Background
Identification of communication signal modulation schemes is an important issue for signal processing research, and is widely applied to military and civil fields. With the rapid development of communication technology, the system and modulation pattern of communication signals become more complex and diversified, and the signal environment becomes increasingly dense, so that the conventional identification method and theory are difficult to adapt to the actual requirements, and the communication signals cannot be effectively identified. At present, common algorithms for identifying AM and 2ASK are as follows:
1. decision identification based on traditional expert characteristics such as signal instantaneous amplitude, decision threshold and the like;
2. identifying the features such as a constellation diagram/combined/blind equalized vector diagram;
the disadvantages of the above modulation identification algorithm are:
1) for the identification mode using traditional characteristics such as signal amplitude and the like, factors such as noise influence change and the like under the real environment are not considered;
2) for the identification mode based on the constellation diagram and the like, the trace frequency offset is difficult to be completely filtered, and the final accumulated frequency offset influence is not considered;
meanwhile, the existing method does not consider the situation that signals are lost when the ratio of the code rate to the sampling rate is unreasonable, and the problem of identification accuracy under low signal-to-noise ratio.
Disclosure of Invention
The invention aims to provide a method for realizing modulation and identification of 2ASK signals and AM signals by combining a neural network, which is used for solving the problems that in the prior art, the AM and 2ASK identification does not consider the factors of amplitude dynamic change, cumulative frequency offset influence, loss signals when the code rate is not matched with the sampling rate and the like, and the identification accuracy is low under the condition of low signal-to-noise ratio.
The invention solves the problems through the following technical scheme:
a method for realizing modulation recognition of 2ASK signals and AM signals by combining a neural network comprises the following steps:
step S100: carrying out IQ demodulation on a modulation signal received by a receiver to respectively obtain discrete data sequences of an I path and a Q path;
step S200: carrying out normalization processing on a complex IQ sequence formed by the path I and the path Q;
step S300: clustering IQ coordinate points which are continuous in time on a complex plane, wherein the clustering result comprises point cluster clustering;
step S400: rotating IQ points on a coordinate plane, and rotationally moving all IQ points near the point clusters by using the offset angle of each point cluster relative to a real-axis positive half shaft to offset frequency deviation;
step S500: converting the rotated IQ points into a signal density map; projecting the IQ points on the rotated coordinate plane into a density map, wherein the density map is darker in color at the position where the IQ points are denser;
step S600: and creating a convolutional neural network model as a signal classification and identification model, and classifying and identifying the density maps of the AM signal and the 2ASK signal, wherein the convolutional neural network model is obtained by training a training data set which is artificially generated according to the texture rule of the density maps of the AM and the 2ASK signals. The artificially generated training data set can be used for learning effective distribution characteristics by a neural network by randomly changing IQ data distribution and adding Gaussian white noise to artificially synthesize enough simulation training density maps.
Further, the step S100 is specifically:
respectively multiplying the received signals by a sine function and a cosine function, and respectively carrying out integral operation to obtain a discrete data sequence as follows: i is1,Q1,I2,Q2,I3,Q3…In,Qn(ii) a n is half the length of the sequence.
Further, the step S200 specifically includes:
will obtain I1,Q1,I2,Q2,I3,Q3…In,QnThe sequences are combined into a complex sequence with the length of n, I1+ j Q1, I2+ j Q2, I3+ j Q3 … In + j Qn, and the complex sequence is processed by adopting maximum value normalization or mean value normalization, wherein the formula of the maximum value normalization is as follows:
each bit of complex IQ sequence is divided by the maximum complex modulus length in the sequence;
the formula of the mean value normalization is as follows:
each bit of complex IQ sequence is complex-divided by the average complex modulus length of the sequence.
