Adaptive demodulation method of digital oscilloscope

Abstract

The invention discloses a self-adaptive demodulation method of a digital oscilloscope, which comprises the following steps of: firstly, training a GRU neural network in a GPU in an off-line manner and generating a coefficient matrix, then loading the coefficient matrix to an FPGA deployed with the GRU neural network, then acquiring a signal to be demodulated through an acquisition system of the digital oscilloscope, multiplying the sampling data by the coefficient matrix, and finally acquiring the signal to be demodulated. Further determining a modulation classification corresponding to the signal to be demodulated; and finally, selecting a corresponding demodulation mode according to a modulation classification result of the to-be-demodulated signal, completing adaptive demodulation of the to-be-demodulated signal, and if a classification error phenomenon occurs in a demodulation process, performing GPU online incremental learning so as to provide demodulation accuracy.

Description

Technical Field

[0001] The invention belongs to the technical field of digital oscilloscopes, and more specifically, relates to an adaptive demodulation method for a digital oscilloscope. Background Art

[0002] In modern communication systems, signals experience various changes and interferences. Channel conditions, such as multipath effects, signal attenuation, interference, etc., will affect the reception of signals. With the development of wireless communication technology, more complex signal processing techniques are needed to cope with the diversity of signals and the uncertainty of channels. The complexity and dynamics of channels are also increasing. In early communication systems, fixed modulation and demodulation schemes were usually adopted, which means that even if the channel conditions change, the modulation method and demodulation strategy of the system remain unchanged. However, fixed demodulation schemes will lead to decreased communication efficiency and increased error rates when faced with problems such as unstable signal strength, frequency selective fading, and time-varying interference.

[0003] Traditional analog and digital demodulation technologies are relatively simple to design, but are often affected by noise and interference. In addition, traditional demodulation usually does not support multiplexing technology, and due to channel conditions or the accuracy of the estimation algorithm, there is a deviation between the estimated parameters and the actual parameters, which will affect the output results of the signal demodulation. These problems are particularly prominent in non-cooperative communication environments because there is no pre-defined protocol to assist in adapting to these changes. It is necessary to estimate signal characteristics and perform blind processing and demodulation on the signal. Cooperative communication provides some mechanisms to improve signal quality, but it also requires the system to have higher flexibility and adaptability, and requires users to manually change the parameter settings of the signal. It can be seen that traditional demodulation technology can no longer meet current needs. In order to process modulated signals more flexibly and universally and improve the reliability and efficiency of communications, a demodulation method that can automatically identify and adapt to these changes is needed. Summary of the invention

[0004] The purpose of the present invention is to overcome the shortcomings of the prior art and provide an adaptive demodulation method for a digital oscilloscope, which selects the most appropriate demodulation scheme according to the real-time information of the channel, thereby improving the spectrum utilization, reducing interference, enhancing the reliability of communication, and effectively improving the blind demodulation quality of the modulated signal and the accuracy of parameter estimation.

[0005] To achieve the above-mentioned object of the invention, the present invention provides an adaptive demodulation method for a digital oscilloscope, characterized in that it comprises the following steps:

[0006] (1) Offline training of a GRU neural network with M neurons in the GPU and generating a coefficient matrix P;

[0007] (2) Deploy the GRU neural network in the FPGA and load the coefficient matrix P;

[0008] (3) Sample the signal to be demodulated and input it to the FPGA;

[0009] (4) estimating the parameters of the sampled signal through a parameter estimation module;

[0010] (5) Determine the modulation classification of the signal to be demodulated;

[0011] (6) Select the corresponding modulation method and demodulate according to the modulation classification result of the signal to be demodulated.

[0012] The object of the invention of the present invention is achieved in this way:

[0013] The invention discloses an adaptive demodulation method for a digital oscilloscope. The method comprises the following steps: firstly training a GRU neural network offline in a GPU and generating a coefficient matrix, then loading the coefficient matrix into an FPGA on which the GRU neural network is deployed, then collecting a signal to be demodulated through an acquisition system of the digital oscilloscope, multiplying the sampled data by the coefficient matrix, and then determining a modulation classification corresponding to the signal to be demodulated; finally, selecting a corresponding modulation mode according to a modulation classification result of the signal to be demodulated, completing adaptive demodulation of the signal to be demodulated, and performing GPU online incremental learning if a classification error occurs during the demodulation process, thereby improving the accuracy of demodulation.

