Continuous learning digital predistortion system and method for intelligent scenarios

Abstract

The present invention discloses a continuous learning digital predistortion system and method for intelligent scenes, including a configuration adjustment module, a continuous learning digital predistortion module, a linearity check module, a training data storage module, a model coefficient extraction module and an objective function module; the configuration adjustment module changes the input signal configuration in real time, the input signal obtains a predistortion signal through the continuous learning digital predistortion module, the predistortion signal is transmitted to the power amplifier to obtain an output signal, the feedback loop collects the output signal, and enters the linearity check module to check the linearity. The present invention can greatly reduce the model coefficient update requirements of the digital predistortion technology in the intelligent scene, thereby reducing the application cost of the digital predistortion technology.

Description

Technical Field

[0001] The present invention belongs to the technical field of digital pre-distortion, and in particular relates to a continuous learning digital pre-distortion system and method for intelligent scenarios. Background Art

[0002] Future intelligent communication systems can adjust transmitter configurations in real time based on user needs and changes in the radio environment to make more flexible use of communication resources. The dynamic changes in these configurations will directly affect the working state of the power amplifier in the transmitter. Like traditional transmitters, power amplifiers usually work in the nonlinear region to ensure high efficiency. The resulting nonlinear characteristics will cause signal transmission distortion, which in turn affects normal communication.

[0003] Digital pre-distortion technology has been widely used to compensate for the nonlinearity of power amplifiers. However, traditional digital pre-distortion technology usually assumes that the power amplifier works in a relatively stable state, while the working state of the power amplifier in the smart scene changes rapidly. If the coefficients of the digital pre-distorter are updated every time the state changes, it will be very inefficient. Therefore, the traditional digital pre-distortion system is not suitable for deployment in future smart communication systems, and it is necessary to develop a digital pre-distortion system that can efficiently cope with smart scenes. Summary of the invention

[0004] The present invention aims to provide a continuous learning digital pre-distortion system and method for intelligent scenarios, so as to solve the technical problem that traditional digital pre-distortion technology is very inefficient.

[0005] In order to solve the above technical problems, the specific technical solutions of the present invention are as follows:

[0006] A continuous learning digital predistortion system for intelligent scenarios comprises a configuration adjustment module, a continuous learning digital predistortion module, a linearity check module, a training data storage module, a model coefficient extraction module and an objective function module; the configuration adjustment module changes the input signal configuration in real time according to user requirements, the input signal obtains a predistortion signal through the continuous learning digital predistortion module, the predistortion signal is transmitted to a power amplifier to obtain an output signal, a feedback loop collects the output signal, and enters the linearity check module to check the linearity, if the linearity meets the requirements, the model can be maintained in its original state without being updated, if the linearity does not meet the requirements, the model updates the coefficients through continuous learning, uses the data in the training data storage module, and adjusts the weights of the model through the model coefficient extraction module according to the objective function calculated by the objective function module.

[0007] Furthermore, the configuration adjustment module adjusts the input signal configuration in real time according to user needs, and the input signal configuration includes: signal power, signal bandwidth and signal type; wherein the signal power is adjusted according to the transmission distance, the signal bandwidth is adjusted according to the spectrum utilization, and the signal type is adjusted according to the signal transmission protocol.

[0008] Furthermore, the continuous learning digital pre-distortion module performs continuous learning through the following process: storing all input and output signals that do not meet the linearity requirements under past configurations, and using both the input and output signals of past configurations and the input and output signals of current configurations when updating model coefficients.

[0009] Furthermore, the linearity checking module calculates the linearity index of the collected output signal, and determines whether the linearity meets the requirements according to the index, and the linearity index includes: normalized mean square error and adjacent channel leakage ratio.

[0010] Furthermore, the model coefficient extraction module is equipped with a digital pre-distortion model, which is mainly based on a neural network and is embedded with a static polynomial module composed of a polynomial model.

[0011] Furthermore, the training data storage module provides training data for continuous learning of the digital predistortion model, and the stored data includes: new input and output data of the power amplifier under the current configuration and old input and output data of the power amplifier stored when all linearity requirements were not met in the past.

