method 4. Detailed Implementation

[0021] Various embodiments of the present invention will now be clearly and completely described with reference to the accompanying drawings. The embodiments described with reference to the drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0022] Implementation Method 1. A deep learning-based adaptive digital predistorter design method, wherein the adaptive digital predistorter includes an offline training module, a forward inference module, a loss calculation module, and a backpropagation module, and the adaptive digital predistorter design method includes the following steps: Step 1: Train the deep learning model using the offline training module to obtain the weight parameters; Step 2: Configure the weight parameters in the forward inference module and use the forward inference module to perform initial digital predistortion processing; Step 3: Use the loss calculation module to evaluate the difference between the PA output signal after initial digital predistortion processing and the ideal reference signal to obtain the real-time loss value; Step 4: Calculate the weight gradient of the real-time loss value using the backpropagation module, backpropagate the weight gradient, update the weight parameters of the forward inference module, and realize the real-time adaptive adjustment of the digital predistorter.

[0023] Most existing deep learning-based DPD methods rely on static modeling techniques and typically complete the fitting of power amplifier characteristics through offline training. In practical applications, these methods are difficult to adapt to changes in the nonlinear characteristics of power amplifiers caused by factors such as changes in ambient temperature, operating frequency drift, and device aging. This leads to a significant decrease in predistortion compensation performance, affecting the linearization effect of the system and communication quality.

[0024] To address the aforementioned technical issues, this embodiment provides a deep learning-based adaptive digital predistorter design method, wherein the adaptive digital predistorter includes an offline training module, a forward inference module, a loss calculation module, and a backpropagation module.

[0025] In this embodiment, the offline training module uses the Python language environment and the PyTorch deep learning framework to build a deep learning digital predistorter model structure, and utilizes the RTX 4090D graphics processing unit (GPU) produced by NVIDIA to accelerate the training process of the deep learning model and obtain fixed model parameters. The forward inference module, loss calculation module, and backpropagation module are implemented in modular hardware using the Vitis HLS high-level synthesis platform provided by Xilinx, and deployed on the Xilinx XCVU11P-FLGA2577-1-E FPGA development board. The comprehensive results show that the LUT (Logical Look-Up Table) resource usage is 720,000, accounting for approximately 55%, the BRAM (Block Memory On-Chip) resource usage is approximately 84%, and the DSP (Digital Signal Processing Unit) resource usage is less than 40%, which meets the actual deployment requirements.

[0026] In this embodiment, the loss calculation module uses RMSE (root mean square error) as the loss function to evaluate the difference between the received digital predistortion processed PA output signal and the ideal reference signal, and calculates the real-time loss value as the basis for model adaptive optimization.

[0027] The RMSE loss function is used to measure the actual output signal. With the target output signal The error between them is calculated using the following formula: ; Where N is the number of samples (i.e., the number of sampling points) participating in each update, and this error is passed into the backpropagation module as a loss metric.

[0028] Therefore, this implementation first trains the deep learning model offline on a high-performance computing platform to obtain the weight parameters of the predistortion model. The trained model weights are then deployed to an FPGA hardware platform, where inference calculations are performed, achieving the traditional "static" form of digital predistortion functionality. Subsequently, the PA output signal after actual digital predistortion processing is obtained through the feedback path of the RF link. The loss calculation module calculates the loss value based on the error between this feedback signal and the desired output signal. This loss result is further fed into the parameter backpropagation module to drive model parameter updates, thereby achieving online adaptive adjustment and closed-loop compensation of the model. This solves the technical problem in existing technologies where fixed model parameters lead to a lack of online adaptive capability in the face of dynamic changes in practical applications, resulting in a significant reduction in compensation effect and an inability to meet actual linearization requirements.

[0029] Implementation Method 2. This implementation method further defines the deep learning-based adaptive digital predistorter design method described in Implementation Method 1. Specifically, in step 1, the deep learning model is trained using an offline training module: The offline training module uses a deep learning model that combines LSTM and FNN modules. The LSTM module is used for time series modeling, and the FNN module is used to capture static nonlinear features to obtain weight parameters.