[0014] At the same time, the adaptive demodulation method of a digital oscilloscope of the present invention also has the following beneficial effects:

[0015] (1) The present invention solves the problem of low accuracy of signal demodulation and parameter estimation when the signal format and parameters are uncertain. In addition, the adaptive demodulation technology can not only handle the diversity of signals and the uncertainty of channels, but also adjust the demodulation strategy in real time to optimize the performance of the communication link.

[0016] (2) The present invention can realize adaptive demodulation more flexibly, achieving the effect of initial offline training, followed by online incremental learning and demodulation. The system can gradually increase the types of processable signals through online incremental learning; in the case of classification errors, it can also use online incremental learning to extract the current signal characteristics, fine-tune parameters, and reduce the probability of subsequent classification errors of the same type. Ultimately, the convenience, practicality, accuracy and reliability are continuously improved with the increase of usage time.

[0017] (3) The present invention can be applied to cooperative communication and non-cooperative communication fields, does not require users to customize the demodulation parameter settings, and can achieve better human-computer interaction performance. It can intelligently demodulate different modulated signal inputs to improve the reliability and efficiency of communication.

[0018] (4) The present invention completes an adaptive demodulation system for modulated signals under the conditions of large amounts of data training and storage resources in the GPU, meets the needs of modern measurement technology for intelligent technology, and has certain guiding significance for the field of modulated signal communication.

[0019] (5) The present invention can demodulate both analog modulation signals and digital modulation signals;

[0020] (6) The present invention trains the neural network by GPU and finally deploys the neural network to FPGA. Compared with the neural network implemented by software, the processing speed of FPGA is faster, up to 100 times faster, and has stronger real-time performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a principle diagram of an adaptive demodulation method of a digital oscilloscope of the present invention;

[0022] Figure 2 It is a block diagram of analog demodulation of sampled signals;

[0023] Figure 3 It is a block diagram of digital demodulation of sampled signals;

[0024] Figure 4 It is a schematic diagram of error vector magnitude EVM;

[0025] Figure 5 This is the 16-QAM demodulation result diagram. DETAILED DESCRIPTION

[0026] The specific implementation of the present invention is described below in conjunction with the accompanying drawings so that those skilled in the art can better understand the present invention. It should be noted that in the following description, when the detailed description of known functions and designs may dilute the main content of the present invention, these descriptions will be omitted here.

[0027] Example

[0028] In this embodiment, if Figure 1 As shown, the present invention provides an adaptive demodulation method for a digital oscilloscope, comprising the following steps:

[0029] (1) Offline training of a GRU neural network with M neurons in the GPU. The GRU neural network completes the classification of N types of modulated signals in each batch, thereby outputting a coefficient matrix P of size M×N, where the element p in the i-th row and j-th column of the matrix P is ij is the neural weight coefficient of the i-th neuron for the j-th modulation type;

[0030] In this embodiment, the number of neurons in the GRU neural network is assumed to be M=100, and 20 types of modulation signals are classified in each batch.

[0031] (2) GPU loads the signal template library, which consists of n known modulated signals with different formats, different parameters, and different noises, as well as Nn random signals. The signal template library is input into the trained GRU neural network to obtain the coefficient matrix P. The coefficient matrix P is then sent to the CPU, which is then deployed to the FPGA.

[0032] In this embodiment, the signal template library includes 11 modulation signals of known modulation formats, including 8 digital modulation signals (BPSK, 8PSK, QPSK, CPFSK, GFSK, 16-QAM, 64-QAM, 4-PAM) and 3 analog modulation signals (AM-DSB, AM-SSB, WBFM); 9 random signals are modulation signals of undetermined modulation types;

[0033] (3) Figure 1 As shown, the demodulated signal S(t) is input into a digital oscilloscope, sampled by ADC to obtain a sampling signal with a length of M, and then the sampling signal is input into FPGA;

[0034] In this embodiment, the signal to be demodulated may be an analog modulation signal and a digital modulation signal, wherein the analog modulation signal can only be a periodic signal, and its baseband signal is one of a sine wave, a square wave and a triangle wave.

[0035] (4) Estimate the parameters of the sampled signal through the parameter estimation module in the FPGA, including: carrier frequency f c , signal bandwidth B, symbol rate R, signal amplitude A, signal frequency f1 and roll-off factor α;

[0036] (5) FPGA multiplies the sampled signal of length M by the coefficient matrix P to obtain a 1*N probability matrix. Each element in the probability matrix represents the probability that the signal to be demodulated belongs to each modulation type signal in the signal template library. Then, the maximum probability is taken as the modulation classification corresponding to the signal to be demodulated.

