A neural network-based method for predicting peak-to-valley values of radio frequency signal envelope

Abstract

This invention proposes a method for predicting peak and valley values ​​of radio frequency (RF) signal envelopes based on neural networks. The method includes: generating an RF signal envelope from a communication baseband signal using a trained fitting neural network; performing label conversion on each data point of the RF signal envelope using a trained classification neural network, outputting RF signal envelope label data; and multiplying the value of each data point in the RF signal envelope by its corresponding label data to obtain the peak and valley value information of the RF signal envelope. The fitting neural network of this invention simplifies the process of generating the RF signal envelope, and the classification neural network uses three parallel backpropagation (BP) neural networks to simultaneously convert all data points in the RF signal envelope into three groups. This simple network structure achieves efficient conversion of RF signal envelope peak and valley value labels, significantly reducing system latency and required hardware resources, effectively improving the speed of obtaining reference signals, and making the power supply of mobile communication base stations more efficient and reliable.

Description

Technical Field

[0001] This invention relates to the field of envelope tracking power control technology, and more specifically to a method for predicting the peak and valley values ​​of radio frequency signal envelopes based on neural networks. Background Technology

[0002] Since its emergence in the last century, mobile communication has developed rapidly. It has become one of the major high-tech industries driving global economic development, profoundly impacting all aspects of people's lives, including clothing, food, housing, and transportation. In recent years, the scale of communication networks has continued to expand, and the number of mobile communication devices has also increased accordingly. As of 2017, the penetration rate of mobile communication in China had reached 100%. Meanwhile, with the construction of 5G base stations, the scale of communication base stations in China will see a greater increase than that of 4G base stations. Currently, the number of 5G base stations in China exceeds 1 million, and the number of 5G terminal connections exceeds 400 million. At the same time, the energy consumption of mobile communication is also increasing. It is estimated that by 2025, the communication industry will consume approximately 20% of the world's electricity. With the increasingly dense deployment of 5G base stations, reducing the energy consumption of communication systems is becoming increasingly important. Communication base stations in communication systems are the largest consumers of electricity, with approximately 80% of the entire communication industry's energy consumption distributed among the widely distributed base stations globally.

[0003] To increase the amount of transmitted information within the same frequency band, modern mobile communication modulation strategies simultaneously modulate the amplitude, phase, and frequency of the high-frequency carrier. This causes the envelope amplitude of the radio frequency (RF) signal to become non-constant. If a constant voltage power supply is continued to power the RF linear power amplifier, most of the energy is wasted as heat, resulting in very low efficiency. Therefore, improving the efficiency of the base station's RF linear power amplifier is the key to energy saving in modern mobile communication base stations. Envelope tracking technology is widely used, and in modern communication systems, broadband envelope tracking technology is required to meet the high-efficiency power supply requirements of the RF power amplifier. Common broadband ET power supplies include single-switch converter structures, parallel structures of multiple switch converters, and switch / linear composite structures. The switch / linear composite structure is currently the mainstream research direction, offering advantages such as lower switching frequencies and simpler circuit topologies compared to the other two structures. Parallel switch / linear composite envelope tracking power supplies use a segmented averaging reference method to control the switch converters in the ET power supply. By using a reference signal to control multiple switch converters in the parallel switch / linear composite envelope tracking power supply, the control complexity is reduced. The generation of the reference signal, which serves as the control signal, requires processing the amplitude of the peak and valley values ​​of the envelope of the radio frequency signal.

[0004] In conventional envelope tracking power supply applications, existing technologies primarily employ two methods to obtain envelope peak and valley values: one using hardware circuits, and the other using convolutional neural networks (CNNs) for prediction. Hardware circuits need to complete envelope extraction and control signal generation within a short timeframe, resulting in complex circuit design, time-consuming computations, and significant computational pressure. Furthermore, they require delays in the communication system to achieve time synchronization between the two systems. While CNN prediction can reduce hardware computational costs, the designed neural network structure is complex, computationally intensive, and, being a single neural network, it lacks the ability to determine the peak and valley types of the first and last data points in the grouped input data. This leads to missing envelope peak and valley information, potentially causing errors in the generation of control signals and impacting the efficiency of the envelope tracking power supply. Therefore, given the current emphasis on communication transmission efficiency, a major breakthrough in this field lies in improving the computational efficiency of the envelope system to increase communication speed. Designing a simple neural network model capable of determining peak and valley values ​​for all envelope data points represents a significant step forward.

