Digital intermediate frequency predistortion system and method based on direct learning network

Abstract

A digital intermediate frequency predistortion system and method based on a direct learning network comprises: a digital intermediate frequency predistortion network unit, radio frequency transmitters respectively connected thereto, a transmitter network parameter extraction unit and an end-to-end network construction and predistortion network parameter extraction unit, and a feedback link unit respectively connected to the radio frequency transmitter output end and the transmitter network parameter extraction unit input end. The present invention utilizes the optimization idea in analog predistortion, introduces a nonlinear neural network structure, and reduces the algorithm complexity by sharing network parameters. A predistortion technology more suitable for digital intermediate frequency signals is designed by jointly optimizing mean square error and intermodulation interference indicators, so that the digital intermediate frequency predistortion module achieves better performance.

Description

Technical Field

[0001] The present invention relates to a technology in the field of power amplifiers, specifically a digital intermediate frequency predistortion system and method based on a direct learning network for introducing intermodulation interference indicators, which is applicable to various cellular network systems, narrowband or broadband systems, etc. having digital intermediate frequency processing modules. Background Art

[0002] The orthogonal frequency division multiplexing waveform used in the fifth generation mobile communication system (5G) is more sensitive to the nonlinear characteristics of the power amplifier due to its high peak-to-average ratio. Predistortion technology is a typical technology in linearization technology and can be applied in digital baseband, digital intermediate frequency and analog RF domains. Predistortion technology is divided into digital predistortion technology and analog predistortion technology, which are usually applied to digital baseband or intermediate frequency and analog RF domains. The optimization goal of existing digital predistortion modules is usually to minimize the mean square error. Due to its circuit characteristics, analog predistortion technology is difficult to model memory characteristics, and its optimization index is usually to minimize odd-order intermodulation interference. At present, digital predistortion technology is often deployed in the digital intermediate frequency domain together with peak reduction technology. Considering that both digital intermediate frequency and RF signals have the characteristics of bandpass signals, the joint optimization of mean square error and intermodulation interference indicators will bring greater gain to predistortion technology.

[0003] Since digital intermediate frequency signals are similar to RF signals and both have the characteristics of bandpass signals, from the frequency domain perspective, what affects the system linearity performance is the intermodulation nonlinear components in the frequency bands adjacent to the useful frequency band. The remaining components can be filtered out by filters, but the existing digital pre-distortion method of minimizing the mean square error will minimize all nonlinear spectral components. Summary of the invention

[0004] In view of the above-mentioned deficiencies in the prior art, the present invention proposes a digital intermediate frequency predistortion system and method based on a direct learning network, utilizes the optimization ideas in analog predistortion, introduces a nonlinear neural network structure to reduce the complexity of the algorithm, and designs a predistortion technology that is more suitable for digital intermediate frequency signals by jointly optimizing mean square error and intermodulation interference indicators, so that the digital intermediate frequency predistortion module achieves better performance.

[0005] The present invention is achieved through the following technical solutions:

[0006] The invention relates to a digital intermediate frequency predistortion system based on a direct learning network for introducing an intermodulation interference index. The system comprises: a digital intermediate frequency predistortion network unit, a radio frequency transmitter respectively connected thereto, a transmitter network parameter extraction unit and an end-to-end network construction and predistortion network parameter extraction unit, and a feedback link unit respectively connected to an output end of the radio frequency transmitter and an input end of the transmitter network parameter extraction unit, wherein: the digital intermediate frequency predistortion network unit receives a digital intermediate frequency signal from a front-stage device, performs predistortion network inference on the signal, and outputs a digital intermediate frequency signal to the radio frequency transmitter and the transmitter network parameter extraction unit; the feedback link unit down-converts the radio frequency signal output by the radio frequency transmitter to a digital intermediate frequency and generates a nonlinear system output data set; the transmitter network parameter extraction unit obtains nonlinear neural network parameters of the radio frequency transmitter through iterative extraction using the collected digital intermediate frequency signal and the nonlinear system output data set; the end-to-end network construction and predistortion network parameter extraction unit establishes an end-to-end system neural network model according to the transmitter network parameters, and obtains the predistortion network parameters through iterative extraction according to a joint optimization target.