[0037] (6) Selecting a corresponding modulation method according to the modulation classification result of the signal to be demodulated. If the signal to be demodulated is an analog modulation signal, proceed to step (7); if the signal to be demodulated is a digital modulation signal, proceed to step (8); otherwise, proceed to step (9);

[0038] (7) Analog demodulation of the sampled signal. The specific demodulation process is as follows: Figure 2 As shown;

[0039] (7.1) In the analog demodulation module, the sampling signal is mixed with the carrier frequency f generated by the local oscillator. cThe carrier is digitally down-converted, and then low-pass filtered with a bandwidth of B, thereby demodulating and obtaining a baseband signal X containing M points = {x1, x2, ..., x i ,…,x M}, x i Represents the quantized value of the i-th sampling point;

[0040] (7.2) Use the KNN algorithm to classify the demodulated baseband signal and obtain the waveform type of the baseband signal;

[0041] In this embodiment, the KNN algorithm has a sample signal library, which contains a sample sequence of complete waveforms of sine waves, square waves, and triangle waves. By calculating the Euclidean distance between the baseband signal and each sample signal, the minimum Euclidean distance is found, and the waveform of the baseband signal is classified into the sample signal corresponding to the minimum Euclidean distance.

[0042] (7.3), the CPU generates a reference waveform Y containing M points based on the signal amplitude A, signal frequency f1, and the waveform type obtained by KNN classification. i ,…,y M},y i represents the quantized value of the i-th reference point;

[0043] (7.4) Calculate the Manhattan distance between the demodulated baseband signal and the reference waveform:

[0044]

[0045] Curve similarity is used to measure the similarity between two or more curves. Manhattan distance is a commonly used curve similarity, which calculates the sum of the absolute differences between two vectors (or curves), taking into account the differences between dimensions, and is suitable for curve data with multiple dimensions.

[0046] (7.5) Compare the Manhattan distance MD with the set threshold. If MD is less than the set threshold, the demodulation and classification results are determined to be correct and the process ends; otherwise, proceed to step (7.6);

[0047] (7.6), GPU online incremental learning;

[0048] (7.6.1) Control the host computer to display the sampled signal and identify the modulation type of the sampled signal through manual retrieval;

[0049] (7.6.2) Determine whether the identified modulation type is a known modulation type in the signal template library. If the modulation type is not a known modulation type, add a known modulation signal to the signal template library, that is, the number of known modulation signals in the signal template library is changed from n to n+1; if the modulation type is a known modulation type, do not add a new modulation type;

[0050] (7.6.3) GPU reloads the signal template library and then inputs it into the trained GRU neural network to obtain a new coefficient matrix P'. The new coefficient matrix P' is then sent to the CPU and redeployed to the FPGA, replacing the original coefficient matrix P to complete GPU online incremental learning.

[0051] (8) Digital demodulation of the sampled signal. The specific demodulation process is as follows: Figure 3 As shown;

[0052] (8.1), the sampling signal and the carrier frequency are c The carrier is digitally mixed to obtain the sum frequency signal and the difference frequency signal, and then the difference frequency signal is retained by low-pass filtering with a bandwidth of B;

[0053] Generate the coefficient of the root raised cosine filter according to the roll-off factor α, and then use the root raised cosine filter to match filter the difference frequency signal to obtain the baseband signal;

[0054] The coefficients of the root raised cosine filter are related to the roll-off factor as follows:

[0055]

[0056] Among them, T R is the symbol period, which is the inverse of the symbol rate R; g R (t) is the continuous-time expression of the root raised cosine filter coefficient; in practical applications, it is also necessary to discretize and truncate it according to actual conditions.

[0057] (8.2) Resample the baseband signal so that the sampling rate is 4 times of R; then perform timing synchronization on the resampled signal, and then perform carrier synchronization to correct the carrier frequency deviation and carrier phase deviation, obtain two corrected I / Q baseband signals, and send them to the CPU;

[0058] (8.3), respectively obtain the binary sequence in the I / Q baseband signal, and then generate I ref / Q ref Two reference signals;

[0059] (8.4), calculate the root mean square of the error vector magnitude EVM RMS :

[0060]

[0061] Where E(i) is the vector difference of the i-th sampling point, I(i) and Q(i) are the quantized values ​​of the i-th sampling point in the I / Q baseband signal, and I ref (i) Q ref (i) is I ref / Q ref The quantized value of the i-th reference point in the reference signal;

[0062] In this embodiment, EVM is the most widely used parameter in vector signal analysis. Figure 4 As shown in the figure, EVM reflects the amplitude of the difference between the measured signal and the ideal signal vector, which can comprehensively reflect the quality of the received signal. EVM measurement has nothing to do with the specific modulation format. Defects in any link of the modulated signal will be reflected in EVM, so EVM is called the total error of vector signal analysis. The smaller the EVM, the smaller the probability of misjudgment of the transmitted signal, indicating that the result of system demodulation is correct.