[0005] A neural network model refers to an artificial neural network, similar to a biological neural network. It's a mathematical model that processes complex information through the interconnections of numerous nodes. Artificial neural networks don't require manual searching for the functional relationship between input and output; they are trained using existing input data and corresponding outputs to determine this relationship. When new inputs arrive, they produce the desired output. The most widely used type is the multilayer feedforward neural network (BP neural network), which employs the backpropagation learning algorithm. The training process of a BP neural network consists of two main parts: forward computation and backward computation. In the forward computation, the input data passes through the hidden and output layers layer by layer; the output of each layer only affects the output of the next layer. If the output obtained through the forward computation differs significantly from the desired output, the backward computation process begins. In the backward computation, the error of the current output is propagated forward, and each node adjusts its weights accordingly to reduce the error. This process is repeated multiple times to obtain the desired output.

[0006] Based on this, an innovative method is proposed to predict the extreme value information of the envelope of radio frequency signals using neural network technology. By introducing a BP neural network, the peak and valley information of the envelope can be obtained more efficiently and accurately, thereby obtaining a more effective reference signal as a control signal, thus realizing fast and accurate tracking of the entire envelope signal. Summary of the Invention

[0007] The purpose of this invention is to obtain tag data of the radio frequency signal envelope from the communication baseband signal through a fitting neural network and a classification neural network, and then obtain the peak and valley value information of the radio frequency signal envelope.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting the peak and valley values ​​of the envelope of a radio frequency signal based on a neural network, comprising the following steps:

[0009] S1: For the communication baseband signal, a pre-trained fitting neural network is applied to obtain the envelope of the radio frequency signal corresponding to the communication baseband signal. The fitting neural network takes the communication baseband signal as input and the envelope of the radio frequency signal corresponding to the communication baseband signal as output.

[0010] S2: Input the radio frequency signal envelope into the trained classification neural network to obtain the label data of the radio frequency signal envelope. The label data of the radio frequency signal envelope is a one-dimensional array composed of the label values ​​of each data point in the radio frequency signal envelope.

[0011] S3: Multiply the values ​​of each data point in the radio frequency signal envelope by their corresponding tag data to obtain the peak and valley values ​​of the radio frequency signal envelope.

[0012] Furthermore, the aforementioned fitting neural network was trained using the following method:

[0013] S11: Use MATALB to generate a preset number of n groups of communication baseband signals, and obtain the corresponding radio frequency signal envelope based on each group of communication baseband signals;

[0014] S12: Combine each group of communication baseband signals with their corresponding radio frequency signal envelopes to form a fitting sample data, thus forming n groups of fitting sample data.

[0015] S13: Construct the neural network to be trained and fitted;

[0016] S14: Use n sets of fitting sample data to train the fitting neural network to obtain a fitting neural network that can modulate and fit the input communication baseband signal into the envelope of the radio frequency signal.

[0017] Furthermore, the process of obtaining the radio frequency signal envelope in S11 includes: performing multi-carrier modulation on n groups of communication baseband signals according to a preset method; performing up-conversion on the modulated signals to obtain n groups of radio frequency signals; and extracting the envelopes of the n groups of radio frequency signals to obtain n radio frequency signal envelopes.

[0018] Furthermore, the neural network to be trained in S13 mentioned above is constructed using a BP neural network.

[0019] Furthermore, the aforementioned classification neural network is trained using the following method:

[0020] S21: For each of the n radio frequency signal envelopes, obtain the tag data of each envelope;

[0021] S22: Combine each radio frequency signal envelope with its corresponding tag data to form a classification sample data; thus forming n sets of classification sample data.

[0022] S23: Construct the classification neural network to be trained;

[0023] S24: Use n sets of classification sample data to train the classification neural network to obtain a classification neural network that can generate corresponding label data by inputting the envelope of radio frequency signals.

[0024] Furthermore, the process of obtaining the corresponding tag data from the radio frequency signal envelope in S21 includes: determining the classification rules for data point tags; dividing all data points into three categories based on their characteristics: peak points, valley points, and others; assigning a tag of 1 to peak points, -1 to valley points, and 0 to the remaining data points; obtaining the characteristic information of each data point in the radio frequency signal envelope using traditional methods of calculating the first and second derivatives; and converting the tags of each data point according to the classification rules based on the characteristic information of each data point, arranging the tags of each data point in chronological order to form a one-dimensional array, which is the tag data of the radio frequency signal envelope.