[0007] The radio frequency transmitter comprises: a digital-to-analog converter, an up-converter and a power amplifier, wherein: the digital-to-analog converter performs digital-to-analog conversion according to a digital intermediate frequency signal, and outputs a radio frequency signal after up-conversion processing and power amplification.

[0008] The end-to-end system neural network model is designed based on a time delay neural network (TDNN), whose input data is current and historical memory data, and whose output data is a multi-layer fully connected network of current data, specifically including: a pre-distortion network and a transmitter network based on a time delay neural network with memory characteristics.

[0009] The pre-distortion network includes: N time-delay neural networks with the same structure, the input layer of each time-delay neural network has M neuron nodes, including the current input and M-1 historical inputs, that is, the input memory depth of each time-delay neural network is M, the number of hidden layers is K, and the output layer has 1 neuron node, which represents the current pre-distorter output.

[0010] The transmitter network is a digital network of a radio frequency transmitter, comprising: a time-delay neural network, an input layer having N neuron nodes, a hidden layer having Q layers, and an output layer having 1 neuron node.

[0011] The iterative extraction obtains the pre-distortion network parameters, which are obtained by calculating the joint optimization cost function and learning the parameters according to the joint optimization target through an end-to-end time-delay neural network architecture with memory characteristics. That is, during the training process, the update of the transmitter network layer parameters is first shielded, and then the pre-distortion network parameters are updated iteratively according to the joint optimization index and the parameters are shared between the pre-distortion sub-networks.

[0012] The transmitter network is first independently trained by the nonlinear neural network parameters output by the transmitter network parameter extraction unit, and then cascaded with the predistortion network for joint training, wherein: during independent training, it is separated from the predistortion network, and the input and output data sets are the collected transmitter digital intermediate frequency signals; during joint training, it is cascaded with the predistortion network, and the input of the transmitter network is the current predistortion network output y(n) and the historical predistortion network output {y(n-1), ..., y(ni), ..., y(nN)}, and each predistortion network output y(ni) corresponds to M inputs {x(ni), ..., x(niM)}.

[0013] In order to further simplify the iteration complexity, the nonlinear characteristics of the transmitter are assumed to be stable within the memory depth M, so that all pre-distortion sub-networks can share a set of weights and bias parameters.

[0014] The predistortion network inference refers to obtaining the corresponding neural network predistortion output signal in real time through the front-stage digital intermediate frequency input signal according to the updated network parameters.

[0015] The iterative extraction obtains the nonlinear neural network parameters of the radio frequency transmitter by using a transmitter neural network architecture of a time-delay neural network with memory characteristics and performing a parameter iteration algorithm based on the mean square error.

[0016] The joint optimization target includes: using odd-order intermodulation interference and mean square error as joint optimization indicators in a dual-tone scenario and using adjacent channel power ratio and mean square error as joint optimization indicators in a multi-tone scenario, specifically: joint optimization target Cost = MSE + μ ANA, where: MSE is the mean square error between the network output and the expected output, μ is the normalization coefficient, and MSE and ANA are adjusted to the same order of magnitude, and intermodulation interference related indicators and They are the odd-order intermodulation interference of the upper sideband and the lower sideband, indicating the size of the odd-order intermodulation spectrum components of the adjacent main tone caused by nonlinear effects, ACPR U and ACPR L They are the adjacent channel power ratios of the upper sideband and the lower sideband, respectively, indicating the power leakage to the adjacent channel, that is, the power in the adjacent channel divided by the power in the useful channel.