[0030] Traditional DPD design methods often employ polynomial models, such as memory polynomials and generalized memory polynomials. These methods offer advantages such as high modeling efficiency and ease of implementation when dealing with static or weakly dynamic nonlinear distortions. However, polynomial models suffer from insufficient modeling capabilities and poor parameter generalization when facing complex modulation schemes, high-bandwidth signals, or highly dynamic changing scenarios, making it difficult to meet the increasingly stringent linearity requirements of modern communication systems.

[0031] To address the aforementioned issues, the offline training module described in this embodiment employs a hybrid network structure combining an LSTM (Long Short-Term Memory) module and a FNN (Feedforward Neural Network) module as the core of DPD modeling. LSTM can preserve the historical state of the input in time series modeling, making it suitable for compensating for the dynamic memory effect present in PA (Power Spectral Density). FNN is responsible for capturing static nonlinear features; the combination of the two significantly improves the model's compensation accuracy. Therefore, this structure possesses both the ability to model time series features and the ability to capture the nonlinear mapping relationship of PA. The model input is the historical sampling sequence of the baseband signal, and the output is the compensated excitation signal. By training this network on a GPU platform, model performance can be quickly optimized to obtain optimal parameters that meet performance requirements such as PSD (Power Spectral Density) and ACPR (Adjacent Channel Power Ratio).

[0032] Adjacent channel power ratio (ACPR) is a key indicator for evaluating the spectral leakage characteristics of a power amplifier's output signal. It reflects the degree of interference from spectral spread caused by nonlinear distortion to adjacent channels. A lower ACPR value indicates better suppression of nonlinear distortion and higher spectral purity of the signal. Table 1 shows the comparison results of the DPD model under the ACPR values ​​of the upper and lower adjacent channels. The results show that without pre-distortion processing, the system's spectral leakage is quite severe, with ACPR values ​​of -22.43 dB for the lower sideband and -22.52 dB for the upper sideband. After introducing the DPD model, the ACPR values ​​are significantly improved, increasing to -45.16 dB and -45.02 dB, respectively.

[0033] Table 1

[0034] Power spectral density measures the power distribution in the frequency domain of a signal and is an important indicator for evaluating the effectiveness of digital predistorters in suppressing nonlinear distortion in power amplifiers. For example... Figure 1 As shown, a comparison of the spectral characteristics with and without a predistorter is presented. From Figure 1 As can be observed, the signal without predistortion compensation exhibits significant spectral spread on both sides of the main carrier frequency, manifested as high adjacent channel leakage power, which is a typical characteristic of power amplifier nonlinear distortion. After introducing a predistorter, the spectral spread phenomenon is suppressed to a certain extent, and the adjacent channel interference level is reduced.

[0035] Implementation Method 3. This implementation method further defines the deep learning-based adaptive digital predistorter design method described in Implementation Method 1. In step 2, the forward inference module includes an LSTM module, which is expanded into multiple cascaded gated unit computation modules according to time steps. Each gated unit computation module shares control logic and schedules historical state information through on-chip block memory to maintain the temporal continuity of the LSTM layer. It also includes an FNN module, which is a parallel structure. Multiple digital signal processing units perform weight matrix multiplication operations in parallel to obtain a digital predistortion signal. The digital predistortion signal is processed by the PA to obtain the PA output signal after digital predistortion processing.

[0036] In this embodiment, the forward inference module adopts a modular design in the FPGA, such as... Figure 2 As shown, the computation process of the LSTM layer and the feedforward neural network is structurally decomposed and mapped into computational units that can be efficiently executed in FPGA. By using the weight parameters obtained from offline training, each sub-module is precisely configured, thereby completing efficient model forward inference and digital predistortion processing on FPGA hardware.

[0037] The LSTM module first expands the LSTM layer into multiple cascaded gated unit computation modules by time step, including input gates. Forgotten Gate Output gate and candidate cell status Each gate unit shares control logic, and historical state information is scheduled through BRAM to maintain the temporal continuity of the LSTM state. The feedforward neural network part is designed as a parallel structure, with the weight matrix multiplication operation of each layer completed in parallel by multiple DSP units, and inter-layer transfer is achieved through pipelined method to ensure a high-throughput model inference process.

