KR101948568B1 - Classification method of modulation based on deep learning - Google Patents

Abstract

The present invention relates to a modulation scheme classification method for predicting a modulation scheme of a received signal. A deep learning based modulation scheme classification method according to the present invention includes the steps of: (a) extracting a plurality of feature values from signal data based on histogram analysis; (b) calculating input data of a deep learning algorithm based on the extracted plurality of feature values; And (c) classifying the modulation scheme of the signal data by performing the deep learning algorithm.

Description

[0001] CLASSIFICATION METHOD OF MODULATION BASED ON DEEP LEARNING [0002]

The present invention relates to a modulation scheme classification method for predicting a modulation scheme of a received signal.

There are two conventional techniques for recognizing a modulation scheme. The first is to classify the modulation scheme in the direction of maximizing the likelihood function based on the statistical model of the received signal.

Although this method can perform modulation classification so as to have a minimum detection error statistically, there is a disadvantage in that a model error occurs and the performance is deteriorated when errors or channel characteristics are changed in a real environment. Assuming a model for various situations, the computational complexity of the algorithm for modulation classification becomes very complicated.

Another approach is to use features that characterize the modulation scheme from the data already obtained. These features are extracted from down-sampled data based on the received data and the estimated symbol rate.

Although this method is not a statistically optimized method, it is relatively easy to implement and performs based on actual data obtained, so that performance degradation due to errors between the theoretical model and actual data is small.

Typical examples of frequently used features are cumulants. In the prior art, a modulation recognition technique using a decision tree or support vector machine (SVM) based on the cumulant feature was used.

Specifically, the decision tree method is a method of classifying a modulation method by separating feature values based on a specific threshold value, and SVM is a method of learning a classification method based on features from various training data.

The most recent technique is to extract the cumulant as well as other features from the received data and apply it to the artificial neural network (ANN) technology to learn and classify the modulation scheme.

The problem with this conventional modulation scheme recognition technique is that the recognition accuracy is sufficiently high in a high SNR (dB) environment and is reliable and usable, but the recognition accuracy is significantly lowered in a low SNR (dB) environment.

However, it is difficult to expect a high SNR (dB) in a special environment requiring the technology. That is, in order to utilize the automatic modulation recognition technology for a special purpose, it should show a certain level of accuracy even in a low SNR (dB) environment. However, the prior art shows low accuracy in the environment, thereby reducing the reliability of the data using the perceived modulation scheme.

The present invention is directed to solving the above-mentioned problems and other problems. Another object of the present invention is to provide a deep-run-based modulation scheme classification method having a high accuracy even in a low SNR environment, by using a large number of various feature values and a deep-run algorithm.

According to an aspect of the present invention, there is provided a method for extracting a plurality of feature values from signal data based on histogram analysis. (b) calculating input data of a deep learning algorithm based on the extracted plurality of feature values; And (c) classifying the modulation scheme of the signal data by performing the deep learning algorithm. [7] According to another aspect of the present invention, there is provided a deep learning based modulation scheme classification method.

In the embodiment, the step (a) may include extracting the plurality of feature values based on a peak value of the histogram.

The step (b) may include calculating the input data by applying a normalization process having a predetermined average and variance values to the extracted plurality of feature values.

In an embodiment, the step (c) may include calculating an estimation probability for each modulation scheme using the deep learning algorithm, and classifying the modulation scheme of the signal data based on the estimation probability .

Effects of the deep learning based modulation method classification method according to the present invention will be described as follows.

It can be seen that at least one of the embodiments of the present invention exhibits high recognition accuracy at high SNR (dB) and at low SNR (dB). That is, the accuracy of the modulation scheme recognition at a low SNR (dB) can be increased to a high level compared with the prior art.

Further scope of applicability of the present invention will become apparent from the following detailed description. It should be understood, however, that the detailed description and specific examples, such as the preferred embodiments of the invention, are given by way of illustration only, since various changes and modifications within the spirit and scope of the invention will become apparent to those skilled in the art.