[0063] (8.5), compare the error vector magnitude RMS E RMS With the set threshold value, if E RMS If it is less than the set threshold, the demodulation and classification results are judged to be correct and the processing ends; otherwise, go to step (8.6);

[0064] (8.6), GPU online incremental learning;

[0065] (8.6.1) Control the host computer to display the sampled signal and identify the modulation type of the sampled signal through manual retrieval;

[0066] (8.6.2) If the modulation type is not a known modulation type, add a known modulation signal to the signal template library, that is, the number of known modulation signals in the signal template library changes from n to n+1; if the modulation type is a known modulation type, do not add a new modulation type;

[0067] (8.6.3) GPU reloads the signal template library and then inputs it into the trained GRU neural network to obtain a new coefficient matrix P'. The new coefficient matrix P' is then sent to the CPU and redeployed to the FPGA, replacing the original coefficient matrix P to complete GPU online incremental learning.

[0068] (9) Online incremental learning of signals with unknown modulation types;

[0069] (9.1) Control the host computer to display the sampled signal and identify the modulation type of the sampled signal through manual retrieval;

[0070] (9.2) Add a known modulation signal to the signal template library, that is, the number of known modulation signals in the signal template library changes from n to n+1;

[0071] (9.3) The GPU reloads the signal template library and then inputs it into the trained GRU neural network to obtain a new coefficient matrix P'. The new coefficient matrix P' is then sent to the CPU and redeployed to the FPGA, replacing the original coefficient matrix P to complete GPU online incremental learning.

[0072] Example Simulation

[0073] In this embodiment, the 16-QAM modulated signal is input into the system, and after being sampled by the ADC, it is input into the FPGA. The FPGA estimates its carrier frequency f c , bandwidth B, symbol rate R and roll-off factor α are classified at the same time, and the classification result is 16-QAM, which belongs to digital modulation. Therefore, next, we use f c , B, R, α are digitally demodulated. The demodulation results are output to the CPU module. After comparison and evaluation by the CPU module, the constellation diagram of the 16-QAM demodulation result of the adaptive demodulation system is obtained as follows Figure 5 As shown in the figure, the horizontal axis is I, the vertical axis is Q, the center of the 16 "+" represents the position of the ideal 16-QAM constellation point, and the coordinates (i, q) of the surrounding points correspond to a pair of demodulated I / Q baseband signals. It can be seen from the figure that the EVM vector deviation value is small and the demodulation is successful.

[0074] Although the above describes the illustrative specific embodiments of the present invention to facilitate those skilled in the art to understand the present invention, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations using the concept of the present invention are protected.

