Frequency hopping signal modulation identification and signal-to-noise ratio estimation based on multi-task learning

Abstract

The method comprises the following steps: firstly, carrying out short-time Fourier transform on IQ original data of received signals to generate a time-frequency graph, carrying out frequency hopping signal detection by using a model trained by YOLOV5 to generate a corresponding tag file, intercepting the signals from the IQ original data according to data in the tag file, uniformly taking a certain length of the intercepted signals as a data set, and carrying out data processing on the data set; and training by using an improved multi-task network, performing modulation identification and signal-to-noise ratio estimation by using the trained model, and outputting a modulation identification result and a signal-to-noise ratio estimation result. Through the improved CGC network, under the condition that too many training parameters are not increased, the performance between two tasks can be balanced, the seesaw phenomenon in multiple tasks is relieved, and the accuracy and generalization of the model are improved by combining large kernel convolution structure re-parameterization.

Description

Technical Field

[0001] The present invention relates to the technical field of frequency hopping signals, and in particular to frequency hopping signal modulation recognition and signal-to-noise ratio estimation based on multi-task learning. Background Art

[0002] Frequency hopping signals have strong anti-interference capabilities and are widely used in military communications. In electronic countermeasures, it is first necessary to estimate the parameters of unknown frequency hopping signals, including hopping speed, hopping frequency, hopping time, symbol rate, modulation style, etc., so as to demodulate and decode to obtain intelligence information, and finally adopt the best interference signal based on the parameter information of the intercepted signal to disrupt the normal communication of the enemy. This is the research focus in modern communication countermeasures.

[0003] First, deep learning can be used for the modulation recognition method of frequency hopping signals. By using the original IQ data of the frequency hopping signal as the feature input network, the modulation recognition accuracy of the trained model is high, but the generalization ability of the model is poor. When the symbol rate or other parameters of the frequency hopping signal change, the recognition accuracy of the model will drop significantly. Secondly, the seesaw phenomenon is prone to occur in multi-task learning, that is, the improvement of one task performance is achieved by losing the performance of another task. The Tencent team proposed the PLE model, in which the idea of ​​the CGC structure is to divide the network into an expert network and a shared network. The final output is obtained by adding the output of the shared network and the output of the expert network through the gating unit to assign corresponding weights, which can effectively alleviate the seesaw phenomenon. However, this structure also significantly increases the number of network parameters. Summary of the invention

[0004] In order to make up for the above shortcomings, the present invention provides frequency hopping signal modulation recognition and signal-to-noise ratio estimation based on multi-task learning, aiming to improve the problem in the prior art that operators are required to make on-site adjustments and the operation is inconvenient.

[0005] In order to achieve the above purpose, the present invention adopts the following technical scheme: frequency hopping signal modulation recognition and signal-to-noise ratio estimation based on multi-task learning. First, the IQ raw data of the received signal is short-time Fourier transformed to generate a time-frequency diagram, and the model trained by YOLOV5 is used to detect the frequency hopping signal and generate the corresponding label file. The signal is intercepted from the IQ original data according to the data in the label file, and a certain length of the intercepted signal is uniformly taken as the data set.

[0006] Use the improved multi-task network for training, use the trained model for modulation recognition and signal-to-noise ratio estimation, and output the modulation recognition results and signal-to-noise ratio estimation results.

[0007] The large kernel convolution structure is used to re-parameterize the MobileNetV3 block structure (MobileBlock) in the multi-task network structure. The designed network has five layers of MobileBlock. The large convolution kernel structure is used to re-parameterize the MobileBlock3 and MobileBlock5 layers, and the residual structure is used. The output of the expert channel is added to the output dimension of the shared channel and input into the task network. Task 1 is a classifier to realize the modulation recognition of the signal, and Task 2 is a regression task to realize the estimation of the signal-to-noise ratio.

