Soft partitioning method and device for multi-dimensional IMSA power amplifier model

Abstract

The invention discloses a soft partition method and device for a multi-dimensional IMSA (International Mobile Subscriber Amplifier) power amplifier model. An MD-IMSA model is a multi-band power amplifier model with low complexity and high performance, the core operation of the MD-IMSA model is to divide the amplitude space of a multi-dimensional input signal into a plurality of regions and fit the nonlinearity of a power amplifier in each region, but how to optimize the regions is always a problem. The invention provides a soft partitioning method for hybrid expert (MoE)-based optimization. Compared with the traditional uniform hard partitioning, the performance is obviously improved.

Description

Technical Field

[0001] The present invention belongs to the field of digital predistortion of radio frequency power amplifiers, and particularly relates to a soft partitioning method and device for a multi-dimensional IMSA power amplifier model. Background Art

[0002] With the increasing demand for high-performance and high-linearity amplifiers in modern communication systems, power amplifiers (PAs) play a particularly important role in radio frequency systems. Multi-band power amplifiers are key components that can operate simultaneously on multiple frequency bands and provide high-power output, and are widely used in 5G, 6G, and other advanced communication systems. It can efficiently support multiple communication standards, reduce system complexity, and improve spectrum utilization. Compared with traditional single-band power amplifiers, multi-band power amplifiers have higher flexibility and adaptability, amplifying power on multiple frequency bands to meet different spectrum requirements. However, due to the non-linear characteristics of power amplifiers, especially at high power outputs, serious signal distortion often occurs, affecting system performance, such as bandwidth limitation, signal distortion, and power efficiency degradation. In multi-band power amplifiers, more serious distortion is caused by the intermodulation of signals in multiple frequency bands. In the design of radio frequency power amplifiers, how to effectively suppress these non-linear distortions has become an urgent technical challenge.

[0003] Digital predistortion (DPD) technology has been widely used in the linearization of power amplifiers. For multi-band power amplifiers, various multi-band power amplifier behavior models have been proposed and widely applied. Among them, the multi-dimensional instantaneous sample index amplitude selection affine function (MD-IMSA) power amplifier model proposed by Wang Kai et al. has the advantages of low complexity and high performance. The core operation of this model is to divide the amplitude space of multi-dimensional input signals into several regions, and fit the non-linearity of the power amplifier separately in each partition. However, how to optimize the partitioning has always been a difficult problem. The original MD-IMSA model is a uniform hard partition in the amplitude space, but the actual communication signals are not uniformly distributed in amplitude. Therefore, this is not an optimized partition. Summary of the Invention

[0004] The purpose of the present invention is to provide an optimized soft partitioning method and device applicable to the MD-IMSA power amplifier model, so as to avoid the overfitting problem caused by too few data samples in some partitions due to uniform hard partitioning, and at the same time improve the performance of the MD-IMSA model under the same number of partitions.

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

[0006] 1. The device proposed by the present invention includes the following modules:

[0007] Digital predistorter based on an optimized soft-partitioned MD-IMSA model: for the input digital baseband signal x of D frequency bands1 (n), x 2 (n), …, x D Perform predistortion processing on (n), and generate digital predistortion signals with characteristics opposite to those of the RF power amplifier in each corresponding frequency band respectively in each frequency band.

[0008] D digital-to-analog converters: Convert the digital predistortion signals of D frequency bands into analog baseband signals.

[0009] D modulators: Up-convert the analog baseband signals of D frequency bands to the corresponding RF frequencies respectively.

[0010] Power combiner: Add the RF signals of D frequency bands and combine them into one path.

[0011] Multi-band power amplifier: Amplify the multi-band RF signal combined by the power combiner and output it.

[0012] Coupler: Couple a small part from the output signal of the power amplifier for feedback.

[0013] Power splitter: Divide the feedback signal into D paths for processing in each frequency band.

[0014] D demodulators: Down-convert the RF signal of each frequency band to obtain the analog baseband signal.