[0012] Furthermore, the input layer vector of the neural network includes a signal input and a configuration input, the signal input is a baseband transmission signal input, the configuration input includes a signal power, a signal bandwidth and a signal type input, the hidden layer uses a fully connected layer and a hyperbolic tangent activation function, the static polynomial module is constructed using a memory polynomial model, and the output of the digital predistortion model is expressed as follows:

[0013]

[0014]

[0015] In the formula, u(n) is the output predistortion signal of the digital predistortion model at time n, u I (n) is the real part of the predistortion signal, u Q (n) is the imaginary part of the predistortion signal, i is the imaginary unit; h i (n) is the value of the i-th input neuron of the neural network, I is the number of neurons in the input layer of the neural network, and w 1ij is the connection weight between the i-th input neuron and the j-th output neuron in the first fully connected layer of the neural network, b 1jis the bias of the jth output neuron in the first layer, f(·) is the hyperbolic tangent activation function, J is the number of output neurons in the first fully connected layer of the neural network, and w 2jl is the connection weight between the jth input neuron and the lth output neuron in the second fully connected layer of the neural network, b 2l is the bias of the lth output neuron in the second layer, L is the number of output neurons in the second fully connected layer of the neural network, w 3l1 is the connection weight between the lth input neuron and the real output neuron in the third fully connected layer of the neural network, b 31 is the bias of the real output neuron in the third layer, w 3l2 is the connection weight between the lth input neuron and the imaginary output neuron in the third fully connected layer of the neural network, b 32 is the bias of the imaginary output neuron in the third layer, x(nq)|x(nq)| k-1 is the basis function of the polynomial model, K is the nonlinear order of the model, Q is the memory depth of the model, Re(·) means taking the real part, Im(·) means taking the imaginary part, w 1kq is the connection weight between the real part of the polynomial basis function and the real part of the output neuron, w 2kq is the connection weight of the imaginary and real output neurons of the polynomial basis function, w 3kq is the connection weight of the real and imaginary output neurons of the polynomial basis function, w 4kq is the connection weight between the imaginary part and the imaginary part output neurons of the polynomial basis function.

[0016] Furthermore, the objective function module calculates a joint objective function, and the model coefficient extraction module updates the model coefficients by minimizing the joint objective function. The joint objective function is expressed as the following expression:

[0017]

[0018] In the formula, J joint (θ) is the joint objective function, J1(θ) is the first objective function, J2(θ) is the second objective function, θ is all the coefficients in the model, including weight w and bias b, α is the weighting factor, is the target output signal of the digital predistortion model.

[0019] The present invention also provides a continuous learning digital predistortion method for intelligent scenarios, comprising the following steps:

[0020] Step S1, the configuration adjustment module intelligently adjusts the configuration of the input signal, and the newly configured input signal generates a predistortion signal through the continuous learning digital predistortion module;

[0021] Step S2, the predistortion signal generates an output signal through a power amplifier and is sent to a linearity inspection module;

[0022] Step S3: The linearity checking module checks the linearity of the output signal. If the linearity meets the requirement, the digital pre-distortion model does not need to be updated, and the process proceeds to step S7; if the linearity does not meet the requirement, the process proceeds to step S4;

[0023] Step S4, storing the predistortion signal and the output signal under the new configuration into a training data storage module;

[0024] Step S5, using the data in the training data storage module and according to the objective function of the objective function module, the model coefficient extraction module optimizes the model and updates the model coefficient;

[0025] Step S6, copying the updated model coefficients to the continuous learning digital pre-distortion module;

[0026] Step S7, continuous learning The digital pre-distortion system operates normally. When the configuration adjustment module is ready to adjust the input signal to a new configuration, step S1 is entered again.

[0027] The intelligent scene-oriented continuous learning digital predistortion system and method of the present invention have the following advantages:

[0028] 1. In the present invention, each update of the model uses both the input and output signals of the past configuration and the input and output signals of the new configuration, so that the model will not forget the information of the past configuration when learning the digital pre-distortion of the new configuration, thereby realizing the function of continuous learning, and enabling the model to have the ability to become more and more complete through continuous learning, greatly reducing the need for updating the model coefficients in the intelligent scene.

[0029] 2. In the present invention, a polynomial model is embedded in the neural network model. The polynomial model fixes part of the nonlinear operation by using the polynomial kernel function, and can fit part of the general nonlinearity, thereby reducing the operation requirements of the entire model and improving the convergence speed.

[0030] 3. The present invention inherits the advantages of the traditional neural network digital pre-distortion model and has high flexibility. Especially in such a dynamic scenario, the neural network can more accurately offset the dynamic nonlinearity, thereby improving the linearization performance.