[0008] Loss function design: This method targets two tasks: modulation mode recognition and signal-to-noise ratio estimation of frequency hopping signals. The modulation mode recognition uses a label smoothing loss function, which changes the true probability of the label to:

[0009] The network outputs a confidence score for each category of the current data. These scores are normalized through softmax to obtain the probability of belonging to each category:

[0010] Then calculate the cross entropy loss function:

[0011] Furthermore, for the signal-to-noise ratio estimation regression problem, the mean square error is used as the loss function, that is, the average of the sum of the squares of the differences between the predicted value and the true value. The formula is as follows: Since the two loss functions are calculated in different ways, when one of the loss functions is too large, the convergence of the other task will slow down. Therefore, the same variance uncertainty loss function is used to balance the loss function values ​​of the two. The specific formula is as follows: The present invention has the following beneficial effects: The improved CGC network can balance the performance between the two tasks without increasing too many training parameters, alleviate the seesaw phenomenon in multi-tasks, and improve the accuracy and generalization of the model by combining large-kernel convolution structure reparameterization. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 The present invention is a flow chart of modulation recognition and signal-to-noise ratio estimation of a frequency hopping signal; Figure 2 This is a schematic diagram of the large kernel convolution re-parameterization principle of the present invention; Figure 3 This is a schematic diagram of the network model structure of the present invention; Figure 4 The accuracy test results of different symbol rates of the present invention are as follows; Figure 5 The mean square error test results of different symbol rates of the present invention; DETAILED DESCRIPTION

[0013] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0014] Frequency hopping signal modulation recognition and signal-to-noise ratio estimation based on multi-task learning include: first, performing short-time Fourier transform on the IQ data of the received signal to generate a time-frequency diagram, using the model trained by YOLOV5 to detect the frequency hopping signal, generating the corresponding label file, intercepting the signal according to the data in the label file, and uniformly taking a certain length of the intercepted signal as a data set. Use the improved multi-task network for training, use the trained model for modulation recognition and signal-to-noise ratio estimation, and output the modulation recognition results and signal-to-noise ratio estimation results. The entire flow chart is as follows: Figure 1 As shown: Improvement of multi-task network structure: The present invention divides the number of channels of the network into expert channels and shared channels, and optimizes parameters such as the number of channels and hidden layer expansion ratio of the network based on MobileNetV3, thereby reducing the amount of training parameters. At the same time, a large-core convolution kernel structure reparameterization method is used. The main principle is to use two convolution kernels, one large and one small, during training, and to fuse the convolution kernel parameters when saving the model. The reason for using a large convolution kernel is that more contextual information can be obtained, and a small convolution kernel can obtain local texture features. At the same time, since large convolution kernels are difficult to train, using a small convolution kernel can avoid smooth transitions. The principle of large convolution kernel structure reparameterization is as follows: Figure 2 As shown: The present invention uses a large kernel convolution structure reparameterization based on the block structure (MobileBlock) of MobileNetV3. The designed network has five layers of MobileBlock. Large convolution kernel structure reparameterization is used in MobileBlock3 and MobileBlock5, and a residual structure is used. The output of the expert channel is added to the output dimension of the shared channel and input into the task network. Task 1 is a classifier to realize signal modulation recognition, and Task 2 is a regression task to realize signal-to-noise ratio estimation. Figure 3 shown.

[0015] Loss function design: This method targets two tasks: modulation mode recognition and signal-to-noise ratio estimation of frequency hopping signals. The modulation mode recognition uses a label smoothing loss function, which changes the true probability of the label to:

[0016] The network outputs a confidence score for each category of the current data. These scores are normalized through softmax to obtain the probability of belonging to each category:

[0017] Then calculate the cross entropy loss function:

[0018] For signal-to-noise ratio estimation regression problems, the mean square error is used as the loss function, which is the average of the sum of the squares of the differences between the predicted value and the true value. The formula is as follows: Since the two loss functions are calculated in different ways, when one of the loss functions is too large, the convergence of the other task will slow down. Therefore, the same variance uncertainty loss function is used to balance the loss function values ​​of the two. The specific formula is as follows:

