A Deep Learning Shortwave Signal Recognition Method Based on Ionospheric Parameter Compensation

Abstract

The present invention discloses a deep learning shortwave signal recognition method based on ionospheric parameter compensation, which includes the following steps: Step 1, construct an overall neural network; Step 2, process the data; Step 3, add ionospheric parameters and prior knowledge; Step 4, perform forward gradient operation and backward gradient derivation to update the parameters; Step 5, obtain the recognition result and obtain the category with the highest probability. The method disclosed by the present invention uses a complex convolution module to extract the features of the I and Q channels, and uses an LSTM module to extract the time series features, fully exploiting and extracting the relevant information of the I and Q channels and the time series to improve the recognition accuracy.

Description

Technical Field

[0001] The present invention belongs to the field of short-wave signal recognition, and particularly relates to a deep learning short-wave signal recognition method based on ionospheric parameter compensation in this field. Background Art

[0002] Nowadays, many methods are considering communication signals, and the research on short-wave signals is relatively less and more basic. Basically, traditional feature extraction methods are still used to identify short-wave signals. Now, deep learning has been widely applied in various fields, further improving the short-wave radiation recognition efficiency. Specific Emitter Identification (SEI) is a technology that identifies an individual by analyzing the unique radiofrequency fingerprinting (RFF) information contained in the electromagnetic waves radiated by the radiation source. It is one of the most difficult tasks in modern electronic intelligence systems and electronic warfare systems. Thus, the individual identification of radiation sources is both challenging and full of opportunities.

[0003] Ding Lida disclosed a method for identifying radiation source individuals based on manifold learning supervised dimensionality reduction and convolutional neural network in "Individual Identification of Communication Radiation Sources Based on Deep Learning". Its composition is to estimate the bispectrum of the steady-state signal of the communication radiation source, use small-sample learning to supervised reduce the dimension of the original bispectrum, and weaken the influence of cross terms, and then construct and optimize a convolutional neural network to identify the compressed bispectrum. Compared with traditional bispectrum-based radiation source individual identification methods, such as local bispectrum integration, selective bispectrum, etc., using a convolutional neural network significantly improves the identification performance. The disadvantage is that this method mainly studies traditional communication signals and does not consider what impact different propagation channels will have on the signal. For short waves, the ionospheric channel is unpredictable and requires certain prior knowledge to improve its identification accuracy. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to overcome the shortcoming that the radiation source identification in the prior art does not actively consider the relevant characteristics of the channel, and provide a deep learning short-wave signal recognition method based on ionospheric parameter compensation.

[0005] The present invention adopts the following technical solutions:

[0006] A deep learning short-wave signal recognition method based on ionospheric parameter compensation, the improvement lies in that it includes the following steps:

[0007] Step 1, construct an overall neural network:

[0008] First, a Transformer neural network is used to construct an ionospheric parameter compensation network. Then, a complex neural network is used to perform corresponding feature correlation on the I and Q signals of the signal, and an LSTM neural network is used to analyze the context correlation before and after the I and Q time series. Finally, a fully connected layer is used to identify and classify the extracted high-dimensional features;

[0009] Step 2: Process the data:

[0010] Perform complex normalization on the data:

[0011]

[0012] In the above formula, real(x) represents the real part and img(x) represents the imaginary part;

[0013] Step 3: Add ionospheric parameters and prior knowledge:

[0014] Perform smooth curve fitting on the detected ionospheric parameters to obtain a polynomial to characterize the variation of ionospheric parameters on the day:

[0015] Variable = ∑a i x i + b, i = 1, 2,..., n

[0016] In the above formula, a is the polynomial coefficient and b is the polynomial constant;

[0017] For the parameter fitting of the ionosphere, the variance λ of the fitting polynomial var is less than the measurement standard threshold α,

[0018] After that, the ionospheric parameters, the real data and the ideal data collected are input into the Transformer neural network, and the multi-head attention mechanism in the Transformer neural network is used:

[0019] V (Q,K,V) = X · W (Q,K,V)

[0020] Multiply the data X by the weight W to generate the query Q, the key K and the value V, and then obtain a new matrix Z through the following transformation:

[0021]

[0022] In the above formula, d k represents the dimension of the key value K;

[0023] Step 4: Perform forward gradient operation and backward gradient derivation to update the parameters:

[0024] The data passes through the network and first undergoes a linear transformation of the matrix:

[0025] z = w T x + b

[0026] a = σ(z)

[0027] In the above formula, w is the weight matrix, b is the bias, and σ is the activation function;

[0028] After that, the cross-entropy loss function is used:

[0029]

[0030] In the above formula, is the predicted label, y is the true label. Through the deviation between the prediction and the truth, and through the reverse gradient operation, the predicted value is then made to approach the true value;

[0031] First, calculate the gradient of the weight w:

[0032]

[0033] And calculate the gradient of the bias:

[0034]

[0035] In the above formula, is the element weight value of the j-th row and k-th column of the l-th layer convolutional kernel; C is the obtained cost function, is the output value of the previous layer; is the error between the prediction and the truth of the previous layer and this layer; is the bias value corresponding to the j-th row of the l-th layer convolutional kernel; is the bias value of the j-th row and k-th column of the l-th layer convolutional kernel; is the value of the data output after passing through the j-th row of the l-th layer convolutional kernel;

[0036] Through the gradient operation, the corresponding weight and bias parameters are obtained. After that, repeat steps 1 - 4 to obtain the weights that can achieve the minimum loss value, and the resulting weights form the trained network;

[0037] Step 5, obtain the recognition result, and obtain the category with the highest probability:

[0038] Use the untrained data. After processing the data format into the same format as the training data, directly input it into the trained network to obtain the output probability vector [β 1 ', β 2 ',..., β n ']. Take the maximum value among the probabilities as the recognition result to obtain the final result of the network's recognition of this data.

[0039] Furthermore, in step 1, the structure of the network includes 6 convolutional modules, and each module includes a complex convolutional layer, a batch normalization layer, and an activation function.

[0040] The formula for complex convolution operation is as follows:

[0041] W*h = (A*x - B*y) + i(B*x + A*y)

[0042] First, for the complex convolution kernel, the complex filter matrix W = A + iB is convolved with the complex vector h = x + iy. By using real numbers to simulate complex operations, in the formula, A and B are real matrices, x and y are real vectors, and the convolution operator is distributive. The filter W convolves the vector h.

[0043] The formula for batch normalization operation is as follows:

[0044]

[0045] In the batch normalization layer, x max represents the largest number in the sequence, and x min represents the smallest number in the sequence.

[0046] The formula for the activation function is as follows:

[0047]

[0048] Finally, after passing through the activation function layer, only the positive values of the data are used in the activation function, while the negative values of the data are all set to 0.

[0049] The beneficial effects of the present invention are as follows:

[0050] The method disclosed by the present invention uses a complex convolution module to extract the features of the I and Q channels, and uses an LSTM module to extract the time series features, giving full play to and extracting the relevant information of the I and Q channels and the time series to improve the recognition accuracy.

[0051] The method disclosed by the present invention adds relevant content and knowledge of the ionosphere to the data input, providing prior information of the ionospheric environment channel for the neural network, and combining relevant propagation characteristic factors to further improve the recognition accuracy.

[0052] The method disclosed by the present invention processes the collected data, and from compensation to network training, it is all autonomously completed by the program, forming a complete short-wave signal recognition system, providing a basis and convenience for carrying out corresponding work in the future. Description of the Drawings

[0053] Figure 1It is the overall architecture diagram of the integrated reconnaissance and identification of short-wave signals in the entire ionosphere;

[0054] Figure 2 It is a schematic diagram of ionospheric compensation using a transformer neural network;

[0055] Figure 3 It is the overall structure diagram of the recognition network;

[0056] Figure 4 It is a sample diagram of Guilin data;

[0057] Figure 5 It is Figure 4 the enlarged diagram of the data;

[0058] Figure 6 It is a sample diagram of Guiyang data;

[0059] Figure 7 It is Figure 6 the enlarged diagram of the data;

[0060] Figure 8 It is a sample diagram of Huizhou data;

[0061] Figure 9 It is Figure 8 the enlarged diagram of the data;

[0062] Figure 10 It is a schematic table of the specific structure of the network;

[0063] Figure 11 It is a curve diagram of the change of the loss function of the training set and the validation set;

[0064] Figure 12 It is a schematic diagram of the confusion matrix of the validation set;

[0065] Figure 13 It is a schematic diagram of the confusion matrix of the test set. Specific implementation manners