[0031] 4. In the present invention, the joint objective function restricts the weights of the neural network, which greatly improves the generalization performance of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a block diagram of a continuous learning digital pre-distortion system for intelligent scenarios in an embodiment of the present invention;

[0033] Figure 2 is a graph showing the linearization performance of digital predistortion according to an embodiment of the present invention as a function of power amplifier configuration;

[0034] Figure 3 is a demodulation constellation diagram of digital predistortion at configuration index 35 in an embodiment of the present invention;

[0035] Figure 4 FIG. 3 is a power spectrum diagram of digital predistortion at configuration index 35 in an embodiment of the present invention. DETAILED DESCRIPTION

[0036] In order to better understand the purpose, structure and function of the present invention, the following further describes in detail a continuous learning digital pre-distortion system and method for intelligent scenarios of the present invention in conjunction with the accompanying drawings.

[0037] The present invention provides a continuous learning digital predistortion system for intelligent scenarios, such as Figure 1 As shown, it includes a configuration adjustment module, a continuous learning digital pre-distortion module, a linearity check module, a training data storage module, a model coefficient extraction module and an objective function module; the configuration adjustment module changes the input signal configuration in real time according to user needs, the input signal obtains a pre-distortion signal through the continuous learning digital pre-distortion module, the pre-distortion signal is transmitted to the power amplifier to obtain an output signal, the feedback loop collects the output signal, and enters the linearity check module to check the linearity. If the linearity meets the requirements, the model can be maintained as it is and does not need to be updated. If the linearity does not meet the requirements, the model updates the coefficients through continuous learning, uses the data in the training data storage module, and adjusts the weights of the model through the model coefficient extraction module according to the objective function calculated by the objective function module.

[0038] The configuration adjustment module adjusts the input signal configuration in real time according to user needs. The input signal configuration includes: signal power, signal bandwidth and signal type. The signal power is adjusted according to the transmission distance, the signal bandwidth is adjusted according to the spectrum utilization, and the signal type is adjusted according to the signal transmission protocol.

[0039] The continuous learning digital pre-distortion module performs continuous learning through the following process: storing all input and output signals that do not meet the linearity requirements under past configurations, and using both the input and output signals of the past configurations and the input and output signals of the current configuration when updating the model coefficients.

[0040] The linearity check module calculates the linearity index of the collected output signal and determines whether the linearity meets the requirements based on the index. The linearity index includes: normalized mean square error and adjacent channel leakage ratio.

[0041] The training data storage module provides training data for the continuous learning of the digital pre-distortion model. The stored data includes: new input and output data of the power amplifier under the current configuration and old input and output data of the power amplifier stored when all linearity requirements were not met in the past.

[0042] The model coefficient extraction module is equipped with a digital pre-distortion model. The digital pre-distortion model is mainly based on a neural network and is embedded with a static polynomial module composed of a polynomial model.

[0043] Specifically, the input layer vector of the neural network includes signal input and configuration input. The signal input is the baseband transmission signal input. The configuration input includes signal power, signal bandwidth and signal type input. The hidden layer uses a fully connected layer and a hyperbolic tangent activation function. The static polynomial module is constructed using a memory polynomial model. The output of the digital predistortion model can be expressed as follows:

[0044]

[0045] In the formula, u(n) is the output predistortion signal of the digital predistortion model at time n, and u I (n) and u Q (n) are the real and imaginary parts of the predistorted signal, respectively, and i is the imaginary unit. i (n) is the value of the i-th input neuron of the neural network, I is the number of neurons in the input layer of the neural network, and w 1ij is the connection weight between the i-th input neuron and the j-th output neuron in the first fully connected layer of the neural network, b 1j is the bias of the jth output neuron in the first layer, f(·) is the hyperbolic tangent activation function, J is the number of output neurons in the first fully connected layer of the neural network, and w 2jl is the connection weight between the jth input neuron and the lth output neuron in the second fully connected layer of the neural network, b 2l is the bias of the lth output neuron in the second layer, L is the number of output neurons in the second fully connected layer of the neural network, w 3l1 is the connection weight between the lth input neuron and the real output neuron in the third fully connected layer of the neural network, b 31 is the bias of the real output neuron in the third layer, w 3l2 is the connection weight between the lth input neuron and the imaginary output neuron in the third fully connected layer of the neural network, b 32 is the bias of the imaginary output neuron in the third layer, x(nq)|x(nq)| k-1 is the basis function of the polynomial model, K is the nonlinear order of the model, Q is the memory depth of the model, Re(·) means taking the real part, Im(·) means taking the imaginary part, w 1kqis the connection weight between the real part of the polynomial basis function and the real part of the output neuron, w 2kq is the connection weight of the imaginary and real output neurons of the polynomial basis function, w 3kq is the connection weight of the real and imaginary output neurons of the polynomial basis function, w 4kq is the connection weight between the imaginary part and the imaginary part output neurons of the polynomial basis function.