의 히스토그램이다.
도 4는 feature PR(peak to rms)의 히스토그램이다.
도 5는 본 발명에서 제안하는 딥 러닝 구조의 실시 예를 설명하기 위한 개념도이다.
도 6은 딥러닝 구조를 이루는 perceptron 구조를 설명하기 위한 개념도이다.
도 7은 본 발명을 평가하기 위한 시뮬레이션 데이터 생성 구조의 실시 예를 설명하기 위한 개념도이다.
도 8은 AWGN 환경에서 딥러닝 구조와 ANN, SVM의 변조 방식 인식 정확도를 정리한 표이다. 1 is a conceptual diagram for explaining an embodiment of a blind demodulation technique system structure.
2 is a conceptual diagram for explaining the structure of a modulation scheme recognition algorithm according to the present invention.
Figure 3

.
4 is a histogram of the feature PR (peak to rms).
5 is a conceptual diagram for explaining an embodiment of a deep learning structure proposed in the present invention.
6 is a conceptual diagram for explaining a perceptron structure forming a deep-running structure.
7 is a conceptual diagram for explaining an embodiment of a simulation data generation structure for evaluating the present invention.
8 is a table summarizing the deep running structure and the recognition accuracy of the modulation method of ANN and SVM in the AWGN environment.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings, wherein like or similar components are denoted by the same reference numerals, and redundant explanations thereof will be omitted. The suffix " module " and " part " for the components used in the following description are given or mixed in consideration of ease of specification, and do not have their own meaning or role. In the following description of the embodiments of the present invention, a detailed description of related arts will be omitted when it is determined that the gist of the embodiments disclosed herein may be blurred. It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory and are intended to provide further explanation of the invention as claimed. , ≪ / RTI > equivalents, and alternatives.

Terms including ordinals, such as first, second, etc., may be used to describe various elements, but the elements are not limited to these terms. The terms are used only for the purpose of distinguishing one component from another.

It is to be understood that when an element is referred to as being "connected" or "connected" to another element, it may be directly connected or connected to the other element, . On the other hand, when an element is referred to as being "directly connected" or "directly connected" to another element, it should be understood that there are no other elements in between.

The singular expressions include plural expressions unless the context clearly dictates otherwise.

In the present application, the terms "comprises", "having", and the like are used to specify that a feature, a number, a step, an operation, an element, a component, But do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, or combinations thereof.

TECHNICAL FIELD The present invention relates to a technology for recognizing an automatic modulation of a digital signal, and a method for predicting a modulation method of a received signal without prior information in wireless communication. In general communication, the receiver knows information about the received signal, but there are special cases where the information is unknown.

This special case is common in military communication situations. For example, a threat signal or a desired signal should be predicted and extracted from a signal received from an enemy without any information.

In such a situation, an automatic blind demodulation technique may be used.

1 is a conceptual diagram for explaining an embodiment of a blind demodulation technique system structure.

Referring to FIG. 1, an automatic blind demodulation technique includes receiving a signal, estimating a symbol rate, classifying a modulation scheme, performing blind demodulation or immediately performing blind demodulation, decoding a line, And then proceed to the extraction step.

The present invention corresponds to the modulation scheme classification step in the automatic blind demodulation process, and recognizes the modulation scheme using the signal data based on the given signal data and the estimated symbol rate.

Hereinafter, the present invention will be described in detail.

In order to increase the reliability of the demodulated final signal, it is necessary to increase the accuracy of the modulation scheme in a low SNR (dB) environment which is a problem of the prior art. To this end, the present invention proposes a technique of recognizing a modulation scheme through the algorithm shown in FIG.

2 is a conceptual diagram for explaining the structure of a modulation scheme recognition algorithm according to the present invention.

Referring to FIG. 2, the present invention uses a plurality of features more than the features used in the prior art, and uses a machine learning algorithm employing a deep learning structure to perform automatic modulation with high accuracy even in a low SNR (dB) Recognition technology can be implemented.