Claims

1. An adaptive demodulation method for a digital oscilloscope, characterized in that: The following steps are involved: (1) Offline training of a GRU neural network with M neurons in the GPU. The GRU neural network completes the classification of N types of modulated signals in each batch, thereby outputting a coefficient matrix P of size M×N, where the element p in the i-th row and j-th column of the matrix P is ij is the neural weight coefficient of the i-th neuron for the j-th modulation type; (2) GPU loads the signal template library, which consists of n known modulated signals with different formats, different parameters, and different noises, as well as Nn random signals. The signal template library is input into the trained GRU neural network to obtain the coefficient matrix P. The coefficient matrix P is then sent to the CPU, which is then deployed to the FPGA. (3) Input the signal to be demodulated into a digital oscilloscope, obtain a sampling signal with a length of M through ADC sampling, and then input the sampling signal into the FPGA; (4) Estimate the parameters of the sampled signal through the parameter estimation module in the FPGA, including: carrier frequency f c , signal bandwidth B, symbol rate R, signal amplitude A, signal frequency f1 and roll-off factor α; (5) FPGA multiplies the sampled signal of length M by the coefficient matrix P to obtain a 1*N probability matrix. Each element in the probability matrix is ​​the probability that the signal to be demodulated belongs to each modulation type signal in the signal template library. Then, the maximum probability is taken as the modulation classification corresponding to the signal to be demodulated. (6) Select the corresponding modulation method according to the modulation classification of the signal to be demodulated. If the signal to be demodulated is an analog modulation signal, proceed to step (7); if the signal to be demodulated is a digital modulation signal, proceed to step (8); otherwise, proceed to step (9); (7) Analog demodulation of the sampled signal; (7.1), the analog demodulation module is based on the carrier frequency f of the sampling signal c The signal bandwidth B is digitally down-converted and low-pass filtered to obtain a baseband signal X containing M points = {x1, x2, ..., x i ,…,x M }, x i Represents the amplitude of the i-th sampling point; (7.2) Use the KNN algorithm to classify the demodulated baseband signal and obtain the waveform type of the baseband signal; (7.3), the CPU generates a reference waveform Y containing M points based on the signal amplitude A, signal frequency f1, and the waveform type obtained by KNN classification. i ,…,y M },y i represents the amplitude of the i-th reference point; (7.4) Calculate the Manhattan distance between the demodulated baseband signal and the reference waveform: (7.5) Compare the Manhattan distance MD with the set threshold. If MD is less than the set threshold, the demodulation and classification results are determined to be correct and the process ends; otherwise, proceed to step (7.6); (7.6), GPU online incremental learning; (7.6.1) Control the host computer to display the sampled signal and identify the modulation type of the sampled signal through manual retrieval; (7.6.2) Determine whether the identified modulation type is a known modulation type in the signal template library. If the modulation type is not a known modulation type, add a known modulation signal to the signal template library, that is, the number of known modulation signals in the signal template library is changed from n to n+1; if the modulation type is a known modulation type, do not add a new modulation type; (7.6.3) GPU reloads the signal template library and then inputs it into the trained GRU neural network to obtain a new coefficient matrix P'. The new coefficient matrix P' is then sent to the CPU and redeployed to the FPGA, replacing the original coefficient matrix P to complete GPU online incremental learning. (8) Digital demodulation of the sampled signal; (8.1), the sampling signal and the carrier frequency are c The carrier is digitally mixed to obtain the sum frequency signal and the difference frequency signal, and then the difference frequency signal is retained by low-pass filtering with a bandwidth of B; Generate the coefficient of the root raised cosine filter according to the roll-off factor α, and then use the root raised cosine filter to match filter the difference frequency signal to obtain the baseband signal; (8.2) Resample the baseband signal so that the sampling rate is 4 times of R; then perform timing synchronization on the resampled signal, and then perform carrier synchronization to correct the carrier frequency deviation and carrier phase deviation, obtain two corrected I / Q baseband signals, and send them to the CPU; (8.3), respectively obtain the binary sequence in the I / Q baseband signal, and then generate I ref / Q ref Two reference signals; (8.4), calculate the root mean square of the error vector magnitude EVM RMS : Where E(i) is the vector difference of the i-th sampling point, I(i) and Q(i) are the quantized values ​​of the i-th sampling point in the I / Q baseband signal, and I ref (i) Q ref (i) is I ref / Q ref The quantized value of the i-th reference point in the reference signal; (8.5), compare the error vector magnitude RMS E RMS With the set threshold value, if E RMS If it is less than the set threshold, the demodulation and classification results are judged to be correct and the processing ends; otherwise, go to step (8.6); (8.6), GPU online incremental learning; (8.6.1) Control the host computer to display the sampled signal and identify the modulation type of the sampled signal through manual retrieval; (8.6.2) Determine whether the identified modulation type is a known modulation type in the signal template library. If the modulation type is not a known modulation type, add a known modulation signal to the signal template library, that is, the number of known modulation signals in the signal template library is changed from n to n+1; if the modulation type is a known modulation type, do not add a new modulation type; (8.6.3) GPU reloads the signal template library and then inputs it into the trained GRU neural network to obtain a new coefficient matrix P'. The new coefficient matrix P' is then sent to the CPU and redeployed to the FPGA, replacing the original coefficient matrix P to complete GPU online incremental learning. (9) Online incremental learning of signals with unknown modulation types; (9.1) Control the host computer to display the sampled signal and identify the modulation type of the sampled signal through manual retrieval; (9.2) Add a known modulation signal to the signal template library, that is, the number of known modulation signals in the signal template library changes from n to n+1; (9.3) The GPU reloads the signal template library and then inputs it into the trained GRU neural network to obtain a new coefficient matrix P'. The new coefficient matrix P' is then sent to the CPU and redeployed to the FPGA, replacing the original coefficient matrix P to complete GPU online incremental learning.

2. The adaptive demodulation method of a digital oscilloscope according to claim 1, characterized in that: The signal to be demodulated includes an analog modulation signal and a digital modulation signal, wherein the analog modulation signal can only be a periodic signal, and its baseband signal is one of a sine wave, a square wave and a triangle wave.

Images