[0066] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0067] A deep learning short-wave signal recognition method based on ionospheric parameter compensation, which uses a complex neural and LSTM network superposition ionospheric parameter compensation network method to perform individual recognition of short-wave signals, includes the following steps:

[0068] Step 1, construct an overall neural network:

[0069] The constructed network should build a network model suitable for specific data. The model structure in this case is mainly divided into three parts: First, a Transformer neural network is used to build an ionospheric parameter compensation network. Then, a complex neural network is used to perform corresponding feature correlation on the I and Q channels of the signal, and an LSTM neural network is used to analyze the context correlation before and after the I and Q time series. Finally, a fully connected layer is used to identify and classify the extracted high-dimensional features;

[0070] The overall block diagram is shown in Figure 1 , and the structure of the network contains 6 convolutional modules, where each module includes a complex convolutional layer, a batch normalization layer, and an activation function.

[0071] The formula for complex convolution operation is as follows:

[0072] W*h = (A*x - B*y) + i(B*x + A*y)

[0073] First, there is a complex convolution kernel. The complex filter matrix W = A + iB is convolved with the complex vector h = x + iy. By using real numbers to simulate complex operations, in the formula, A and B are real matrices, x and y are real vectors, and the convolution operator is distributive. The filter W convolves the vector h;

[0074] The formula for batch normalization operation is as follows:

[0075]

[0076] In the batch normalization layer, x max represents the largest number in the sequence, and x min represents the smallest number in the sequence. The normalized data will, to a certain extent, prevent gradient vanishing and explosion;

[0077] The formula for the activation function is as follows:

[0078]

[0079] Finally, after passing through the activation function layer, only the positive values of the data are used in the activation function, while the negative values of the data are all set to 0, enhancing the nonlinear ability of the network.

[0080] Figure 2 is a schematic diagram of ionospheric compensation using a transformer neural network; Figure 3 is the overall structure diagram of the recognition network; Figure 10 is a schematic table of the specific network structure;

[0081] Step 2, process the data:

[0082] For the data just obtained, first, the program processes the data into a structure of (batch size, channel, width, length). Here, batch size represents the total number of data; channel represents the number of channels of the data. Here, it represents two channels of I and Q, so channel = 2; width represents the width of each channel, that is, the width of the data in each of the I and Q channels. Since it is a time series, width = 1; length represents the length of each channel. Here, the length usually refers to the number of sampling points of the I and Q channels. After that, the data is normalized for complex numbers:

[0083]

[0084]

[0085] In the above formula, real(x) represents the real part, and img(x) represents the imaginary part; through the above formula, the two normalized signals x I and x Q . The normalized data is more likely to converge and can better enable the neural network to reach the lowest point of the gradient.

[0086] Step 3, add ionospheric parameters and prior knowledge:

[0087] Regarding the variability of the ionosphere, smooth curve fitting is performed on the detected ionospheric parameters to obtain a polynomial to characterize the change of ionospheric parameters on the day:

[0088] Variable = ∑a i x i +b, i = 1, 2,..., n

[0089] In the above formula, a is the polynomial coefficient and b is the polynomial constant;

[0090] For the parameter fitting of the ionosphere, the variance λ of the fitting polynomial var is less than the measurement standard threshold α,

[0091] In this way, prior information is added to the original short-wave signal. Then, the ionospheric parameters, the collected real data, and the ideal data are input into the Transformer neural network, and the multi-head attention mechanism in the Transformer neural network is used:

[0092] V (Q,K,V) = X·W (Q,K,V)

[0093] Multiply the data X by the weight W to generate the query Q, the key K, and the value V, and then obtain a new matrix Z through the following transformation:

[0094]

[0095] In the above formula, d k represents the dimension of the key value K;

[0096] The multi-head attention mechanism in the Transformer can focus on different regions, enabling the network to capture the information of the ionosphere and then perform compensation.

[0097] Step 4: Perform forward gradient calculation and backward gradient derivation to update the parameters:

[0098] The data passes through the network and first undergoes a linear transformation of the matrix:

[0099] z = w T x + b

[0100] a = σ(z)

[0101] In the above formula, w is the weight matrix, b is the bias, and σ is the activation function;

[0102] The output of the last layer of the activation function is a probability vector, that is, [β 1 , β 2 ,..., β n , where n represents the total number of categories to be recognized, and the relationship satisfied is β 1 + β 2 +... + β n = 1.