[0015] D analog-to-digital converters: Convert the analog baseband signals of D frequency bands into the corresponding digital baseband signals y 1 (n), y 2 (n), …, y D (n);

[0016] Model soft partitioning and parameter training algorithm: Calculate the MD-IMSA model optimization soft partitioning and model parameters according to the input and output signals x 1 (n), x 2 (n), …, x D (n) and y 1 (n), y 2 (n), …, y D (n).

[0017] 2. On the basis of 1, taking the forward model of the power amplifier corresponding to the signal of frequency band 1 as an example for derivation, and so on for the remaining frequency bands. At this time, the corresponding MD-IMSA model expression is:

[0018]

[0019] In formula (1), \(n\in[1,N]\) represents the index position of the sampling points of the digital baseband signal, and the total length of the baseband signal is \(N\) sampling points; \(m\) is the memory depth and \(m\in[1,M]\). \(M\) is the maximum memory depth. \(p\in[1,P]\) is the index position of the partition, and \(P\) is the total number of partitions. \(d\in[1,D]\) is the index label of the frequency band, and \(D\) is the total number of frequency bands of the power amplifier. \(x 1 (n), \(x 2 (n),…, \(x D (n) is the input complex signal for each frequency band, and \(||\) is defined as the operation of taking the modulus value of the complex number. is the output complex signal of the model, is the output of the sub-model of the \(p\)-th partition, \(a md,p and \(b m,p are the parameters of the sub-model of the \(p\)-th partition. Define the \(D\)-dimensional amplitude vector \(A(n)=[|x 1 (n)|,|x 2 (n)|,…,|x D (n)|]\). For the MD-IMSA model, the partitioning method is to evenly divide the amplitude space corresponding to \(A(n)\) into \(P\) partitions, and these partitions are represented as \(g p (n)\) is the gating function that selects each partition, and is defined as

[0020]

[0021] It can be seen that MD-IMSA can be regarded as selecting the output of the sub-model of the corresponding \(p\)-th partition through \(g p (n)\), so it is a uniform hard partition. Since the actual communication signals are not evenly distributed in amplitude, and only the sub-model of one partition works at the same time, if the partition can be optimized and multiple sub-models work simultaneously, the model performance can be further improved. Use a probability model to model the model error and the gating function, and the expression is as shown in (3)

[0022]

[0023] where \(x(n)=[x 1 (n),x 2 (n),…,x D (n)]\), \(y 1 (n)\) is the actual output of the power amplifier in frequency band 1. The probability of the error of the sub-model of the \(p\)-th partition is Use a complex Gaussian process to model it, as shown in (4)

[0024]

[0025] where is the variance of this Gaussian process, \(\theta pis the parameter set of the probability model, defined as where c p =[a 00,p ,b 00,p …,a MD,p ,b MD,p is the parameter vector of the sub-model of the p-th partition. Regarding the probability model of g p (n), a variable z p (n) is defined. If A(n) belongs to partition p, the value is taken as 1, otherwise 0. a p is defined as the prior probability that any data belongs to the p-th partition, is a multi-dimensional Gaussian distribution, which represents the probability that A(n) appears at the corresponding position in the p-th partition. The expression is as shown in (5)

[0026]

[0027] In the formula, α p ={μ p ,∑ p} is the parameter set of this Gaussian distribution. μ p and ∑ p are the mean vector and covariance matrix of this Gaussian process respectively. T represents the transpose operation of a vector or matrix, -1 represents the inverse operation of a matrix. It can be seen that at this time, the gating function g p (n) of each partition is not equal to 0. Therefore, soft partitioning is achieved. This method of weighting multiple sub-models with a probability model is called the Mixture of Experts (MoE) method. If the gating function g p (n) defined in (3) is substituted into (1), the MD-IMSA model (MoE-MD-IMSA) with soft partitioning optimized based on the Mixture of Experts method is obtained.