In the prior art, only about six to eight features are utilized in the automatic modulation method. On the other hand, in the present invention, more than 21 different features can be used as inputs.

These multiple features are selected from features that show high correlation with the modulation method through histogram analysis among various features.

FIGS. 3 and 4 are examples of a histogram for selecting such a feature, which is a feature extracted assuming a situation of 0 dB in an AWGN channel.

의 히스토그램이다. Figure 3

.

Referring to FIG. 3, it can be seen that the distribution is easy to distinguish between 2FSK and 4FSK.

4 is a histogram of the feature PR (peak to rms).

Referring to FIG. 4, it is easy to distinguish the remaining modulation schemes except 2FSK and 4FSK.

In order to extract 20 selected features from the signal data in this way, the signal data is primarily calculated based on the following equations (1) to (8) and processed into values for various components of the signal.

는 베이스밴드 수신신호 샘플의 실수부,

는 허수부를 나타낸다. here,

The real part of the baseband received signal sample,

Represents the imaginary part.

는 신호 샘플의 위상을 의미한다. here,

Means the phase of the signal sample.

Next, from the processed values based on the above equations, 20 feature values are calculated through the following equations (9) to (28).

Equation (9) means a power ratio between an imaginary part and a real part of a baseband received signal sample calculated by Equation (1).

는 신호 샘플의 위상을 의미하며, 수학식 10과 수학식 11은, 이렇게 구해진

와 계산에 이용하는 샘플의 개수 C와 신호 크기에 대한 특정한 임계값

를 통해 계산되는 베이스밴드 수신신호 샘플 위상의 통계값을 의미한다. As described above,

(10) and (11) represent the phases of the signal samples

And the number of samples C used in the calculation and a specific threshold value

Which is a statistical value of the baseband received signal sample phase calculated through the above-described method.

Here, Equation (12) denotes the standard deviation of the absolute value of the signal magnitude.

Here, Equation (13) means the ratio of the mixed-order moments.

Here, Equation (14) means an average value of the signal magnitudes.

Equation (15) represents the maximum power spectral density (PSD) of the standardized signal value magnitude obtained through Equation (3).

Here, Equations (16) to (22) mean various combinations of cumulants.

Here, Equation (23) denotes the kurtosis of the signal magnitude.

Equation 24 represents the asymmetry of the signal magnitude.

Here, Equation (25) represents an average value with respect to a peak value.

Equation (26) represents the average power ratio to the peak power.

Equation (27) represents the absolute value standard deviation of the normalized value of the signal.

Since each of the calculated features has a large difference in size and value range, the feature value f _n is used as an input to the deep learning structure so as to have an average of 0 and a variance of 1, Data is provided.

Next, the deep running structure proposed by the present invention will be described.

FIG. 5 is a conceptual view for explaining an embodiment of the deep learning structure proposed in the present invention, and FIG. 6 is a conceptual diagram for explaining a perceptron structure forming a deep learning structure.

The deep running structure is a multi-layered structure of the perceptron structure shown in FIG. Unlike conventional ANNs, which use a perceptron structure, unlike ANN, which is a single layer structure, the deep learning structure is designed to increase the number of perceptrons and to make a multi-layer structure to learn more complex environments.

Referring to FIG. 5, the deep learning structure proposed in the present invention has one input, output layer, three hidden layers, and can be composed of 20, 500, 200, 40, and 7 nodes, respectively.

The remaining layers except for the last layer can use ReLU as an activation function and softmax as an activation function of the last layer to output the estimation probability for each modulation method.

Learning of the deep learning structure can use a negative log-likelihood as a cost function, and an adaptive learning rate of 0.01 as the initial value and halving if the training error no longer decreases can be applied.

In addition, the weight of each layer can be optimized by using stochastic gradient descent, and the mini-batch size can be set to 20 at this time.