[0017] The predistortion network parameters are updated and iterated in the following manner:

[0018] 1) The feedback link unit performs down-conversion and analog-to-digital conversion on the RF transmission signal to obtain a digital intermediate frequency signal. The transmitter network parameter extraction unit obtains the parameters of the transmitter neural network, including the weight and bias of the network, based on the collected RF transmitter input and output data set through the standard neural network training process. The cost function is the mean square error between the network output and the actual collected transmitter output signal.

[0019] 2) The end-to-end network construction and pre-distortion network parameter extraction unit constructs an end-to-end nonlinear neural network structure based on the transmitter network parameters obtained by the transmitter network parameter extraction unit; then the network structure is used to iterate the parameters of the pre-distortion network according to the joint optimization objective, and the update of the transmitter neural network parameters needs to be shielded during the iteration process.

[0020] 3) The digital intermediate frequency predistortion network unit performs real-time predistortion processing on the digital intermediate frequency signal based on the end-to-end network construction and the predistortion network parameters obtained by the predistortion network parameter extraction unit, and sends the generated predistorted digital intermediate frequency signal to the RF transmitter.

[0021] The predistortion network parameters are first determined based on the system output linearity to determine whether to update the predistortion network parameters. When no update is required, the historical predistortion network parameters are directly used to perform real-time digital intermediate frequency predistortion algorithm reasoning.

[0022] Technical Effects

[0023] The present invention solves the problem that the existing digital intermediate frequency predistortion method based on neural network does not consider the optimization of nonlinear components at specific frequency points. The present invention uses a dual-delay neural network cascade structure to perform end-to-end modeling of the predistorter and the transmitter, wherein the predistortion network uses parameter sharing to reduce the training complexity. The architecture design fully considers current data and historical data, which can well compensate for the performance degradation caused by memory characteristics; using the neural network architecture and the concept of direct learning, through iterative updates of input and output data sets, based on the designed joint optimization cost function, the relevant network parameters are dynamically obtained.

[0024] Compared with the prior art, the present invention has improved performance on third-order intermodulation and adjacent channel power ratio indicators for digital intermediate frequency signals, and has lower model complexity than the existing Volterra series-based model. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a schematic diagram of the system of the present invention;

[0026] Figure 2 It is a schematic diagram of the neural network model of the end-to-end system;

[0027] Figure 3 It is a flow chart of the present invention;

[0028] Figure 4 Schematic diagram of the structure of a time-delay neural network (TDNN) according to an embodiment. DETAILED DESCRIPTION

[0029] like Figure 3 As shown, this embodiment relates to a predistortion method for introducing intermodulation interference indicators for digital intermediate frequency signals based on a direct learning network, comprising:

[0030] Step 1: Determine whether the pre-distortion network parameters need to be updated: perform error calculation on the system output or monitor the frequency domain nonlinear characteristics of the output signal. When the nonlinearity exceeds a certain threshold, start the network parameter update process.

[0031] Step 2: Obtain an output data set through a feedback link unit: The feedback link unit converts the RF transmission signal into a digital intermediate frequency domain to generate an output data set with nonlinear characteristics.

[0032] Step 3, parameter extraction of the transmitter network parameter extraction unit: The transmitter network parameter extraction unit constructs a network architecture based on a time-delay neural network to fit the nonlinear characteristics of the RF transmitter. In order to fully consider the memory characteristics, the input of the network is the current digital intermediate frequency input and historical input of the RF transmitter, and the output is the current output, such as Figure 2 As shown. Finally, the network parameters are iterated according to the above collected RF transmitter input and output data sets, and the loss function is the mean square error between the network output and the actual collected data. When the error is less than the threshold, the iteration is stopped and the network weights and bias parameters are saved.