[0028] 3. According to the MoE-MD-IMSA model of multi-band soft partitioning described in 2, the following introduces the method for solving the model parameters. This model uses the Expectation-Maximization (EM) method to solve. First, an auxiliary variable h p (n) is defined as

[0029]

[0030] The model parameters in equation (3) can be solved successively according to equation (7)

[0031]

[0032] where Y 1 =[y 1 (1),…,y 1 (N)] Tis the vector composed of output signals, Ψ is the corresponding model matrix in (1), and H represents the conjugate transpose of the complex matrix. W p is a diagonal weighting matrix with diagonal elements being {h p (1), …, h p (N)}. To achieve the best performance, the solution process requires multiple iterations. The specific process is as follows:

[0033] Step 3.1: Initialize the model parameters in formula (3).

[0034] Step 3.2: Substitute the current model parameters and the power amplifier input and output data into equations (3) and (4) to calculate the model error probability corresponding to each partition and the gate function probability g p (n).

[0035] Step 3.3: Substitute the current model parameters and the power amplifier input and output data into equation (6) to calculate the auxiliary variable h p (n) corresponding to each partition.

[0036] Step 3.4: Update the model parameters according to equation (7).

[0037] Step 3.5: Substitute the updated model parameters into equation (1) to obtain the model output under the current parameters, and determine whether the modeling error meets the requirements. If it meets, the parameter solution process is completed. Otherwise, go to Step 3.2 to start the next iteration.

[0038] Beneficial effects: A soft partitioning method and device for optimizing the MD-IMSA power amplifier model according to the present invention has the following advantages: Novelly modeling the contributions of sub-models in each partition by using multi-dimensional Gaussian distribution probability, realizing soft partitioning, enabling the sub-models corresponding to multiple partitions to work together, and improving the final model accuracy. At the same time, it effectively avoids problems such as overfitting caused by too few data samples in some partitions due to uniform hard partitioning. Brief Description of the Drawings

[0039] Figure 1 is a schematic structural diagram of the power amplifier digital predistortion device of the present invention;

[0040] Figure 2 is the power spectrum of the power amplifier output signal before and after performing predistortion on a dual-band signal with two 30 MHz bandwidths, 168.84 MSPS sampling rate, and 7 dB peak-to-average ratio according to the present invention. Detailed Embodiment

[0041] To better understand the purpose, structure, and function of the present invention, the following further describes in detail a soft partitioning device and method for a multi-dimensional IMSA power amplifier model of the present invention in conjunction with the accompanying drawings.

[0042] As shown Figure 1 in the figure, considering the scenario of a D-band power amplifier, the power amplifier digital predistortion device of the present invention includes a predistorter based on a soft-partitioned MD-IMSA model, analog-to-digital converters for frequency bands 1 to D, modulators for frequency bands 1 to D, a power combiner, a multi-band power amplifier, a coupler, a power divider, demodulators for frequency bands 1 to D, analog-to-digital converters for frequency bands 1 to D, and a model soft partitioning and parameter training algorithm and other modules.

[0043] The overall process of the present invention is as follows. Specifically, first, the multi-band baseband signals x 1 (n), x 2 (n), …, x D (n) sequentially pass through the digital-to-analog converters and quadrature modulators of their respective frequency bands to obtain analog radio frequency signals of their respective frequency bands. After being aggregated by the power combiner, they drive the multi-band radio frequency power amplifier. A part of the distorted radio frequency signal output by the power amplifier is coupled through the coupler for feedback. The coupled signal is output to the quadrature demodulator and analog-to-digital converter corresponding to each frequency band through the power divider to obtain the distorted digital baseband signals y 1 (n), y 2 (n), …, y D (n) output by each frequency band of the power amplifier. Then, set the model configuration parameters used, such as the number of partitions and the memory depth, etc. Finally, use the soft partitioning method proposed by the present invention to establish a power amplifier model for the output signals of each frequency band. Taking frequency band 1 as an example, this process will be introduced in detail.