[0103] Then use the cross-entropy loss function:

[0104]

[0105] In the above formula, is the predicted label, y is the true label, and through the deviation between the prediction and the truth, through the backward gradient operation, the predicted value is made as close as possible to the true value;

[0106] First, calculate the gradient of the weight w:

[0107]

[0108] And calculate the gradient of the bias:

[0109]

[0110] In the above formula, is the element weight value at the j-th row and k-th column of the l-th layer convolutional kernel; C is the obtained cost function, i.e., the loss function; is the output value of the previous layer; is the error between the prediction and the true value of the previous layer and this layer; is the bias value corresponding to the j-th row of the l-th layer convolutional kernel; is the bias value at the j-th row and k-th column of the l-th layer convolutional kernel; is the value output after the data passes through the j-th row of the l-th layer convolutional kernel;

[0111] Through gradient operations, obtain the corresponding weight and bias parameters, and then repeat steps 1 - 4 to obtain the weights that can reach the minimum loss value to form a trained network;

[0112] Step 5, obtain the recognition result, and obtain the category with the highest probability:

[0113] Use the untrained data. After processing the data format into the same format as the training data, directly input it into the trained network to obtain the output probability vector [β 1 ', β 2 ',..., β n '], and take the maximum value in the probabilities as the recognition result to obtain the final result of the network's recognition of this data.

[0114] Example 1, the method of this example is mainly implemented through python. The main libraries used are torch (deep learning library), numpy (scientific and mathematical operation library), and matplotlib (mainly used for drawing). The following are the specific implementation steps:

[0115] The data is obtained through on-site collection. Wuhan is used as the receiving point, and Huizhou, Guiyang, and Guilin are used as the sending points for transmission. After receiving the corresponding signals through the ionosphere, they are recognized. Figure 4 —9 are examples of some data collected in Guilin, Guiyang, and Huizhou respectively, where the red is the I-channel signal and the blue is the Q-channel signal.

[0116] First, organize the data set. The dimension of the collected data is channel = 2, and the shape is (2, 30000), that is, the time series length of the I and Q channels is 30000 points. Combine all the signals. Among them, there are 1333 pieces of Huizhou data, 1028 pieces of Guilin data, and 1155 pieces of Guiyang data. Integrate the data into a three-dimensional shape (1333 + 1028 + 1155, 2, 30000) = (3516, 2, 30000) conger to form a data set.

[0117] Label the data and set the labels in the form of one-hot encoding:

[0118] y label = [0, 0, ..., 1 (i) , ..., 0]

[0119] As can be seen, when the data is for the i-th device, the i-th column is marked as 1 and the rest are 0. The total number of columns is the total number of device types in the data. After encoding the labels, they are integrated into three-dimensional label data (3516, 1, 3), where 3516 represents the total number of labels; (1, 3) represents a one-hot encoding with a label value of 1 row and 3 columns.

[0120] Randomly sample the data. Shuffle the data to ensure the randomness of the data and that the random experiments conducted are not affected by the differences between categories, only focusing on the specific internal values.

[0121] Divide the training set, validation set, and test set according to a ratio. Divide them into a training set, validation set, and test set in a ratio of 8:1:1. The training set is used to update the network parameters; the validation set is used to select the optimal network; the test set is used to test the overall situation of the network.

[0122] Set the basic network parameters. Set the loss function of the network to the cross-entropy loss function, and adopt an initial learning rate of 0.0001 for training updates:

[0123] w n+1 = w n - lr·▽

[0124] In the above formula, w n+1 is the weight of the n + 1 layer, w n is the weight of the n layer, lr is the learning rate, and ▽ is the gradient calculated through the loss function. Figure 11 is the curve graph of the change in the loss function of the training set and the validation set;

[0125] Set batches for training and save the model. Divide the data into batches of 32 data for training to ensure that the graphics card memory can accommodate the corresponding quantity. Then set the number of training epochs to train the network. After each epoch of training, use the validation set to verify the network, obtain the recognition rate, ensure that overfitting of the network can be detected in a timely manner and stop training, and finally obtain the optimal network for corresponding storage.