7 is a conceptual diagram for explaining an embodiment of a simulation data generation structure for evaluating the present invention.

Referring to FIG. 7, modulation symbols are generated for seven modulation schemes of BPSK, QPSK, 8PSK, 16QAM, 63QAM, 2FSK, and 4FSK, and modulated symbols are passed through a root raised cosine filter to generate signal samples. At this time, the oversampling ratio is set to 10 and the roll-off factor is set to 0.5.

Assuming that the generated signal is a transmission signal, the receiver adds the Gaussian noise to the transmission signal, uses the same root raised cosine filter, and then samples the reception signal at the same oversampling ratio.

Features were calculated from 200,000 samples of received signal and 40,000 feature vectors were used for training data and 1/4 of them were used for validation. The performance evaluation of the learned model was evaluated by the modulation classification error ratio using 10,000 feature vectors.

8 is a table summarizing the deep running structure and the recognition accuracy of the modulation method of ANN and SVM in the AWGN environment.

Specifically, it summarizes the deep running structure using the features proposed in the present invention in the AWGN environment and the accuracy of the recognition of the seven modulation methods according to the change of the SNR (dB) of the conventional ANN and SVM.

Referring to FIG. 8, when the accuracy of each technique is examined, it can be seen that the accuracy of ANN and SVM is high in the case of low SNR (dB) although the accuracy of SNR (dB) is observed.

On the other hand, it can be seen that the accuracy according to the present invention shows a high recognition accuracy even at a high SNR (dB) or at a low SNR (dB). In other words, it can be seen that the technique proposed by the present invention has raised the accuracy of modulation scheme recognition at a low SNR (dB) to a high level compared with the prior art.

Effects of the deep learning based modulation method classification method according to the present invention will be described as follows.

It can be seen that at least one of the embodiments of the present invention exhibits high recognition accuracy at high SNR (dB) and at low SNR (dB). That is, the accuracy of the modulation scheme recognition at a low SNR (dB) can be increased to a high level compared with the prior art.

The present invention described above can be embodied as computer-readable codes on a medium on which a program is recorded. The computer readable medium includes all kinds of recording devices in which data that can be read by a computer system is stored. Examples of the computer readable medium include a hard disk drive (HDD), a solid state disk (SSD), a silicon disk drive (SDD), a ROM, a RAM, a CD-ROM, a magnetic tape, a floppy disk, , And may also be implemented in the form of a carrier wave (e.g., transmission over the Internet). Also, the computer may include a control unit 180 of the terminal. Accordingly, the above description should not be construed in a limiting sense in all respects and should be considered illustrative. The scope of the present invention should be determined by rational interpretation of the appended claims, and all changes within the scope of equivalents of the present invention are included in the scope of the present invention.

Claims (4)

(a) extracting a plurality of feature values from signal data based on histogram analysis;
(b) calculating input data of a deep learning algorithm based on the extracted plurality of feature values; And
(c) classifying the modulation scheme of the signal data by performing the deep learning algorithm,
Wherein the plurality of feature values comprise:
A value for the different components of the signal data,
The power ratio of the imaginary and real components of the baseband received signal samples, the phase of the signal samples, the standard deviation of the signal magnitude absolute values, the ratio of the mixed order moments, the mean value of the signal magnitudes, and the PSD and a standard deviation of the absolute value of the standardized value of the signal, which is a predetermined number of values including the maximum value of the signal intensity, the amplitude of the signal amplitude, the asymmetry of the signal amplitude, the average value of the peak values, Wherein the depth learning based modulation method classification method comprises: The method according to claim 1,
The step (a)
And extracting the plurality of feature values based on a peak value of the histogram. The method according to claim 1,
The step (b)
And calculating the input data by applying a normalization process having a predetermined average and a variance value to the extracted plurality of feature values. The method according to claim 1,
The step (c)
Calculating an estimation probability for each modulation scheme using the depth learning algorithm and classifying the modulation scheme of the signal data based on the estimation probability.

Images