[0046] The objective function module calculates a joint objective function, and the model coefficient extraction module updates the model coefficients by minimizing the joint objective function. The joint objective function can be expressed as follows:

[0047]

[0048] In the formula, J joint (θ) is the joint objective function, J1(θ) is the first objective function, J2(θ) is the second objective function, θ is all the coefficients in the model, including weight w and bias b, α is the weighting factor, is the target output signal of the digital predistortion model.

[0049] This specific implementation also discloses a continuous learning digital predistortion method for intelligent scenarios, comprising the following steps:

[0050] Step S1, the configuration adjustment module intelligently adjusts the configuration of the input signal, and the newly configured input signal generates a predistortion signal through the continuous learning digital predistortion module;

[0051] Step S2, the predistortion signal generates an output signal through a power amplifier and is sent to a linearity inspection module;

[0052] Step S3: The linearity checking module checks the linearity of the output signal. If the linearity meets the requirement, the digital pre-distortion model does not need to be updated, and the process proceeds to step S7; if the linearity does not meet the requirement, the process proceeds to step S4;

[0053] Step S4, storing the predistortion signal and the output signal under the new configuration into a training data storage module;

[0054] Step S5, using the data in the training data storage module and according to the objective function of the objective function module, the model coefficient extraction module optimizes the model and updates the model coefficient;

[0055] Step S6, copying the updated model coefficients to the continuous learning digital pre-distortion module;

[0056] Step S7, continuous learning The digital pre-distortion system operates normally. When the configuration adjustment module is ready to adjust the input signal to a new configuration, step S1 is entered again.

[0057] The dynamic configuration in the intelligent communication scenario is simulated by input signals of different power, bandwidth and signal category. The continuous learning digital pre-distortion system proposed by the present invention is used to linearize the power amplifier in the intelligent scenario. The change of linearization performance during the continuous learning process is as follows: Figure 2 As shown in Figure 1, as the learning continues, the update frequency of the model becomes longer, indicating that the model is becoming more and more complete, thereby reducing the need for coefficient updates in the intelligent scene; the demodulation constellation diagram and power spectrum density diagram of configuration index 35 are shown in Figure 1. Figure 3 and Figure 4 As shown, there is an obvious improvement compared with the situation before digital pre-distortion is adopted, which proves the effectiveness of the present invention.

[0058] The matters not described in detail in the present invention are all known technologies to those skilled in the art.

[0059] It is to be understood that the present invention is described by some embodiments, and it is known to those skilled in the art that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the scope of protection of the present invention.