[0033] Step 4: Parameter extraction of the end-to-end network construction and pre-distortion network parameter extraction unit: The end-to-end network construction and pre-distortion network parameter extraction unit constructs an end-to-end system neural network according to the network parameters obtained above, and iterates the pre-distortion network parameters according to the joint optimization target to obtain the final result, specifically including:

[0034] Step 4.1. Construct an end-to-end system neural network based on the transmitter network parameters: The predistortion network is also based on the architecture of the time-delay neural network. Similar to the transmitter network, considering the memory characteristics, the input of this network is the current input and the historical input, and the output is the current output. It contains a total of N sub-networks. This network is cascaded with the transmitter network obtained above to form an end-to-end neural network structure. The first few layers of the network are the predistortion network, and the last few layers are the transmitter network, such as Figure 2 Shown

[0035] Step 4.2, the learning process of the predistortion network parameters according to the joint optimization index: first, the parameter update of the transmitter network layer is shielded to ensure that the transmitter nonlinear model remains unchanged. Then, the predistortion network parameters of the first few layers are iteratively updated according to the joint optimization target.

[0036] Step 4.3, determine whether the optimization target is less than the threshold and update the parameters: monitor whether the optimization target is less than the preset threshold during the parameter update iteration process. If not, continue to iterate the parameters. If less than the threshold, update the obtained network parameters to the digital intermediate frequency predistortion network unit.

[0037] Step 5: The digital intermediate frequency predistortion network unit performs real-time predistortion processing: The digital intermediate frequency predistortion network unit infers the predistortion output signal based on the latest network parameters, and outputs the predistortion signal to the RF transmitter.

[0038] In the dual-tone signal scenario of this embodiment, if the interval between the dual tones is close, the nonlinear characteristics of the frequency domain are relatively stable, so the dual-tone signal with a fixed frequency can be used as the input and output data set to extract the parameters of the pre-distortion network. In this embodiment, the pre-distortion network parameter extraction unit first uses the dual-tone training signal to extract parameters, and then uses the dual-tone test signal to perform performance testing. The frequencies of the dual-tone training signals used are 12.1MHz and 14.1MHz; the frequencies of the dual-tone test signals are 12.3 / 14.1MHz and 13.1 / 13.8MHz.

[0039] In the dual-tone signal scenario, third-order intermodulation is used as the performance indicator, which is defined as the third-order intermodulation frequency power divided by the fundamental frequency power. The specific mean square error and third-order intermodulation interference index performance are shown in Table 1.

[0040] Table 1 Performance indicators of various methods in dual-tone scenarios

[0041]

[0042] In this embodiment, in a multi-tone signal scenario, the nonlinear characteristics of the frequency domain are relatively complex, so a broadband signal is used as an input and output data set to extract parameters of the predistortion network. In this embodiment, the predistortion network parameter extraction unit first uses a broadband training signal to extract parameters, and then uses a broadband test signal to perform performance testing.

[0043] In the multi-tone signal scenario, the adjacent channel power ratio is used as the performance indicator. The specific mean square error and adjacent channel power ratio indicators are shown in Table 2.

[0044] Table 2 Performance indicators of various methods in multi-tone scenarios

[0045]

[0046] Compared with the prior art, in the dual-tone signal scenario, this method has a 7-10dB gain in the third-order intermodulation index over the prior algorithm. In the multi-tone signal scenario, this method has a 4-10dB gain in the adjacent channel power ratio index over the prior algorithm.

[0047] The above-mentioned specific implementation can be partially adjusted in different ways by those skilled in the art without departing from the principle and purpose of the present invention. The protection scope of the present invention shall be based on the claims and shall not be limited by the above-mentioned specific implementation. Each implementation scheme within its scope shall be subject to the constraints of the present invention.