[0019] Working principle: The simulation generates four phase shift keying modulation signals, namely BPSK, QPSK, 8PSK and 16QAM. The frequency range of the frequency hopping signal is 5~100MHz, with a total of 20 hopping frequencies, the signal-to-noise ratio ranges from -10dB to 10dB, the hopping rate is 300 hops per second, the sampling frequency is set to 100MHz, and the symbol rate is set to 1e6 / 130 as the training data set. In addition, three frequency hopping signals with different symbol rates are generated as test sets. The test results are shown in the following table: Comparison table of model structure experimental results

[0020] Compared with the MMOE_small model, the modulation recognition accuracy of the CGC_small model increased, and the mean square error of the signal-to-noise ratio estimation increased slightly. Compared with the MMOE_large model, the classification task accuracy of the CGC_large model decreased by 0.84%, while the mean square error of the signal-to-noise ratio estimation decreased from 7.568 to 4.064. This shows that the CGC structure model can better balance the performance of the two tasks than the MMOE structure model, and can avoid the situation where one task has good performance and the other has poor performance, effectively alleviating the seesaw phenomenon in multi-task training. Comparing the results of the CGC_small model with those of the CGC_large model, the modulation recognition accuracy of the CGC_large network has been greatly improved, and the MSELoss value has decreased. It can be concluded that the method of reparameterization of the large kernel convolution structure can improve the accuracy of the model.

[0021] The large convolution kernel sizes are [81, 1], [121, 1], [169, 1], and [225, 1], respectively. The data sets are tested at different symbol rates. The accuracy of the classification task and the MSELoss value of the regression task are as follows: Figure 4-5 The 5_81 curve in the figure represents the training results of the model with a small convolution kernel size of [5, 1] ​​and a large convolution kernel size of [81, 1].

[0022] Depend on Figure 4-5 It can be seen that the greater the difference between the symbol rate and the symbol rate of the training set, the worse the performance of the model. Under the condition of the same symbol rate, the larger the convolution kernel size, the higher the model modulation recognition accuracy and the smaller the mean square error value, that is, the better the generalization ability of the model.

[0023] The above experiments show that the improved CGC network can balance the performance between the two tasks without increasing too many training parameters, alleviate the seesaw phenomenon in multi-tasks, and improve the accuracy and generalization of the model by combining large kernel convolution structure reparameterization.

[0024] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. Frequency hopping signal modulation recognition and signal-to-noise ratio estimation based on multi-task learning, characterized in that: First, the IQ raw data of the received signal is short-time Fourier transformed to generate a time-frequency diagram, and the model trained by YOLOV5 is used to detect the frequency hopping signal and generate the corresponding label file. The signal is intercepted from the IQ original data according to the data in the label file, and a certain length of the intercepted signal is uniformly taken as the data set; Use the improved multi-task network for training, use the trained model for modulation recognition and signal-to-noise ratio estimation, and output the modulation recognition results and signal-to-noise ratio estimation results; The designed network has five layers of MobileBlocks. The large convolution kernel structure is used to re-parameterize the MobileNetV3 block structure (MobileBlock) in the multi-task network structure. The MobileBlock3 and MobileBlock5 layers are re-parameterized with large convolution kernel structures. The residual structure is used. The output of the expert channel is added to the output dimension of the shared channel and input into the task network. Task 1 is a classifier to realize signal modulation recognition, and Task 2 is a regression task to realize signal-to-noise ratio estimation. Loss function design: This method targets two tasks: modulation mode recognition and signal-to-noise ratio estimation of frequency hopping signals. The modulation mode recognition uses a label smoothing loss function, which changes the true probability of the label to: ， The network outputs a confidence score for each category of the current data. These scores are normalized through softmax to obtain the probability of belonging to each category: ， Then calculate the cross entropy loss function: 。 2. The frequency hopping signal modulation recognition and signal-to-noise ratio estimation based on multi-task learning according to claim 1, characterized in that: For the signal-to-noise ratio estimation regression problem, the mean square error is used as the loss function, which is the average of the sum of the squares of the differences between the predicted value and the true value. The formula is as follows: ， Since the two loss functions are calculated in different ways, when the value of one loss function is too large, the convergence of the other task will slow down. Therefore, the same variance uncertainty loss function is used to balance the loss function values ​​of the two. The specific formula is as follows: 。

Images