Claims

1. A continuous learning digital pre-distortion system for intelligent scenarios, characterized in that: It includes a configuration adjustment module, a continuous learning digital pre-distortion module, a linearity check module, a training data storage module, a model coefficient extraction module and an objective function module; the configuration adjustment module changes the input signal configuration in real time according to user needs, the input signal obtains a pre-distortion signal through the continuous learning digital pre-distortion module, the pre-distortion signal is transmitted to the power amplifier to obtain an output signal, the feedback loop collects the output signal, and enters the linearity check module to check the linearity. If the linearity meets the requirements, the model can be maintained in its original state and does not need to be updated. If the linearity does not meet the requirements, the model updates the coefficients through continuous learning, uses the data in the training data storage module, and adjusts the model weights through the model coefficient extraction module according to the objective function calculated by the objective function module; The model coefficient extraction module is equipped with a digital pre-distortion model, which is mainly based on a neural network and is embedded with a static polynomial module composed of a polynomial model; The training data storage module provides training data for continuous learning of the digital predistortion model, and the stored data includes: new input and output data of the power amplifier under the current configuration and old input and output data of the power amplifier stored when all linearity requirements were not met in the past; The input layer vector of the neural network includes signal input and configuration input. The signal input is a baseband transmission signal input. The configuration input includes signal power, signal bandwidth and signal type input. The hidden layer adopts a fully connected layer and a hyperbolic tangent activation function. The static polynomial module is constructed using a memory polynomial model. The output of the digital predistortion model is expressed as follows: In the formula, u(n) is the output predistortion signal of the digital predistortion model at time n, and u I (n) is the real part of the predistortion signal, u Q (n) is the imaginary part of the predistortion signal, i is the imaginary unit; h i (n) is the value of the i-th input neuron of the neural network, I is the number of neurons in the input layer of the neural network, and w 1ij is the connection weight between the i-th input neuron and the j-th output neuron in the first fully connected layer of the neural network, b 1j is the bias of the jth output neuron in the first layer, f(·) is the hyperbolic tangent activation function, J is the number of output neurons in the first fully connected layer of the neural network, and w 2jl is the connection weight between the jth input neuron and the lth output neuron in the second fully connected layer of the neural network, b 2l is the bias of the lth output neuron in the second layer, L is the number of output neurons in the second fully connected layer of the neural network, w 3l1 is the connection weight between the lth input neuron and the real output neuron in the third fully connected layer of the neural network, b 31 is the bias of the real output neuron in the third layer, w 3l2 is the connection weight between the lth input neuron and the imaginary output neuron in the third fully connected layer of the neural network, b 32 is the bias of the imaginary output neuron in the third layer, x(nq)|x(nq)|k -1 is the basis function of the polynomial model, K is the nonlinear order of the model, Q is the memory depth of the model, Re(·) means taking the real part, Im(·) means taking the imaginary part, w 1kq is the connection weight between the real part of the polynomial basis function and the real part of the output neuron, w 2kq is the connection weight of the imaginary and real output neurons of the polynomial basis function, w 3kq is the connection weight of the real and imaginary output neurons of the polynomial basis function, w 4kq is the connection weight between the imaginary part and the imaginary part output neurons of the polynomial basis function.

2. The continuous learning digital predistortion system for intelligent scenarios according to claim 1, characterized in that: The configuration adjustment module adjusts the input signal configuration in real time according to user needs. The input signal configuration includes: signal power, signal bandwidth and signal type; the signal power is adjusted according to the transmission distance, the signal bandwidth is adjusted according to the spectrum utilization, and the signal type is adjusted according to the signal transmission protocol.

3. The continuous learning digital predistortion system for intelligent scenarios according to claim 1, characterized in that: The continuous learning digital predistortion module performs continuous learning through the following process: storing all input and output signals that do not meet the linearity requirements under past configurations, and using both the input and output signals of past configurations and the input and output signals of current configurations when updating model coefficients.

4. The continuous learning digital predistortion system for intelligent scenarios according to claim 1, characterized in that: The linearity checking module calculates the linearity index of the collected output signal, and determines whether the linearity meets the requirements according to the index. The linearity index includes: normalized mean square error and adjacent channel leakage ratio.

5. The continuous learning digital predistortion system for intelligent scenarios according to claim 1, characterized in that: The objective function module calculates a joint objective function, and the model coefficient extraction module updates the model coefficients by minimizing the joint objective function. The joint objective function is expressed as the following expression: In the formula, J joint (θ) is the joint objective function, J1(θ) is the first objective function, J2(θ) is the second objective function, θ is all the coefficients in the model, including weight w and bias b, α is the weighting factor, is the target output signal of the digital predistortion model.

6. A method for continuous learning digital pre-distortion for intelligent scenarios, the method being applied to the continuous learning digital pre-distortion system for intelligent scenarios as claimed in any one of claims 1 to 5, characterized in that: The steps include: Step S1, the configuration adjustment module intelligently adjusts the configuration of the input signal, and the newly configured input signal generates a predistortion signal through the continuous learning digital predistortion module; Step S2, the predistortion signal generates an output signal through a power amplifier and is sent to a linearity inspection module; Step S3: The linearity checking module checks the linearity of the output signal. If the linearity meets the requirement, the digital predistortion model does not need to be updated, and the process proceeds to step S7. If the linearity does not meet the requirements, proceed to step S4; Step S4, storing the predistortion signal and the output signal under the new configuration into a training data storage module; Step S5, using the data in the training data storage module and according to the objective function of the objective function module, the model coefficient extraction module optimizes the model and updates the model coefficient; Step S6, copying the updated model coefficients to the continuous learning digital pre-distortion module; Step S7, continuous learning The digital pre-distortion system operates normally. When the configuration adjustment module is ready to adjust the input signal to a new configuration, step S1 is entered again.

Images