CN117407785B

CN117407785B - Training method of radar signal recognition model, radar signal recognition method and device

Info

Publication number: CN117407785B
Application number: CN202311728693.1A
Authority: CN
Inventors: 郭琦; 武斌; 门兰宁
Original assignee: Xi'an Shengxin Technology Co ltd
Current assignee: Xi'an Shengxin Technology Co ltd
Priority date: 2023-12-15
Filing date: 2023-12-15
Publication date: 2024-03-01
Anticipated expiration: 2043-12-15
Also published as: CN117407785A

Abstract

The embodiment of the invention relates to the technical field of radars, and discloses a training method of a radar signal recognition model, a radar signal recognition method and a device, wherein the method comprises the following steps: acquiring a sample set, wherein the sample set comprises at least two types of sample radar signals and sample tags of each sample radar signal in the at least two types of sample radar signals; obtaining a signal type prediction result of the sample radar signals in the sample set according to the sample radar signals in the sample set and the initial radar signal identification model; determining a loss value according to a signal type prediction result of the sample radar signal in the sample set and a sample tag of the sample radar signal in the sample set; the loss value is determined according to a target loss function, and the target loss function can adjust the loss value corresponding to the sample radar signal according to the identification difficulty of the sample radar signal; and iteratively updating the initial radar signal identification model according to the loss value to obtain the target radar signal identification model.

Description

Training method of radar signal recognition model, radar signal recognition method and device

Technical Field

The embodiment of the invention relates to the technical field of radars, in particular to a training method of a radar signal recognition model, a radar signal recognition method and a radar signal recognition device.

Background

With the development of the electronic information field, electronic countermeasure plays an important role in electronic intelligence reconnaissance, electronic support and threat alert systems, and radar radiation source signal identification is an important link in electronic countermeasure.

In recent years, with the rapid development of deep learning technology, the radar signals can be classified and identified by using a neural network model, so that the problems of low efficiency and low precision existing in the traditional manual identification mode can be solved.

In the related art, in the process of classifying and identifying radar signals through the existing neural network model, a great number of training samples are needed to train the neural network model to ensure certain classification accuracy, however, when the training samples are insufficient, the neural network model can be fitted, so that the classification accuracy of the neural network model is lower.

Disclosure of Invention

In view of the above problems, embodiments of the present invention provide a method for training a radar signal recognition model, a method for radar signal recognition, and a device for radar signal recognition, which are used for solving the problem that when a training sample is insufficient, the existing radar signal recognition model is over-fitted, resulting in lower classification accuracy of the model.

According to a first aspect of an embodiment of the present application, there is provided a training method of a radar signal identification model, including: acquiring a sample set, wherein the sample set comprises at least two types of sample radar signals and sample tags of each sample radar signal in the at least two types of sample radar signals; obtaining a signal type prediction result of the sample radar signals in the sample set according to the sample radar signals in the sample set and the initial radar signal identification model; determining a loss value according to a signal type prediction result of the sample radar signal in the sample set and a sample tag of the sample radar signal in the sample set; the loss value is determined according to a target loss function, and the target loss function can adjust the loss value corresponding to the sample radar signal according to the identification difficulty of the sample radar signal; and iteratively updating the initial radar signal identification model according to the loss value to obtain the target radar signal identification model.

In an alternative way, the above-mentioned loss value is obtained by the following formula:wherein (1)>For loss value, +_>The result is predicted for the signal type of the sample radar signal in the sample set.

In an alternative manner, the loss value is determined according to a target loss function and attention weight coefficients, and the attention weight coefficients corresponding to different sample radar signals are different.

In an alternative way, the above-mentioned loss value is obtained by the following formula:wherein (1)>For loss value, +_>Attention weight coefficient corresponding to sample radar signal in sample set, < ->The result is predicted for the signal type of the sample radar signal in the sample set.

In an alternative manner, iteratively updating the initial radar signal identification model according to the loss value to obtain the target radar signal identification model, including: iteratively updating the initial radar signal recognition model according to the loss value to obtain a trained radar signal recognition model; under the condition that a first preset condition is met, determining the trained radar signal recognition model as a target radar signal recognition model; the first preset condition comprises: the loss value is less than or equal to a preset threshold; or the iteration times of the initial radar signal identification model reach the preset times.

In an alternative way, the sample set comprises N sample subsets, each sample subset of the N sample subsets comprising at least two types of sample radar signals, and a sample tag for each sample radar signal; after acquiring the sample set, the method further comprises: taking any sample subset in the N sample subsets as a test sample subset, and taking sample subsets except the test sample subset in the N sample subsets as training sample subsets; obtaining a signal type prediction result of the sample radar signal in the sample set according to the sample radar signal in the sample set and the initial radar signal identification model, wherein the method comprises the following steps: obtaining a signal type prediction result of the sample radar signals in the training sample subset according to the sample radar signals in the training sample subset and the initial radar signal identification model; determining a loss value according to a signal type prediction result of the sample radar signal in the sample set and a sample tag of the sample radar signal in the sample set, including: and determining a loss value according to the signal type prediction result of the sample radar signals in the training sample subset and the sample labels of the sample radar signals in the training sample subset.