Further, the step S300 specifically includes:
for several IQ points A, B, C, … … and M adjacent in continuous time, firstly, calculating the vector formed by the point A and the origin of coordinates
Then calculating the vector formed by the point B and the origin of coordinates
If it is
If the included angle of the two vectors is smaller than the threshold value, putting the point A, B into the same point cluster; then, the vector formed by the point C and the origin of coordinates is calculated continuously
And vector
If the included angle is still smaller than the threshold value, the point C is also placed into the same point cluster, and other IQ points of the same point cluster are continuously searched according to the same method; completing the searching of the clustering class of the point until one IQ point is found not to meet the threshold limit;
and continuously searching the next point cluster by the same method until all IQ sample points are calculated and IQ points of all point clusters are found and put into the same point cluster.
Further, the step S400 is to aggregate β and β for two adjacent points
Firstly, the average coordinate position of all IQ points in the β point clustering class is calculated, and the vector formed by the average coordinate position and the coordinate origin is calculated
An included angle omega with a real shaft positive half shaft; then will follow the pointLast IQ Point-to-Point Cluster in Cluster β
All IQ points between the last IQ points in the IQ graph rotate clockwise by an omega angle; at this time, the IQ point is moved to the vicinity of the real-axis positive half shaft;
after the same operation is sequentially carried out on all the point cluster clusters and the IQ points, all the IQ sample points are rotationally moved to a positive half shaft of a real coordinate axis.
Preferably, step S600 is specifically: taking as input a scattergram whose density distribution plane size is 50 × 50 (or other values), the 1 st layer convolution uses 3 × 3 convolution kernels, 16 feature maps extract input features, the output is subjected to ReLU activation and then pooled 3 × 3, the 2 nd layer convolution uses 3 × 3 convolution kernels, 16 feature maps outputs 3 × 3 pooled after ReLU activation, the 3 rd layer convolution uses 3 × 3 convolution kernels, 32 feature maps outputs 3 × 3 pooled after ReLU activation, the output data subjected to 3 layer convolution pooled further passes through a Dropout layer, and finally output recognition results through a full connection layer of 32 neurons and a Sigmoid layer of 2 neurons, the Sigmoid layer outputs a vector with a length of 2, each element in the vector represents the probability that the input is recognized as the class, and assuming that the output vector in this example is (0.1, 0.9), the probability that the sample is identified as the second class by the neural network is the highest, and is 90%, which is the identification result of the neural network.
Compared with the prior art, the invention has the following advantages and beneficial effects:
(1) the invention firstly utilizes a characteristic extraction algorithm to preprocess signals, learns the relevant characteristics of the signals by utilizing the high nonlinear mapping capability and the strong pattern recognition capability of a neural network, recognizes the signals and automatically optimizes the recognition result, because the method based on the effective data cumulant can resist the noise data interference to a certain degree, the frequency deviation is effectively counteracted by the rotary motion, the effective data capturing and clustering is not influenced by the code rate sampling rate, and the invention can effectively complete the recognition of the 2ASK signals and AM signal modulation modes under the non-ideal conditions of low signal-to-noise ratio, frequency deviation of the received signals, incomplete signals and the like.
(2) The invention is based on the technology of neural network image identification, identifies the brand new characteristics of AM and 2ASK signal density graphs and then realizes modulation, the influence of the actual complex communication environment on the signals is reflected on the signal density graphs to cause distortion such as image stretching, displacement, rotation, scaling and the like, but the characteristics of basic outline texture and the like are reserved, the neural network-based image identification can well learn the basic texture characteristics without the influence of the distortion, and the modulation identification of the AM and 2ASK signals can be realized under the actual communication conditions of low signal-to-noise ratio, under-sampling, over-sampling and the like.