[0025] Furthermore, in the aforementioned S23, the classification neural network to be trained is constructed using three backpropagation (BP) neural networks. These three BP neural networks are combined in parallel and are referred to as the first neural network, the second neural network, and the third neural network, respectively. The first, second, and third neural networks simultaneously receive the radio frequency (RF) signal envelope from the fitting neural network. The first neural network is responsible for the label conversion of the intermediate data points of each subgroup under the first group of the RF signal envelope, obtaining the first label group; the second neural network is responsible for the label conversion of the intermediate data points of each subgroup under the second group of the RF signal envelope, obtaining the second label group; and the third neural network is responsible for the label conversion of the RF signal envelope... The labels of the intermediate data points in each group under the third group are converted to obtain the third label group; the first group is: starting from the first data point in the RF signal envelope, every 3 data points are grouped into a group; the second group is: starting from the second data point in the RF signal envelope, every 3 data points are grouped into a group; the third group is: starting from the third data point in the RF signal envelope, every 3 data points are grouped into a group; the first label group, the second label group, and the third label group are arranged into a row to form 3 rows of label groups, and then the number of each label is extracted in column order to obtain one-dimensional label data, which is used as the label data of the RF signal envelope.

[0026] The method for predicting peak and valley values ​​of radio frequency signal envelope based on neural networks described in this invention has the following technical advantages compared with existing technologies:

[0027] 1. This invention uses a fitting neural network to quickly generate the radio frequency signal envelope from the communication baseband signal, without waiting for the communication system to complete the multi-carrier modulation, up-conversion, and envelope extraction of the baseband signal. This simplifies the existing process of generating the radio frequency signal envelope and significantly reduces system latency and required hardware resources.

[0028] 2. The classification neural network of the present invention is constructed by three BP neural networks in parallel, which divides all data points in the envelope of the radio frequency signal into three groups and performs label conversion simultaneously. With a simple network structure, it realizes efficient conversion of peak and valley value labels of the radio frequency signal envelope, effectively improves the speed of obtaining the reference signal required for the envelope tracking power switch control signal, and makes the power supply of mobile communication base stations more efficient and reliable.

[0029] 3. The radio frequency signal envelope tag data obtained by this invention has a multiplexing function. It can be processed with the envelope information predicted by the BP neural network to obtain the predicted peak and valley value information of the envelope, or it can wait for the actual communication system to generate the radio frequency signal envelope information and then process it with the tag data to obtain the accurate peak and valley value information of the envelope, thus meeting the needs of different application scenarios. Attached Figure Description

[0030] Figure 1 This is a system framework diagram of the fitting neural network and classification neural network of this invention;

[0031] Figure 2 This is a training example diagram of the fitting neural network and the classification neural network of this invention;

[0032] Figure 3 This is an example diagram of the architecture of the present invention applied to an envelope tracking power supply implementation. Detailed Implementation

[0033] To better understand the technical content of the present invention, specific embodiments are described below in conjunction with the accompanying drawings.

[0034] In this invention, various aspects of the invention are described with reference to the accompanying drawings, in which numerous illustrative embodiments are shown. Embodiments of the invention are not limited to those depicted in the drawings. It should be understood that the invention is implemented through any of the various concepts and embodiments described above, as well as the concepts and embodiments described in detail below, because the concepts and embodiments disclosed herein are not limited to any particular implementation. Furthermore, some aspects of the invention disclosed may be used alone or in any suitable combination with other aspects of the invention disclosed.

[0035] like Figure 1As shown, in this embodiment, the fitting neural network is constructed using one BP neural network, and the classification neural network is constructed using a parallel combination of three BP neural networks, referred to as the first neural network, the second neural network, and the third neural network, respectively. A method for predicting the peak and valley values ​​of the envelope of radio frequency signals based on neural networks includes the following steps:

[0036] Step 1: The pre-trained fitting neural network is used to generate the predicted radio frequency signal envelope from the input communication baseband signal.

