Method for updating training parameters of digital pre-distortion system based on normalized mean square error

Abstract

The application provides a normalized mean square error (NMSE) based parameter updating method for a digital predistortion system, and belongs to the technical field of adaptive digital predistortion. The application is used for controlling the updating of parameters in the iterative training process of the digital predistortion system. Firstly, node data is collected from the digital predistortion system, and two normalized mean square errors NMSE1 and NMSE2 are calculated. Then, a threshold th1 for measuring the nonlinearity of a power amplifier and a threshold th2 for measuring the performance of the digital predistortion are set, and the parameter updating control is performed according to the relationship among NMSE1, NMSE2, th1 and th2. The application can adopt different strategies such as updating, retaining but not updating and discarding for the parameters obtained by training the digital predistortion system, guarantees the effectiveness of parameter transmission, ensures the stability of system performance, and thus improves the robustness of performance.

Description

Technical Field

[0001] This invention relates to the field of adaptive digital predistortion technology in wireless communication technology, and in particular to a method for updating training parameters of a digital predistortion system based on normalized mean square error. Background Technology

[0002] Currently, wireless communication is developing at an increasingly rapid pace. To achieve massive data transmission and the Internet of Things, it is necessary to not only transmit signals as efficiently as possible (with minimal signal distortion) but also to utilize energy efficiently during wireless communication. However, in practical communication systems, these two points are often contradictory. This is because the nonlinearity of the transmitter's core component, the radio frequency power amplifier (PA), is inevitable, especially when the PA operates near the saturation region. In this case, the input and output signal power amplitudes cannot maintain a linear relationship, leading to significant data transmission error rates and interference with adjacent channels.

[0003] To effectively correct the nonlinearity of RF power amplifiers and achieve both high linearity and high efficiency, Digital Pre-Distortion (DPD) technology has been proposed in existing technologies. With its advantages of flexible programming, moderate cost, and high correction accuracy, this technology has become the most mainstream linearization technology used in RF front-ends.

[0004] The core idea of ​​DPD is to first extract an inverse model of an RF PA, and then cascade the extracted model at the front end of the PA, thereby making the system exhibit linear characteristics, such as... Figure 1 As shown. Commonly used DPD structures include two types: direct learning structures and indirect learning structures, such as... Figure 2 As shown, the two methods select different nodes when calculating model parameters, and the indirect learning structure does not need to predict model coefficients or wait for a round in the initial iteration, allowing more time for model training and algorithm convergence. However, regardless of whether it is a direct or indirect learning structure, the trained parameters are often directly passed, without considering the following possible situations:

[0005] (1) If the nonlinearity of PA is too severe or the parameters are incorrect, the obtained inverse model is at great risk.

[0006] (2) The performance of the parameters in a certain training session is poor or average, and it is not advisable or necessary to update them;

[0007] (3) During iterative training, if the performance is not good after multiple iterations, it needs to be discarded and recalculated.

[0008] Therefore, when the parameter passing strategy is inappropriate, it can easily lead to a sudden deterioration or continuous decrease in DPD performance, or even a failure to converge. Contents of the Invention

[0009] In view of the defects in the prior art, the present invention provides a digital predistortion system training parameter updating method based on normalized mean square error. The present invention can adopt different strategies such as updating, retaining without updating, and discarding the parameters obtained from training the digital predistortion system, so as to ensure the effectiveness of parameter transmission, guarantee the stability of system performance, and thereby improve the robustness of performance.

[0010] To achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0011] The digital predistortion system training parameter updating method based on normalized mean square error is used to control parameter updating in the iterative training process of the digital predistortion system, and includes the following steps:

[0012] Step 1: Collect node data from the digital predistortion system, and calculate two normalized mean square errors NMSE1 and NMSE2.

[0013] For the direct learning structure, NMSE1 and NMSE2 are respectively:

[0014]

[0015]

[0016] For the indirect learning structure, NMSE1 and NMSE2 are respectively:

[0017]

[0018]

[0019] Wherein, x(i) represents the input of the pre-predistorter, y(i) represents the output of the predistorter or the input of the power amplifier, z(i) represents the output of the pre-predistorter or the input of the power amplifier, k(i) represents the output of the post-predistorter, and T is the data length.

[0020] Step 2: Set a threshold th1 for measuring the nonlinearity of the power amplifier, and a threshold th2 for measuring the performance of the digital predistortion, and perform parameter update control according to the relationships among NMSE1, NMSE2, th1 and th2:

[0021] If NMSE1>NMSE2, reset the parameters to initial values and upload them to the pre-predistorter.

[0022] If NMSE1<NMSE2 and NMSE2>th1, stop the current iteration and start the next iteration.

[0023] If NMSE1<NMSE2 and NMSE2<th1 and NMSE1<th2, perform parameter iteration and upload the parameters to the pre-distorter;

[0024] If NMSE1<NMSE2 and NMSE2<th1 and NMSE1>th2, perform parameter iteration, but do not upload the parameters to the pre-distorter.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] 1. Improve the robustness of the system

[0027] Since digital pre-distortion is a model that undergoes continuous iterative training, the calculated parameters are the key factors of system performance. Each operation often affects the next operation, forming a correlated process. The error and influence in a certain iteration may be continuously amplified. Therefore, selecting a corresponding strategy according to different situations can improve the stability of system performance, help the algorithm converge to the expected value more quickly and stably, and avoid scenarios where performance becomes increasingly worse.