[0038] In this embodiment, the calculation process of LSTM in the forward inference module is as follows: Let the current time step input be The hidden state in the previous time step was The memory state of the previous time step is Then, the intermediate vector generated by the weighted connection between the input and the hidden state is... The calculation formula is: ; in, For the current time step, For the previous time step, For input, In hidden state, For weight parameters, This is the bias parameter.

[0039] Will It is divided into four parts, and the multiple current time step gating variables are the input gates. Forgotten Gate Candidate cell status and output gate : ; ; ; ; in, For the Sigmoid function, It is the hyperbolic tangent function.

[0040] Then calculate the memory state at the current time step. and the hidden state of the current time step : ; ; in, This is element-wise multiplication.

[0041] To support the implementation of the subsequent backpropagation module, this implementation maintains a cache structure at each time step of the forward inference, caching the input gates of the gating variables from the current time step and the previous time step. Forgotten Gate Output gate and candidate cell status and memory state and hidden state It also caches the data from the previous time step. This mechanism effectively improves data accessibility and reduces off-chip access pressure.

[0042] Implementation Method 4. This implementation method further defines the deep learning-based adaptive digital predistorter design method described in Implementation Method 1. In step 4, the weight gradient is calculated using the backpropagation module to obtain the real-time loss value, and the weight gradient is backpropagated. Specifically: Error initialization is performed based on real-time loss values. Gradients are backpropagated to the FNN module. Historical data from the LSTM module in the forward propagation module are cached in the on-chip block memory. Gradients are backpropagated to the LSTM module to obtain weight parameters for updating the forward inference module.

[0043] Furthermore, this embodiment proposes a backpropagation computation method based on a combination of LSTM and FNN modules suitable for FPGA platforms. Addressing the online training requirements of deep learning models in resource-constrained scenarios, the gradient backpropagation process of the LSTM and FNN modules is divided into three main computational stages. Based on a modular design approach, efficient parallel computation and resource reuse are achieved on the FPGA.

[0044] The three stages include: forward activation caching and error initialization stage, FNN module gradient backpropagation and parameter update stage, and LSTM module gradient backpropagation stage. The mathematical models and calculation processes of each stage are described below.

[0045] 1. Activation of Cache and Error Initialization Phase: After the forward propagation phase is completed, the system will input the current time step. The memory state at the current time step The hidden state of the current time step The gate vectors are cached in BRAM for dependency calculation during backpropagation. Meanwhile, if the network endpoint is the output of an FNN module, its output error can be expressed as: ; in, The output layer error of the FNN module. For loss function, The PA output signal after digital predistortion processing. This is a real label.

[0046] 2. FNN Module Gradient Backpropagation and Parameter Update Stage: The FNN module is typically located after the LSTM module output and is used for feature classification or numerical regression. Its backpropagation process can be modeled as a standard MLP (Multilayer Perceptron) backpropagation structure. The input is defined as the final hidden state of the LSTM module. Then the hidden layer error is: ; in, For the hidden layer error of the FNN module, The output layer weights of the FNN module, The gradients of the output layer error and the hidden layer error of the FNN module are respectively: (The weighted sum of the inputs to the hidden layer of the FNN module is used.) ; ; in, These are the hidden layer weights of the FNN module. This represents the hidden layer state of the FNN module.

[0047] 3. LSTM Module Gradient Backpropagation Stage: The backpropagation of the LSTM module needs to proceed step by step from... Traverse backwards up to t=1. Let the error term at the current time step be... Error propagation in an LSTM module includes the hidden state at the current time step. Memory state at the current time step The partial derivatives are propagated. According to the chain rule, the gradients of the parameters of each gate vector in the LSTM module are calculated as follows: Obtain the error terms of multiple gated variables of the LSTM module, as follows: ; ; ; ; ; in, For the output gate error term, This is the unit state error term. For the input gate error term, Forget gate error term, For candidate unit error terms, for The derivative of The error term for a time step in a given unit state; Based on the error terms of multiple gated variables in the LSTM module, their corresponding gradients are obtained as follows: ; ; ; ; in, For the output gate weights, For the input gate weights, For the weight of the forget gate, The weights are those of the candidate memory units.