[0037] Step 2: Apply the pre-trained classification neural network to perform label conversion on each data point of the predicted RF signal envelope, outputting RF signal envelope label data. The classification neural network receives the RF signal envelope from the fitted neural network and first groups the data points on the RF signal envelope: starting from the first data point, every three data points are grouped together as the first group. The first neural network is responsible for the label conversion of the middle data points in each subgroup within the first group. That is, the first neural network converts the labels of the 2nd, 5th, 8th, 11th... data points in the RF signal envelope to 1, 0, -1, 0..., denoted as the first label group; starting from the first data point, every three data points are grouped together as the first subgroup. Starting with two data points, the data points are grouped into groups of three, forming the second group. The second neural network is responsible for the label conversion of the intermediate data points in each subgroup of the second group. That is, the second neural network converts the labels of the 3rd, 6th, 9th, 12th... data points in the RF signal envelope into 0, 1, 1, -1..., which is denoted as the second label group. Starting with the 3rd data point in the RF signal envelope, the data points are grouped into groups of three, forming the third group. The third neural network is responsible for the label conversion of the intermediate data points in each subgroup of the third group. That is, the third neural network converts the labels of the 4th, 7th, 10th, 13th... data points in the RF signal envelope into -1, 0, 0, 1..., which is denoted as the third label group. The first, second, and third label groups are arranged into three rows, forming three label groups. Then, the number of each label is extracted in column order to obtain one-dimensional label data: 1, 0, -1, 0, 1, 0, -1, 1, 0, 0, -1, 1..., which is the label data of the RF signal envelope.

[0038] Step 3: Multiply the value of each data point in the predicted RF signal envelope by its corresponding tag data. The peak value is converted to a positive value, the valley value to a negative value, and the rest of the data points are converted to 0. Arrange them in chronological order to form one-dimensional data, which is the peak and valley value information of the predicted RF signal envelope.

[0039] like Figure 2As shown, the training sample data for training the fitting neural network consists of 100 sets of communication baseband signals generated by MATLAB and the corresponding radio frequency (RF) signal envelopes for each set of communication baseband signals. The RF signal envelopes for each set of communication baseband signals are obtained through the following method: The 100 sets of communication baseband signals are multi-carrier modulated; the modulated signals are up-converted to obtain 100 sets of RF signals; envelope extraction is performed on each set of 100 RF signal envelopes to obtain 100 RF signal envelopes; then, the communication baseband signals are used as input, and the extracted RF signal envelopes are used as output to train the fitting neural network, resulting in a fitting neural network capable of generating corresponding RF signal envelopes from input communication baseband signals.

[0040] Meanwhile, the following methods were used to extract labels from 100 radio frequency signal envelopes: The classification rules for data point labels were determined, and all data points were divided into three categories based on their characteristics: peak points, valley points, and others. Peak points were labeled as 1, valley points as -1, and the remaining data points as 0. The characteristic information of each data point in the radio frequency signal envelope was obtained using traditional methods of calculating the first and second derivatives. Based on the characteristic information of each data point, the labels of each data point were converted according to the classification rules, and the labels of each data point were arranged sequentially in chronological order to form a one-dimensional array. This one-dimensional array is the label data of the radio frequency signal envelope.

[0041] The label data of 100 obtained radio frequency signal envelopes are extracted and used to form training sample data for a classification neural network. The radio frequency signal envelopes are used as input and the obtained label data are used as output to train the classification neural network to obtain a classification neural network that can generate corresponding label data by inputting radio frequency signal envelopes.

[0042] like Figure 3As shown, in this embodiment, a fitting neural network is used to obtain RF signal envelope data from the baseband signal, and then a classification neural network is used to obtain RF signal envelope tag data. The RF signal envelope data and tag data are then multiplied to obtain the peak-valley information of the RF signal envelope. These peak-valley values ​​are then fed into a reference signal generation module to obtain a reference signal, and finally, the corresponding control signal is input to a switching converter. The switching converter generates the power supply voltage required by the linear power amplifier based on the control signal and inputs it to the linear power amplifier. The method described in this invention eliminates the need to wait for the communication system to complete baseband signal multi-carrier modulation, up-conversion, envelope extraction, and envelope peak-valley value extraction, significantly reducing system latency and hardware resource requirements. Furthermore, the obtained envelope tag information has multiplexing capabilities; it can be processed with the envelope information predicted by the BP neural network to obtain predicted envelope peak-valley information, or it can wait for the actual communication system to generate the RF signal envelope information and then process it with the tag data to obtain accurate envelope peak-valley information, meeting the needs of different application scenarios. Using a BP neural network to predict the peak and valley values ​​of the RF signal envelope can effectively improve the speed of obtaining the reference signal required for the control signal of the power switch tube for envelope tracking, reduce RF signal delay and hardware resource requirements, and make the power supply of mobile communication base stations more efficient and reliable.

[0043] While the present invention has been described above with reference to preferred embodiments, it is not intended to limit the invention. Those skilled in the art can make various modifications and refinements without departing from the spirit and scope of the invention. Therefore, the scope of protection of the present invention shall be determined by the claims.

