阅读说明：本技术 基于时频分析和卷积神经网络的qpsk信号解调方法 (QPSK signal demodulation method based on time-frequency analysis and convolutional neural network ) 是由 李月 叶亮 贺梦利 黄刚 朱倩倩 郑鑫宇 于 2020-07-29 设计创作，主要内容包括：一种基于时频分析和卷积神经网络的QPSK信号解调方法,具体涉及一种QPSK信号解调方法。本发明是为了解决QPSK信号受同频干扰与噪声污染严重问题,首先对接收到的QPSK脉冲成型信号进行时频分析得到时频图,并对时频图进行预处理；然后输入预先训练好的卷积神经网络对接收信号时频图进行分类,并依据分类结果选择滤波通域,控制时变滤波器对接收信号进行滤波及解调,得到解调数据；最后依据分类结果可得到接收信号所属符号样本集,可通过符号样本集对解调数据进一步纠错。通过本方法可以有效抑制同频干扰,提升接收端的信干噪比性能,降低误比特率。本发明适用于QPSK信号的解调。(A QPSK signal demodulation method based on time-frequency analysis and a convolutional neural network specifically relates to a QPSK signal demodulation method. In order to solve the problem that a QPSK signal is seriously polluted by same frequency interference and noise, the method comprises the steps of firstly carrying out time-frequency analysis on a received QPSK pulse forming signal to obtain a time-frequency diagram and preprocessing the time-frequency diagram; then inputting a pre-trained convolutional neural network to classify the time-frequency graph of the received signal, selecting a filtering pass domain according to a classification result, and controlling a time-varying filter to filter and demodulate the received signal to obtain demodulation data; finally, a symbol sample set to which the received signal belongs can be obtained according to the classification result, and further error correction can be carried out on the demodulated data through the symbol sample set. The method can effectively inhibit same frequency interference, improve the signal-to-interference-and-noise ratio performance of a receiving end and reduce the bit error rate. The invention is suitable for demodulation of QPSK signals.)

1. The QPSK signal demodulation method based on time-frequency analysis and convolutional neural network is characterized by comprising the following steps:

(1) receiving signals at a receiving end, and intercepting the received signals;

(2) performing time-frequency analysis on the intercepted signals to generate a time-frequency distribution graph, and preprocessing the generated time-frequency distribution graph;

(3) inputting the preprocessed time-frequency distribution graph into a trained convolutional neural network for classification, and determining a filtering pass domain corresponding to a received signal according to a classification result;

(4) generating a time-varying filtering template according to the time-frequency distribution map of the corresponding category, and performing time-varying filtering on the received signal to finally obtain a demodulation symbol;

(5) and further calculating Hamming distances between the demodulation symbols and 4 symbol samples corresponding to the received signal time-frequency distribution diagram, and correcting the demodulation symbols according to the symbol samples corresponding to the minimum Hamming distance.

2. The QPSK signal demodulation method based on time-frequency analysis and convolutional neural network of claim 1, wherein the length of the signal truncation in step (1) is 4 symbol periods.

3. The QPSK signal demodulation method based on time-frequency analysis and convolutional neural network of claim 1, wherein the preprocessing of step (2) comprises the following steps: and copying the images in the first half period of the time of each time-frequency distribution graph to the tail part of the time-frequency distribution graph on the time dimension, so that the time dimension of each image is extended from 4 symbol periods to 6 symbol periods.

4. The QPSK signal demodulation method according to claim 1, wherein the training method of the trained convolutional neural network in step (3) comprises the following steps:

(3.1) generating n time-frequency distribution maps according to I, Q paths of all possible transmitted data of QPSK signals, classifying the data to be transmitted with the same time-frequency distribution maps into one class, and generating n' class time-frequency distribution maps together;

(3.2) preprocessing the n' class time-frequency distribution graph of the classified QPSK signals;

and (3.3) training the neural network, and taking the m samples of each class of the preprocessed n' class time-frequency distribution graph as the input of the convolutional neural network to obtain the optimal weight of the neural network.

5. The QPSK signal demodulation method based on time-frequency analysis and convolutional neural network of claim 4, wherein n time-frequency distribution maps are generated in step (3.1), wherein n is 64.

6. The QPSK signal demodulation method according to claim 4, wherein the n 'class of time-frequency distribution patterns are generated in step (3.1), wherein n' is 16.

7. The QPSK signal demodulation method based on time-frequency analysis and convolutional neural network of claim 4, wherein the preprocessing of step (3.2) comprises the following steps: and copying the images in the first half period of the time of each time-frequency distribution graph to the tail part of the time-frequency distribution graph on the time dimension, so that the time dimension of each image is extended from 4 symbol periods to 6 symbol periods.

8. The QPSK signal demodulation method based on time-frequency analysis and convolutional neural network of claim 4, wherein the value of m in step (3.3) satisfies that m is greater than or equal to 100.

9. The QPSK signal demodulation method based on time-frequency analysis and convolutional neural network of claim 4, wherein the m samples in step (3.3) satisfy a clean time-frequency distribution map without interference and noise, and are obtained by duplicating each time-frequency distribution map after preprocessing m times.

Technical Field

The invention relates to the field of mobile communication, in particular to a QPSK signal demodulation method.