Claims

1. A digital intermediate frequency predistortion system based on a direct learning network and introducing intermodulation interference indicators, characterized in that: include: A digital intermediate frequency predistortion network unit, an RF transmitter, a transmitter network parameter extraction unit and an end-to-end network construction and predistortion network parameter extraction unit respectively connected thereto, and a feedback link unit respectively connected to the output end of the RF transmitter and the input end of the transmitter network parameter extraction unit, wherein: the digital intermediate frequency predistortion network unit receives a digital intermediate frequency signal from a preceding device and performs predistortion network inference on the signal to output a digital intermediate frequency signal to the RF transmitter and the transmitter network parameter extraction unit; the feedback link unit down-converts the RF signal output by the RF transmitter to a digital intermediate frequency and generates a nonlinear system output data set; the transmitter network parameter extraction unit obtains the nonlinear neural network parameters of the RF transmitter through iterative extraction from the collected digital intermediate frequency signal and the nonlinear system output data set; the end-to-end network construction and predistortion network parameter extraction unit establishes an end-to-end system neural network model according to the transmitter network parameters, and obtains the predistortion network parameters through iterative extraction according to the joint optimization target; The end-to-end system neural network model is designed based on a time-delay neural network, whose input data is current and historical memory data, and whose output data is a multi-layer fully connected network of current data, specifically including: a pre-distortion network and a transmitter network based on a time-delay neural network with memory characteristics.

2. The digital intermediate frequency predistortion system for introducing intermodulation interference index based on direct learning network according to claim 1 is characterized in that: The iterative extraction obtains the pre-distortion network parameters, and through an end-to-end time-delay neural network architecture with memory characteristics, the joint optimization cost function is calculated and the parameters are learned according to the joint optimization target. That is, during the training process, the update of the transmitter network layer parameters is first shielded, and then the pre-distortion network parameters are updated iteratively according to the joint optimization index and the parameters are shared between the pre-distortion sub-networks.

3. The digital intermediate frequency predistortion system for introducing intermodulation interference index based on direct learning network according to claim 1 or 2, characterized in that: The joint optimization target includes: using odd-order intermodulation interference and mean square error as joint optimization indicators in a dual-tone scenario and using adjacent channel power ratio and mean square error as joint optimization indicators in a multi-tone scenario, specifically: joint optimization target Cost = MSE + μ ANA, where: MSE is the mean square error between the network output and the expected output, μ is the normalization coefficient, and MSE and ANA are adjusted to the same order of magnitude, and intermodulation interference related indicators and They are the odd-order intermodulation interference of the upper sideband and the lower sideband, indicating the size of the odd-order intermodulation spectrum components of the adjacent main tone caused by nonlinear effects, ACPR U and ACPR L They are the adjacent channel power ratios of the upper sideband and the lower sideband, respectively, indicating the power leakage to the adjacent channel, that is, the power in the adjacent channel divided by the power in the useful channel.

4. The digital intermediate frequency predistortion system based on direct learning network with introduction of intermodulation interference index according to claim 1 is characterized in that: The pre-distortion network comprises: N time-delay neural networks with the same structure, each time-delay neural network has an input layer with M neuron nodes, including current input and M-1 historical inputs, that is, each time-delay neural network has an input memory depth of M, a hidden layer number of K, and an output layer with 1 neuron node, representing the current pre-distorter output; The transmitter network is a digital network of a radio frequency transmitter, comprising: a time-delay neural network, an input layer having N neuron nodes, a hidden layer having Q layers, and an output layer having 1 neuron node.

5. The digital intermediate frequency predistortion system for introducing intermodulation interference index based on direct learning network according to claim 1 or 4, characterized in that: The transmitter network is first independently trained by the nonlinear neural network parameters output by the transmitter network parameter extraction unit, and then cascaded with the predistortion network for joint training, wherein: during independent training, it is separated from the predistortion network, and the input and output data sets are the collected transmitter digital intermediate frequency signals; during joint training, it is cascaded with the predistortion network, and the input of the transmitter network is the current predistortion network output y(n) and the historical predistortion network output {y(n-1), ..., y(ni), ..., y(nN)}, and each predistortion network output y(ni) corresponds to M inputs {x(ni), ..., x(niM)}.