[0044] Step 1. The expression of the MD-IMSA model corresponding to frequency band 1 at this time is:

[0045]

[0046] In formula (1), n ∈ [1, N] represents the index position of the digital baseband signal sampling points, and the total length of the baseband signal is N sampling points; m is the memory depth and m ∈ [1, M]. M is the maximum memory depth. p ∈ [1, P] is the index position of the partition, and P is the total number of partitions. d ∈ [1, D] is the index label of the frequency band, and D is the total number of frequency bands of the power amplifier. x 1 (n), x 2 (n), …, x D (n) are the input complex signals of each frequency band, and || is defined as the operation of taking the modulus value of a complex number. is the output complex signal of the model, is the output of the sub-model of the p-th partition, a md,p and b m,p are the parameters of the sub-model of the p-th partition. Define the D-dimensional amplitude vector A(n) = [|x 1(n)|,|x 2 (n)|,…,|x D (n)|]. For the MD - IMSA model, the partitioning method is to evenly divide the amplitude space corresponding to A(n) into P partitions, and these partitions are represented as g p (n) is the gating function for selecting each partition, defined as

[0047]

[0048] It can be seen that MD - IMSA can be regarded as passing through gp ( (n) to select the output of the sub - model corresponding to the p - th partition, so it is a uniform hard partition. Since the actual communication signals are not uniformly distributed in amplitude, and only the sub - model of one partition works at the same time, if the partitions can be optimized and multiple sub - models work simultaneously, the model performance can be further improved. A probability model is used to model the model error and the gating function, and the expression is as shown in (3)

[0049]

[0050] where x(n)=[x 1 (n),x 2 (n),…,x D (n)], y 1 (n) is the actual output of the power amplifier in frequency band 1. The probability of the error of the sub - model of the p - th partition is It is modeled using a complex Gaussian process, as shown in (4)

[0051]

[0052] where is the variance of this Gaussian process, and θ p is the parameter set of this probability model, defined as where c p =[a 00,p ,b 00,p …,a MD,p ,b MD,p is the parameter vector of the sub - model of the p - th partition. Regarding the probability model of g p (n), a variable z p (n) is defined. If A(n) belongs to partition p, this value takes 1, otherwise 0. a p is defined as the prior probability that any data belongs to the p - th partition, is a multi - dimensional Gaussian distribution, which represents the probability that A(n) appears at the corresponding position of the p - th partition, and the expression is as shown in (5)

[0053]

[0054] where α p ={μ p , ∑ p} is the parameter set of the Gaussian distribution, μ p and ∑ p are the mean vector and covariance matrix of the Gaussian process respectively, T represents the transpose operation of a vector or matrix, and -1 represents the inverse operation of a matrix. It can be seen that at this time, the gating function g p (n) of each partition is not equal to 0, so soft partitioning is achieved. This method of weighting multiple sub-models with a probability model is called the Mixture of Experts (MoE) method. If the gating function g p (n) defined in (3) is substituted into (1), the MoE-MD-IMSA model (MoE-MD-IMSA) of soft partitioning optimized based on the MoE method is obtained.

[0055] Step 2. According to the MoE-MD-IMSA model of multi-band soft partitioning described in Step 1, the following introduces the method for solving the model parameters. This model uses the Expectation-Maximization (EM) method to solve. First, define an auxiliary variable h p (n) as

[0056]

[0057] The model parameters in equation (3) can be solved sequentially according to equation (7)

[0058]

[0059] where Y 1 =[y 1 (1), …, y 1 (N)] T is the vector composed of the output signals, ψ is the corresponding model matrix in (1), and H represents the conjugate transpose of a complex matrix. W p is a diagonal weighting matrix with diagonal elements {h p (1), …, h p (N)}.

[0060] Step 3. To make the parameter solution in Step 2 achieve the best performance, the solution process needs to be iterated multiple times. The specific process is as follows:

[0061] Step 3.1. Initialize the model parameters in formula (3).

[0062] Step 3.2. Substitute the current model parameters and the power amplifier input and output data into equations (3) and (4) to calculate the model error probability corresponding to each partition and the gating function probability g p (n).