[0028] 2. Reduce unnecessary resource overhead

[0029] The implementation of a digital pre-distortion system usually adopts a lookup table manner, and each set of coefficients obtained from training corresponds to one lookup table. Iterative training requires matrix operations. If iteration or update is performed every time, a large amount of time or computing resources will inevitably be consumed, which is a burden even in the offline mode. Therefore, flexible control can be implemented to prune the results in a timely manner, reducing unnecessary time overhead, storage overhead and operation overhead, so as to alleviate the system burden.

[0030] 3. Flexible design for easy maintenance

[0031] Different power amplifiers have different characteristics, and systems also differ from each other. By adjusting the thresholds in the design, corresponding design can be flexibly performed for different systems, which facilitates achieving the expected system performance. Brief Description of the Drawings

[0032] Figure 1 This is a graph of the input and output characteristics of PA and DPD.

[0033] Figure 2 This is a schematic diagram of the learning structure of DPD, wherein the upper diagram shows a direct learning structure, and the lower diagram shows an indirect learning structure.

[0034] Figure 3 This is a flowchart of the method of the present invention. Detailed Description of the Embodiments

[0035] The present invention will be further described in detail below with reference to the accompanying drawings.

[0036] A method for updating training parameters of a digital predistortion system based on normalized mean square error, which is used to control parameter update during the iterative training process of the digital predistortion system, as Figure 3 shown, includes the following steps:

[0037] Step 1: Collect node data from the digital predistortion system, and calculate two normalized mean square errors NMSE1 and NMSE2;

[0038] See Figure 2 , for the direct learning structure, NMSE1 and NMSE2 are respectively:

[0039]

[0040]

[0041] For the indirect learning structure, NMSE1 and NMSE2 are respectively:

[0042]

[0043]

[0044] Wherein, x(i) represents the input of the front predistorter, y(i) represents the output of the predistorter or the input of the power amplifier, z(i) represents the output of the front predistorter or the input of the power amplifier, k(i) represents the output of the post-predistortion, and T is the data length;

[0045] Step 2: Set a threshold th1 for measuring the nonlinearity of the power amplifier and a threshold th2 for measuring the performance of the digital predistortion, and perform parameter update control according to the relationships among NMSE1, NMSE2, th1 and th2:

[0046] If NMSE1>NMSE2, reset the parameters to initial values and upload them to the front predistorter;

[0047] If NMSE1<NMSE2 and NMSE2>th1, stop the current round of iteration and start the next round of iteration;

[0048] If NMSE1<NMSE2, NMSE2<th1 and NMSE1<th2, perform parameter iteration and upload the parameters to the front predistorter;

[0049] If NMSE1<NMSE2, NMSE2<th1 and NMSE1>th2, perform parameter iteration but do not upload the parameters to the front predistorter.

[0050] The following is a more specific example:

[0051] For a digital predistortion system with an indirect learning structure, the parameter updates are controlled using the following steps:

[0052] Step 1: Based on the input and output data of the pre-distorter, the output data of the power amplifier, and the post-distortion data, select the data with larger amplitude and calculate to obtain NMSE1 and NMSE2.

[0053] Step 2: Set threshold th1 to -15 and threshold th2 to -30.

[0054] Step 3, according to Figure 3 The update strategy should be selected and implemented.

[0055] Step 4: If the parameters are updated, the pre-distorter model is updated accordingly; otherwise, the previous model is used, and the next round of data pre-distortion is performed.

[0056] Step 5: Repeat Step 1 to control the next iteration.

[0057] From the above examples, it can be seen that this method does not require updating the model every time, saving some time and computational overhead, and ensuring the stability of system performance.

Claims

1. A method for updating training parameters of a digital predistortion system based on normalized mean square error, characterized in that, for controlling parameter update during the iterative training process of a digital predistortion system, comprising the following steps: Step 1, collecting node data from the digital predistortion system, and calculating two normalized mean square errors NMSE1 and NMSE2; For a direct learning structure, NMSE1 and NMSE2 are respectively: For an indirect learning structure, NMSE1 and NMSE2 are respectively: where x(i) denotes the input of the predistorter, y(i) denotes the output of the power amplifier, z(i) denotes the output of the predistorter or the input of the power amplifier, k(i) denotes the output of the post-predistorter, and T is the data length; Step 2, setting a threshold th1 for measuring the nonlinearity of the power amplifier and a threshold th2 for measuring the performance of the digital predistortion, and performing parameter update control according to the relationship among NMSE1, NMSE2, th1 and th2: If NMSE1>NMSE2, reset the parameters to initial values, and upload the parameters to the predistorter; If NMSE1<NMSE2, NMSE2>th1, stop the current iteration and start the next iteration; If NMSE1<NMSE2, NMSE2<th1 and NMSE1<th2, perform parameter iteration, and upload the parameters to the predistorter; If NMSE1<NMSE2, NMSE2<th1 and NMSE1>th2, perform parameter iteration, without uploading the parameters to the predistorter.

Images