[0048] In this embodiment, the backpropagation module uses a stochastic gradient descent optimizer. After receiving B sample data, it performs gradient averaging and updates the weight parameters, where B is the set number of sample data received.

[0049] The backpropagation module employs a stochastic gradient descent optimizer to update the parameters of each layer. Let the weight parameters be... The learning rate is The gradient is The parameter update rule is as follows: ; Where, gradient The calculation is obtained by passing the partial derivative of the error with respect to the current layer output to the corresponding input, and then combining it with the historical gating state. The calculation process is implemented in the FPGA through the reverse data flow, which activates the reverse gradient propagation path in sequence.

[0050] The backpropagation module implements error backpropagation and stochastic gradient descent algorithms on the FPGA platform. Based on the loss value output by the loss calculation module and the gradient calculated by the parameter weights, the backpropagation module updates the weight parameters in the deep learning model online, thereby constructing a deep learning digital predistortion system with real-time closed-loop training capability.

[0051] Furthermore, to control FPGA resource consumption and computational latency, a batch accumulation mechanism is integrated into the SGD optimizer. Specifically, after receiving B sample data, a gradient averaging operation is performed and an update is triggered, effectively reducing the update frequency and improving hardware utilization efficiency. Simultaneously, it prevents model weights from getting trapped in local optima during frequent updates, thus avoiding impacting the compensation effect. This batch processing mechanism is cached in BRAM, and a control state machine determines whether the data volume has reached a threshold, activating the gradient accumulation module and the parameter update module accordingly.

[0052] Hardware structure division as follows Figure 3 As shown, the system includes an input buffer, a model weight storage area, a forward inference module, an output buffer, a loss calculation module, a gradient calculation module, and a weight update process. The input buffer receives data streams from the ADC or a host computer. The forward inference module performs inference for the LSTM and FNN modules. The loss calculation module executes the RMSE loss function to calculate the loss. The gradient calculation module calculates the gradient based on the loss and updates the weights. Finally, the weight update process corrects the weights based on the output gradient and the SGD algorithm.

[0053] Furthermore, to verify the effectiveness of the online adaptive mechanism of this invention, continuous testing was conducted after deployment. Under stable conditions, the model's loss test curve is shown below. Figure 4 As shown, the loss function decreases steadily, enabling closed-loop training and stable convergence.

[0054] In summary, the backpropagation method for FPGA-based LSTM+FNN network structure proposed in this invention, through three-stage task partitioning and pipeline optimization, not only significantly improves the hardware parallelism and computational efficiency of the training process, but also provides a feasible path for online training on resource-constrained platforms, possessing high engineering practical value and promising prospects for promotion.

[0055] Implementation Method 5. This implementation method describes a deep learning-based adaptive digital predistorter design system. The system is constructed based on the method described in Implementation Method 1 and includes the following modules: The offline training module is used to train the deep learning model and obtain the weight parameters. The forward inference module configures the weight parameters and uses the forward inference module to perform initial digital predistortion processing. The loss calculation module evaluates the difference between the PA output signal after initial digital predistortion processing and the ideal reference signal to obtain the real-time loss value. The backpropagation module calculates the weight gradient of the real-time loss value, performs backpropagation on the weight gradient, updates the weight parameters of the forward inference module, and realizes the real-time adaptive adjustment of the digital predistorter.

[0056] The above provides a detailed description of the design method and system for an adaptive digital predistorter based on deep learning proposed in this invention. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A deep learning based adaptive digital pre-distorter design method, characterized in that, The adaptive digital pre-distorter comprises an offline training module, a forward reasoning module, a loss calculation module and a back propagation module. Step 1, training the deep learning model based on the offline training module to obtain weight parameters; Step 2, configuring the weight parameters to the forward reasoning module and performing initial digital pre-distortion processing by using the forward reasoning module; Step 3, performing difference evaluation on the PA output signal after initial digital pre-distortion processing and the ideal reference signal by using the loss calculation module to obtain a real-time loss value; Step 4, calculating the weight gradient for the real-time loss value by using the back propagation module, performing back propagation on the weight gradient, updating the weight parameters of the forward reasoning module, and realizing real-time adaptive adjustment of the digital pre-distorter.