[0063] Step 3.3: Substitute the current model parameters and the power amplifier input and output data into Equation (6) to calculate the auxiliary variable h corresponding to each partition p (n).

[0064] Step 3.4: Update the model parameters according to Equation (7).

[0065] Step 3.5: Substitute the updated model parameters into Equation (1) to obtain the model output under the current parameters, and determine whether the modeling error meets the requirements. If it meets, the parameter solution process is completed. Otherwise, go to Step 3.2 to start the next iteration.

[0066] Through the above soft partition optimization and parameter training process, a forward model of the power amplifier frequency band 1 can be obtained. By swapping the positions of the above power amplifier output and input, the basis function optimization and coefficient estimation of the inverse model of the power amplifier, that is, the predistorter, can be completed. Update the predistorter with the finally selected basis function and model coefficients, and then inject the input signal into the predistorter to obtain the pre-distorted signal Complete the entire predistortion process, and so on for other frequency bands.

[0067] The baseband signal is a complex signal including a real part and an imaginary part. The tested power amplifier is a dual-band power amplifier, and the center frequencies of the two frequency bands are 3.42 GHz and 3.58 GHz respectively. Two baseband signals with a 30 MHz bandwidth, a 168.84 MSPS sampling rate, and a 7 dB peak-to-average ratio are modulated to the corresponding radio frequencies.

[0068] In formula (1), the configuration of the model is dimension D = 2, memory depth M = 5, and partition number P = 25. Then, optimize the soft partition of the model and train the parameters according to the steps proposed by the present invention, and generate the corresponding predistortion signal to linearize the power amplifier. Table 1 gives the normalized mean square error (NMSE) of the two frequency band modeling and the adjacent channel leakage ratio (ACLR) before and after predistortion. Figure 2 Shows the power spectra of the power amplifier output before and after predistortion in the two frequency bands. Combining Table 1 and Figure 2 it can be seen that after predistortion using the model proposed by the present invention, the signals in both frequency bands can obtain an improvement in NMSE performance of about 20 dB, and the ACLR of the corresponding sidebands in the two frequency bands is reduced by about 19 - 22 dB, indicating that the linearity of the power amplifier has been greatly improved.

[0069] Table 1. NMSE and ACPR of two frequency bands

[0070]

[0071] It is understood that the present invention is described by way of some embodiments, and those skilled in the art will be aware that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the present invention. Additionally, under the teaching of the present invention, these features and embodiments can 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 belong to the scope protected by the present invention.

Claims

1. A soft partitioning method for a multi-dimensional IMSA power amplifier model, characterized in that: The steps include: The power amplifier forward model corresponding to the signal of frequency band 1 is used as an example for derivation, and the rest of the frequency bands are deduced in the same way. The corresponding MD-IMSA power amplifier model expression is: In formula (1), n∈[1, N] represents the index position of the digital baseband signal sampling point, and the total length of the baseband signal is N sampling points; m is the memory depth and m∈[1,M], M is the maximum memory depth; p∈[1,P] is the index position of the partition, P is the total number of partitions; d∈[1,D] is the index of the frequency band; x1(n), x2(n), …, x D (n) is the input complex signal for each frequency band; | | is defined as the modulus operation on the complex number; is the complex output signal of the model, is the output of the sub-model of the pth partition, a md，p and b m，p is the parameter of the sub-model of the pth partition; Define a D-dimensional magnitude vector A(n) = [|x1(n)|, |x2(n)|, ..., |x D (n)|], for the MD-IMSA amplifier model, the partitioning method is to evenly divide the amplitude space corresponding to A(n) into P partitions, which are represented as g p (n) is the gate function for selecting each partition, defined as The probability model is used to model the MD-IMSA power amplifier model error and gate function. The expression is shown in (3): where x(n)=[x1(n),x2(n),…,x D (n)], y1(n) is the actual output of the power amplifier in band 1; the probability of error of the sub-model of the pth partition is Use the complex Gaussian process to model it, as shown in (4) in is the variance of the Gaussian process, θ p is the set of parameters of the probability model, defined as where c p =[a 00,p , b 00,p …, a MD,p , b MD，p ] is the parameter vector of the sub-model of the pth partition; p (n) probability model, define a variable z p (n), if A(n) belongs to partition p, the variable value is 1, otherwise it is 0; a p Defined as the prior probability that any data belongs to the pth partition, is a multidimensional Gaussian distribution, which represents the probability that A(n) appears at the corresponding position in the pth partition, as shown in (5) Where α p ={μ p ,∑ p } is the parameter set of Gaussian distribution, μ p and∑ p are the mean vector and covariance matrix of the Gaussian process, T Represents the transpose operation of a vector or matrix, -1 Represents the inversion operation of a matrix; The gate function g defined in (3) p Substituting (n) into (1), we obtain the MD-IMSA power amplifier model based on optimized soft partitioning.