6. The digital intermediate frequency predistortion system for introducing intermodulation interference index based on direct learning network according to claim 1 or 2, characterized in that: The predistortion network parameters are updated and iterated in the following manner: 1) The feedback link unit down-converts the RF transmission signal and performs analog-to-digital conversion to obtain a digital intermediate frequency signal. The transmitter network parameter extraction unit obtains the parameters of the transmitter neural network, including the weight and bias of the network, according to the collected RF transmitter input and output data set through a standard neural network training process; the cost function is the mean square error between the network output and the actual collected transmitter output signal; 2) After the end-to-end network construction and pre-distortion network parameter extraction unit constructs an end-to-end nonlinear neural network structure based on the transmitter network parameters obtained by the transmitter network parameter extraction unit, the network structure is used to iterate the parameters of the pre-distortion network according to the joint optimization objective. During the iteration process, the update of the transmitter neural network parameters needs to be shielded; 3) The digital intermediate frequency predistortion network unit performs real-time predistortion processing on the digital intermediate frequency signal based on the end-to-end network construction and the predistortion network parameters obtained by the predistortion network parameter extraction unit, and sends the generated predistorted digital intermediate frequency signal to the RF transmitter.

7. The digital intermediate frequency predistortion system for introducing intermodulation interference index based on direct learning network according to claim 6 is characterized in that: The predistortion network parameters are first determined based on the system output linearity to determine whether to update the predistortion network parameters. When no update is required, the historical predistortion network parameters are directly used to perform real-time digital intermediate frequency predistortion algorithm reasoning.

8. A direct learning network based predistortion method based on the system of any one of claims 1 to 7, characterized in that: include: Step 1: Determine whether the pre-distortion network parameters need to be updated: perform error calculation on the system output or monitor the frequency domain nonlinear characteristics of the output signal. When the nonlinearity exceeds a certain threshold, start the network parameter update process. Step 2, obtaining an output data set through a feedback link unit: the feedback link unit converts the RF transmission signal into a digital intermediate frequency domain to generate an output data set with nonlinear characteristics; Step 3, parameter extraction of the transmitter network parameter extraction unit: The transmitter network parameter extraction unit constructs a network architecture based on a time-delay neural network to fit the nonlinear characteristics of the RF transmitter. In order to fully consider the memory characteristics, the input of the network is the current digital intermediate frequency input and historical input of the RF transmitter, and the output is the current output; finally, the network parameters are iterated according to the above-collected RF transmitter input and output data sets, and the loss function is the mean square error between the network output and the actual collected data. When the error is less than the threshold, the iteration is stopped and the network weights and bias parameters are saved; Step 4, parameter extraction of the end-to-end network construction and pre-distortion network parameter extraction unit: The end-to-end network construction and pre-distortion network parameter extraction unit constructs an end-to-end system neural network according to the network parameters obtained above, and iterates the pre-distortion network parameters according to the joint optimization target to obtain the final result; Step 5: The digital intermediate frequency predistortion network unit performs real-time predistortion processing: The digital intermediate frequency predistortion network unit infers the predistortion output signal based on the latest network parameters, and outputs the predistortion signal to the RF transmitter.

9. The predistortion method according to claim 8, characterized in that: The step 4 specifically includes: Step 4.1, constructing an end-to-end system neural network according to the transmitter network parameters; Step 4.2, the learning process of the pre-distortion network parameters is carried out according to the joint optimization index: first, the parameter update of the transmitter network layer is shielded to ensure that the transmitter nonlinear model remains unchanged; then, the parameters of the first few layers of pre-distortion network are iteratively updated according to the optimization target; Step 4.3, determine whether the optimization target is less than a threshold and update the parameters: when the optimization target is monitored to be less than a preset threshold during the parameter update iteration process, the obtained network parameters are updated to the digital intermediate frequency predistortion network unit.

Images