2. The method of claim 1, wherein, In step 1, the deep learning model is trained based on the offline training module, specifically: The offline training module adopts a deep learning model combining an LSTM module and an FNN module, uses the LSTM module to model time series, and captures static nonlinear features through the FNN module to obtain weight parameters.

3. The method of claim 1, wherein, The forward reasoning module in step 2 comprises an LSTM module, which is expanded into multiple serial gate unit calculation modules according to time steps, shares control logic between each gate unit calculation module, and maintains the time continuity of the LSTM layer by scheduling historical state information through an on-chip block memory.

4. The method of claim 1 or 3, wherein, The forward reasoning module further comprises an FNN module, which is a parallel structure and comprises multiple digital signal processing units that perform weight matrix multiplication operations in parallel to obtain a digital pre-distortion signal, process the digital pre-distortion signal by using a PA, and obtain a PA output signal after digital pre-distortion processing.

5. The method of claim 1, wherein, In step 4, the weight gradient is calculated for the real-time loss value by using the back propagation module, and the weight gradient is back propagated, specifically: Error initialization is performed based on the real-time loss value, the gradient of the FNN module is back propagated, historical data of the LSTM module in the forward transmission module is cached through the on-chip block memory, the gradient of the LSTM module is back propagated, and the weight parameters for updating the forward reasoning module are obtained.

6. The method of claim 5, wherein the method further comprises: The error initialization based on the real-time loss value is specifically: ； wherein, is the FNN module output layer error, is the loss function, is the PA output signal after digital pre-distortion processing, is the true label.

7. The method of claim 5, wherein the method further comprises: The gradient back propagation of the FNN module is specifically: ； where, is the FNN module hidden layer error, is the FNN module output layer weight, is the FNN module hidden layer input weighted sum, the corresponding gradients of the FNN module output layer error and hidden layer error are respectively: ； ； wherein, is the FNN module hidden layer weight, is the FNN module hidden layer state, is the LSTM module final hidden state.

8. The method of claim 5, wherein the method further comprises: The gradient back propagation of the LSTM module is specifically: The error terms of multiple gate variables of the LSTM module are obtained, which are respectively: ； ； ； ； ； in, For the current time step, In hidden state, This is the error term for the current time step. For the output gate error term, This is the unit state error term. For the input gate error term, Forget gate error term, For candidate unit error terms, for The derivative, It is the hyperbolic tangent function. This represents the memory state at the current time step. For output gate, The candidate cell state is... For input gate, This refers to the memory state of the previous time step. For the Gate of Oblivion For the forgetting gate of the next time step, This represents the error term for a single time step within a given unit state. Based on the error terms of multiple gate variables of the LSTM module, the corresponding gradients are obtained, which are respectively: ； ； ； ； wherein, is an output gate weight, is an input gate weight, is a forget gate weight, is a candidate memory cell weight, is an input for all time steps of the current sequence.

9. The method of claim 1 or 5, wherein, The back propagation module adopts a stochastic gradient descent optimizer, and performs gradient averaging and updates the weight parameters once every B sample data are received, wherein B is the set number of received sample data.

10. A deep learning based adaptive digital pre-distorter design system, said system is constructed based on the method of claim 1, characterized in that, The following modules are included: An offline training module, which trains a deep learning model based on the offline training module to obtain weight parameters; A forward reasoning module, which configures the weight parameters to the forward reasoning module and performs initial digital pre-distortion processing by using the forward reasoning module; The loss calculation module is configured to evaluate the difference between the PA output signal after the initial digital pre-distortion processing and the ideal reference signal to obtain a real-time loss value. The back propagation module is configured to calculate a weight gradient based on the real-time loss value, perform back propagation on the weight gradient, update the weight parameter of the forward reasoning module, and realize real-time adaptive adjustment of the digital pre-distorter.

Images