Background

In various non-orthogonal resource sharing modes, signals of different users are overlapped on a time-frequency plane, and a receiving end distinguishes the signals of the different users through the difference of a code domain, a space domain and a power domain to realize useful signal recovery. However, through time-frequency analysis of signals, we find that energy concentration areas of different user signals on a time-frequency plane are different, and can further combine a time-varying filtering method to filter received signals according to time-frequency distribution of useful signals, so as to improve the signal-to-interference-and-noise ratio of the signals and further improve the error rate performance. However, when time-varying filtering is used in a communication system, a filtering pass-band that conforms to the time-frequency distribution characteristics of useful signals cannot be accurately obtained from signals polluted by interference and noise. This patent selects the filtering pass-through field that matches with the received signal through the convolutional neural network, and then accomplishes the time-varying filtering, realizes interference and noise suppression. In addition, the symbol sample set corresponding to the matched filtering pass domain can further correct the error of the QPSK demodulation data, and the error rate performance is improved.

Disclosure of Invention

The invention aims to solve the problem that QPSK signals are seriously polluted by same frequency interference and noise. The QPSK signal demodulation method based on time-frequency analysis and convolutional neural network includes the following steps:

(1) receiving signals at a receiving end, and intercepting the received signals;

(2) performing time-frequency analysis on the intercepted signals to generate a time-frequency distribution graph, and preprocessing the generated time-frequency distribution graph;

(3) inputting the preprocessed time-frequency distribution graph into a trained convolutional neural network for classification, and determining a filtering pass domain corresponding to a received signal according to a classification result;

(4) generating a time-varying filtering template according to the time-frequency distribution map of the corresponding category, and performing time-varying filtering on the received signal to finally obtain a demodulation symbol;

(5) and further calculating Hamming distances between the demodulation symbols and 4 symbol samples corresponding to the received signal time-frequency distribution diagram, and correcting the demodulation symbols according to the symbol samples corresponding to the minimum Hamming distance.

Further, the signal truncation described in step (1) has a length of 4 symbol periods.

Further, the pretreatment process in the step (2) comprises the following steps: and copying the images in the first half period of the time of each time-frequency distribution graph to the tail part of the time-frequency distribution graph on the time dimension, so that the time dimension of each image is extended from 4 symbol periods to 6 symbol periods.

Further, the training method of the trained convolutional neural network in the step (3) includes the following steps:

(3.1) generating n time-frequency distribution maps according to I, Q paths of all possible transmitted data of QPSK signals, classifying the data to be transmitted with the same time-frequency distribution maps into one class, and generating n' class time-frequency distribution maps together;

(3.2) preprocessing the n' class time-frequency distribution graph of the classified QPSK signals;

and (3.3) training the neural network, and taking the m samples of each class of the preprocessed n' class time-frequency distribution graph as the input of the convolutional neural network to obtain the optimal weight of the neural network.

Preferably, n time-frequency distribution graphs are generated in the step (3.1), wherein n is 64.

Preferably, the step (3.1) of generating an n 'class time-frequency distribution graph is performed, wherein n' has a value of 16.

Further, the pretreatment process of step (3.2) comprises the following steps: and copying the images in the first half period of the time of each time-frequency distribution graph to the tail part of the time-frequency distribution graph on the time dimension, so that the time dimension of each image is extended from 4 symbol periods to 6 symbol periods.

Preferably, the value of m in the step (3.3) satisfies that m is more than or equal to 100.

Further, the m samples in step (3.3) satisfy that all the samples are pure time-frequency distribution graphs without interference and noise, and each time-frequency distribution graph after preprocessing is copied m times to obtain the pure time-frequency distribution graph.

The invention has the beneficial effects that:

because the energy concentration areas of different user signals on a time-frequency plane are different, the invention provides a QPSK signal demodulation method based on time-frequency analysis and a convolutional neural network, and a filtering pass-domain which accords with the time-frequency distribution characteristics of useful signals is accurately obtained from signals polluted by interference and noise, thereby realizing the function of an error correcting code to a certain degree, effectively inhibiting same-frequency interference, improving the signal-to-interference-and-noise ratio performance of a receiving end and reducing the bit error rate.

Drawings

FIG. 1 is a flow chart of the present invention.

Fig. 2 is a 16-class time-frequency distribution diagram of a QPSK signal; wherein FIG. 2a represents signal 000001,001111,110000,111110; FIG. 2b shows signal 000010,010111,101000,111101; FIG. 2c shows signal 000011,011111,100000,111100; FIG. 2d shows signal 000100,011000,100111,111011; FIG. 2e shows signal 000101,010000,101111,111010; FIG. 2f shows signal 000110,001000,110111,111001; FIG. 2g shows signal 000111,111000,111111,000000; FIG. 2h shows signal 001101,010001,101110,110010; FIG. 2i shows signal 001001,001110,110001,110110; FIG. 2j shows signal 001001,001110,110001,110110; FIG. 2k shows signal 001011,011110,100001,110100; FIG. 2l shows signal 001100,011001,100110,110011; FIG. 2m shows signal 010010,010101,101010,101101; FIG. 2n shows signal 010011,011101,100010,101100; FIG. 2o shows signal 010100,011010,100101,101011; FIG. 2p shows signal 011011,011100,100011,100100;

FIG. 3 is a CNN classification result for a time-frequency distribution graph without preprocessing;

FIG. 4 is a time-frequency distribution graph after three types of error classification sample preprocessing; wherein FIG. 4a represents signal 000011,011111,100000,111100; FIG. 4b shows signal 000110,001000,110111,111001; FIG. 4c shows signal 001001,001110,110001,110110;

fig. 5 shows the result of CNN classification of the preprocessed time-frequency distribution map.

Detailed Description