[0126] Conduct data testing. Directly input the test set into the network according to the structure of the training set without training, label the output one-hot encoding, and calculate the recognition accuracy rate accordingly:

[0127]

[0128] In the above formula, Acc is the recognition accuracy rate, Num AccTo identify the correct number of samples, Num is the number of samples in the test set. This is the end of the experiment.

[0129] Figure 12 This is a diagram of the confusion matrix of the validation set after training; Figure 13 This is a diagram of the confusion matrix after testing with the test set. The recognition effect is 92%.

[0130] In summary, most of the traditional shortwave signal recognition methods can only achieve a recognition accuracy of about 50%-60%. After using ionospheric parameter compensation and then performing deep learning training, the recognition effect will tend to improve with the increase of ionospheric parameters, and eventually reach a recognition accuracy of more than 90%.

Claims

1. A deep learning shortwave signal recognition method based on ionospheric parameter compensation, characterized in that: The steps include: Step 1, construct the overall neural network: First, the Transformer neural network is used to build an ionospheric parameter compensation network. Then, a complex neural network is used to associate the corresponding features of the I and Q signals. An LSTM neural network is used to analyze the contextual association of the I and Q time series. Finally, a fully connected layer is used to identify and classify the extracted high-dimensional features. Step 2: Process the data: Normalize the data to complex numbers: In the above formula, real(x) represents the real part and img(x) represents the imaginary part; Step 3, add ionospheric parameters and prior knowledge: The detected ionospheric parameters are smoothed by curve fitting to obtain a polynomial to characterize the changes of ionospheric parameters on the day: Variable=∑a i x i +b,i=1,2,...,n In the above formula, a is the polynomial coefficient and b is the polynomial constant; For ionospheric parameter fitting, the variance λ of the fitting polynomial var is less than the measurement threshold α, Then the ionospheric parameters, the collected real data and the ideal data are input into the Transformer neural network, using the multi-head attention mechanism in the Transformer neural network: V (Q,K,V) =X·W (Q,K,V) Multiplying the data X by the weight W generates the query Q, key K and value V, which are then transformed into a new matrix Z by the following formula: In the above formula, d k Represents the dimension of the key value K; Step 4: Perform forward gradient calculation and reverse gradient derivation to update the parameters: The data passes through the network and first undergoes a linear transformation of the matrix: z=w T x+b a=σ(z) In the above formula, w is the weight matrix, b is the bias, and σ is the activation function; Then use the cross entropy loss function: In the above formula, is the predicted label, y is the real label, and the predicted value is brought close to the real value through the deviation between the prediction and the real value through reverse gradient operation; First calculate the gradient of weight w: And calculate the gradient of the bias: In the above formula, is the element weight value of the jth row and kth column of the lth layer convolution kernel; C is the obtained cost function, is the output value of the previous layer; is the error between the prediction of the previous layer and this layer and the true value; is the bias value corresponding to the jth row of the convolution kernel of the lth layer; is the bias value of the jth row and kth column of the lth layer convolution kernel; is the value output by the jth row of the data after passing through the lth layer of convolution kernel; Obtain the corresponding weights and bias parameters through gradient calculation, and then repeat steps 1-4 to obtain the weights that can achieve the minimum loss value to form a trained network; Step 5: Get the recognition result and the category with the highest probability: Use untrained data, process the data format into the same format as the training data, and directly input it into the trained network to obtain the output probability vector [β1',β2',...,β n '], take the maximum value of the probability as the recognition result, and obtain the final result of the network's recognition of this data.

2. The deep learning shortwave signal recognition method based on ionospheric parameter compensation according to claim 1 is characterized in that: In step 1, the network structure contains 6 layers of convolutional modules, each of which includes a complex convolutional layer, a batch normalization layer, and an activation function. The formula for complex convolution operation is as follows: W*h=(A*xB*y)+i(B*x+A*y) First, the complex convolution kernel convolves the complex filter matrix W = A + iB with the complex vector h = x + iy. The operation of complex numbers is simulated by using real numbers. In the formula, A and B are real matrices, x and y are real vectors, the convolution operator is distributed, and the filter W convolves the vector h. The formula for batch normalization operation is as follows: In the batch normalization layer, x max Represents the largest number in the sequence, x min Represents the smallest number in the sequence; The formula for the activation function is as follows: Finally, after the activation function layer, the activation function only uses the values ​​of the data that are positive, and all the negative values ​​of the data are taken as 0.

Images