2. A soft partitioning device for a multi-dimensional IMSA power amplifier model, characterized in that: include: A digital predistorter based on the optimized soft partitioned MD-IMSA power amplifier model obtained by the method of claim 1: for input digital baseband signals x1(n), x2(n), ..., x D (n) Perform pre-distortion processing to generate a digital pre-distortion signal in each frequency band that is opposite to the characteristics of the RF power amplifier in the corresponding frequency band. Indicates the index position of the digital baseband signal sampling point. The total length of the baseband signal is N sampling points. D digital-to-analog converters: converting the digital predistortion signals of D frequency bands into analog baseband signals; D modulators: up-convert the analog baseband signals of D frequency bands to the corresponding RF frequencies respectively; Power combiner: adds the RF signals of D frequency bands into one; Multi-band power amplifier: amplifies and outputs the multi-band RF signal synthesized by the power synthesizer; Coupler: couples part of the output signal from the power amplifier for feedback; Power divider: divides the feedback signal into D paths for processing in each frequency band; D demodulators: down-convert the RF signal of each frequency band to obtain an analog baseband signal; D analog-to-digital converters: convert the analog baseband signals of D frequency bands into corresponding digital baseband signals y1(n), y2(n), ..., y D (n); Model soft partitioning and parameter training algorithm module: According to the D frequency band input and output signals of the power amplifier x1(n), x2(n), ..., x D (n) and y1(n), y2(n), …, y D (n) Optimize the soft partitioning of the MD-IMSA power amplifier model and calculate the model parameters.

3. A soft partitioning device for a multi-dimensional IMSA power amplifier model according to claim 2, characterized in that: The expectation maximization method is used to solve the parameters of the MD-IMSA power amplifier model based on optimized soft partitioning.

4. A soft partitioning device for a multi-dimensional IMSA power amplifier model according to claim 3, characterized in that: The specific steps of solving the parameters of the MD-IMSA power amplifier model based on optimized soft partitioning using the expectation maximization method include: Step 4.1, initialize the MD-IMSA power amplifier model parameters in formula (3); Step 4.2: Substitute the current MD-IMSA amplifier model parameters and amplifier input and output data into formulas (3) and (4) to calculate the MD-IMSA amplifier model error probability corresponding to each partition. AND gate function probability g p (n); Step 4.3: Substitute the current MD-IMSA amplifier model parameters and amplifier input and output data into formula (6) to calculate the auxiliary variable h corresponding to each partition. p (n); Step 4.4, update the MD-IMSA power amplifier model parameters according to formula (7); where Y1=[y1(1),…,y1(N)] T is the vector of output signals, Ψ is the corresponding MD-IMSA amplifier model matrix in (1), H represents the conjugate transpose of the complex matrix, W p The diagonal elements are {h p (1),…,h p (N)} diagonal weighted matrix; Step 4.5: Substitute the updated MD-IMSA power amplifier model parameters into formula (1) to obtain the MD-IMSA model output under the current parameters, and determine whether the modeling error meets the requirements. If so, the parameter solution process is completed; otherwise, enter step 4.2 to start the next iteration.

Images