Example 1:
with reference to fig. 1, a method for implementing modulation identification of 2ASK signals and AM signals in combination with a neural network specifically includes:
1. IQ demodulation is carried out on the modulation signal received by the receiver, and discrete data sequences of an I path and a Q path are respectively obtained: specifically, the received carrier signal is multiplied by a sine function and a cosine function respectively, and integration operations are performed respectively, in this example, for convenience of demonstration and explanation, it is assumed that an IQ data sequence with a length of 2 × n is obtained as I1,Q1,I2,Q2,I3,Q3…In,Qn;
2. Normalizing the complex IQ sequence formed by the path I and the path Q, specifically:
will obtain I1,Q1,I2,Q2,I3,Q3…In,QnThe sequences are combined into a complex sequence of length n: i is1+j*Q1,I2+j*Q2,I3+j*Q3…In+j*QnAnd performing modulo length normalization on each complex number in the sequence, wherein the normalization can be performed using maximum normalization, and the maximum normalization formula is:
each complex number in the sequence is divided by the largest complex number in the sequence modulo length or normalized by mean, the mean normalization formula being:
mean complex modulus length of sequence
Taking the maximum normalization as an example, first find the complex number with the largest length in the complex number sequence, and let the complex number be Ik+j*QkThe length of the plural module is LkThe normalized complex sequence is: (I)1+j*Q1)/Lk,(I2+j*Q2,)/Lk,(I3+j*Q3)/Lk…(In+j*Qn)/Lk。
3. Clustering IQ coordinate points in time continuity on a complex plane
For three IQ points A, B and C adjacent in continuous time, firstly, calculating a vector formed by the point A and a coordinate origin
Then calculating the vector formed by the point B and the origin of coordinates
If it is
If the included angle between the two vectors is smaller than the threshold value, the two vectors are close enough, logically represent the 'code origin' of the same code element, and then the point A, B is put into the same point cluster; then, the vector formed by the point C and the origin of coordinates is calculated continuously
And vector
If the included angle is still smaller than the threshold value, the point C is also placed into the same point cluster; if the point cluster does not meet the threshold limit, the point cluster is searched completely, the next point cluster is searched by the same method, and so on until all IQ sample points are calculated and all point cluster clusters are found.
4. Rotating IQ points on the coordinate plane according to the clustered point groups to offset frequency deviation
For two adjacent clustering point groups β and
firstly, the average coordinate position of all IQ points in the point group is calculated β, and the vector formed by the average coordinate and the coordinate origin is calculated
Angle omega with the positive half axis of the real axis, then from the last IQ point in the point cloud β to the point cloud
All IQ points between the last IQ points in the IQ graph rotate clockwise by an omega angle; at this time, the IQ point is moved to the vicinity of the real-axis positive half shaft; after the operations are sequentially performed on all the point groups and the IQ points, all the IQ sample points are rotated and moved to the positive half axis of the real coordinate axis.
5. Generation of a signal density map from rotated signal data
And projecting the IQ points on the coordinate plane after rotation into a density map, wherein the density map is darker in color at places where the IQ points are denser. The IQ points on the density map of the AM signal will be in a band shape, reflecting the range of continuous amplitude interval conversion, as shown in fig. 2 (b); the IQ points on the 2ASK signal density map will show a distinct 2-cluster distribution, reflecting the clustering of two discrete amplitudes, as shown in fig. 2 (a);
6. a convolutional neural network (convolutional neural network) classifier model is designed for classifying two signals, namely 2ASK and AM.
Specifically, as shown in fig. 3, with a 50 × 50 scatter diagram as input, the 1 st layer convolution uses 3 × 3 convolution kernels, 16 feature maps extract input features, the output is subjected to 3 × 3 pooling after ReLU activation, the 2 nd layer convolution uses 3 × 3 convolution kernels, 16 feature maps, the output is subjected to 3 × 3 pooling after ReLU activation, the 3 rd layer convolution uses 3 × 3 convolution kernels, 32 feature maps, the output is subjected to 3 × 3 pooling after ReLU activation, the output data after 3 layers of pooling is further subjected to Dropout, finally, an identification result is output through a full connection layer of 32 neurons and a Sigmoid layer of 2 neurons, the Sigmoid layer outputs a vector with a length of 2, each element in the vector represents a probability that the input is identified as the class, and assuming that the output vector in this example is (0.1, 0.9), the sample is identified as the second highest probability by the neural network, 90% as the recognition result of the neural network.
Although the present invention has been described herein with reference to the illustrated embodiments thereof, which are intended to be preferred embodiments of the present invention, it is to be understood that the invention is not limited thereto, and that numerous other modifications and embodiments can be devised by those skilled in the art that will fall within the spirit and scope of the principles of this disclosure.