Claims

1. A method for predicting peak and valley values ​​of radio frequency signal envelopes based on neural networks, used to obtain peak and valley value information of the corresponding radio frequency signal envelope from an input communication baseband signal, characterized in that, Includes the following steps: S1: For the communication baseband signal, a pre-trained fitting neural network is applied to obtain the envelope of the radio frequency signal corresponding to the communication baseband signal. The fitting neural network takes the communication baseband signal as input and the envelope of the radio frequency signal corresponding to the communication baseband signal as output. S2: Input the radio frequency signal envelope into the trained classification neural network to obtain the label data of the radio frequency signal envelope. The label data of the radio frequency signal envelope is a one-dimensional array composed of the label values ​​of each data point in the radio frequency signal envelope. The classification neural network is trained using the following method: S21: For each of the n radio frequency signal envelopes, obtain the tag data of each envelope; S22: Combine each radio frequency signal envelope with its corresponding tag data to form a classification sample data; thus forming n sets of classification sample data. S23: Construct the classification neural network to be trained; In step S23, the classification neural network to be trained is constructed using three backpropagation (BP) neural networks. These three BP neural networks are combined in parallel and are referred to as the first neural network, the second neural network, and the third neural network, respectively. The first neural network, the second neural network, and the third neural network simultaneously receive the radio frequency (RF) signal envelope from the fitted neural network. The first neural network is responsible for the label conversion of the intermediate data points of each subgroup under the first group of the RF signal envelope, obtaining the first label group. The second neural network is responsible for the label conversion of the intermediate data points of each subgroup under the second group of the RF signal envelope, obtaining the second label group. The third neural network is responsible for the label conversion of the intermediate data points of each subgroup under the third group of the RF signal envelope, obtaining the third label group. The first grouping is as follows: starting from the first data point in the radio frequency signal envelope, every 3 data points are grouped together. The second grouping is as follows: starting from the second data point in the radio frequency signal envelope, every three data points are grouped together. The third grouping is as follows: starting from the third data point in the radio frequency signal envelope, every three data points are grouped together. Arrange the first tag group, the second tag group, and the third tag group into a row to form 3 rows of tag groups. Then, extract the number of each tag in column order to obtain one-dimensional tag data, which is used as the tag data of the radio frequency signal envelope. S24: Use n sets of classification sample data to train the classification neural network to obtain a classification neural network that can generate corresponding label data by inputting the envelope of radio frequency signals; S3: Multiply the values ​​of each data point in the radio frequency signal envelope by their corresponding tag data to obtain the peak and valley values ​​of the radio frequency signal envelope.

2. The method for predicting peak and valley values ​​of radio frequency signal envelope based on neural networks according to claim 1, characterized in that, The fitted neural network was trained using the following method: S11: Use MATLAB to generate a preset number of n sets of communication baseband signals, and obtain the corresponding radio frequency signal envelope based on each set of communication baseband signals; S12: Combine each group of communication baseband signals with their corresponding radio frequency signal envelopes to form a fitting sample data, thus forming n groups of fitting sample data. S13: Construct the neural network to be trained and fitted; S14: Use n sets of fitting sample data to train the fitting neural network to obtain a fitting neural network that can modulate and fit the input communication baseband signal into the envelope of the radio frequency signal.

3. The method for predicting peak and valley values ​​of radio frequency signal envelope based on neural networks according to claim 2, characterized in that, The process of obtaining the radio frequency signal envelope in S11 includes: The n groups of communication baseband signals are modulated by multiple carriers according to a preset method; The modulated signals are up-converted to obtain n sets of radio frequency signals; Envelopes of n groups of radio frequency signals are extracted to obtain n radio frequency signal envelopes.

4. The method for predicting peak and valley values ​​of radio frequency signal envelope based on neural networks according to claim 2, characterized in that, The neural network to be trained in S13 is constructed using a single backpropagation (BP) neural network.

5. The method for predicting peak and valley values ​​of radio frequency signal envelope based on neural networks according to claim 1, characterized in that, The process of obtaining the corresponding tag data from the radio frequency signal envelope in S21 includes: Determine the classification rules for data point labels. Based on the characteristics of each data point in the envelope, divide all data points into three categories: peak points, valley points, and others. Determine the label for peak points as 1, the label for valley points as -1, and the label for the remaining data points as 0. The characteristic information of each data point in the envelope of the radio frequency signal is obtained by using the traditional method of calculating the first and second derivatives; Based on the feature information of each data point, the data points are labeled according to the classification rules of the data point labels. The labels of each data point are arranged in chronological order to form a one-dimensional array, which is the label data of the radio frequency signal envelope.

Images