In an alternative way, determining the trained radar signal identification model as the target radar signal identification model includes: obtaining a signal type prediction result of the sample radar signals in the test sample subset according to the sample radar signals in the test sample subset and the trained radar signal identification model; determining a verification score according to a signal type prediction result of the sample radar signals in the test sample subset and sample labels of the sample radar signals in the test sample subset; and under the condition that the verification score meets the second preset condition, determining the trained radar signal identification model as a target radar signal identification model.

In an alternative manner, determining the verification score according to the signal type prediction result of the sample radar signal in the test sample subset and the sample tag of the sample radar signal in the test sample subset includes: under the condition that the training period of the initial radar signal identification model reaches a first preset period, determining a current verification score according to a signal type prediction result of the sample radar signals in the test sample subset and sample labels of the sample radar signals in the test sample subset; under the condition that the verification score meets the second preset condition, determining the trained radar signal identification model as a target radar signal identification model comprises the following steps: and under the condition that the current verification score is smaller than or equal to the last verification score, determining the trained radar signal recognition model as a target radar signal recognition model.

In an optional manner, the initial radar signal identification model includes at least two convolution layers and a classification layer, and according to the sample radar signals in the sample set and the initial radar signal identification model, a signal type prediction result of the sample radar signals in the sample set is obtained, including: processing the sample radar signals in the sample set according to at least two convolution layers to obtain a target radar signal characteristic diagram; and processing the target radar signal feature map according to the classification layer to obtain a signal type prediction result of the sample radar signal in the sample set.

In an alternative, the at least two convolution layers include three first convolution layers, three second convolution layers, three third convolution layers, and three fourth convolution layers; the size of the convolution kernels of the second convolution layer is the same as that of the convolution kernels of the first convolution layer, the number of the convolution kernels of the second convolution layer is larger than that of the convolution kernels of the first convolution layer, the size of the convolution kernels of the third convolution layer is larger than that of the second convolution layer, the number of the convolution kernels of the third convolution layer is equal to that of the second convolution layer, the size of the convolution kernels of the fourth convolution layer is the same as that of the convolution kernels of the third convolution layer, and the number of the convolution kernels of the fourth convolution layer is larger than that of the third convolution layer; processing the sample radar signals in the sample set according to at least two convolution layers to obtain a target radar signal feature map, wherein the processing comprises the following steps: processing the sample radar signals in the sample set according to the three first convolution layers to obtain a first radar signal characteristic diagram; processing the first radar signal feature map according to the three second convolution layers to obtain a second radar signal feature map; processing the second radar signal feature map according to the three third convolution layers to obtain a third radar signal feature map; and processing the third radar signal feature map according to the three fourth convolution layers to obtain a target radar signal feature map.

In an alternative way, iteratively updating the initial radar signal identification model according to the loss value includes: under the condition that the training period of the initial radar signal identification model does not reach the second preset period, iteratively updating the initial radar signal identification model according to the loss value based on the first learning rate; and under the condition that the training period of the initial radar signal recognition model reaches a second preset period, iteratively updating the initial radar signal recognition model according to the loss value based on a second learning rate, wherein the second learning rate is smaller than the first learning rate.

According to a second aspect of embodiments of the present application, there is provided a radar signal identification method, the method further comprising: acquiring a radar signal to be identified; determining the signal type of the radar signal to be identified according to the radar signal to be identified and a target radar signal identification model, wherein the target radar signal identification model is obtained by training according to the training method of the radar signal identification model in any one of the first aspect.

According to a third aspect of embodiments of the present application, there is provided a training apparatus for a radar signal identification model, the apparatus including: the acquisition module is used for acquiring a sample set, wherein the sample set comprises at least two types of sample radar signals and sample labels of the sample radar signals in the at least two types of sample radar signals; the prediction module is used for obtaining a signal type prediction result of the sample radar signals in the sample set according to the sample radar signals in the sample set and the initial radar signal identification model; the determining module is used for determining a loss value according to a signal type prediction result of the sample radar signals in the sample set and a sample tag of the sample radar signals in the sample set; the loss value is determined according to a target loss function, and the target loss function can adjust the loss value corresponding to the sample radar signal according to the identification difficulty of the sample radar signal; and the training module is used for iteratively updating the initial radar signal recognition model according to the loss value to obtain the target radar signal recognition model.

According to a fourth aspect of embodiments of the present application, there is provided a radar signal identifying apparatus, the apparatus comprising: the acquisition module is used for acquiring radar signals to be identified; the recognition module is used for determining the signal type of the radar signal to be recognized according to the radar signal to be recognized and a target radar signal recognition model, wherein the target radar signal recognition model is obtained by training the training device of the radar signal recognition model according to the third aspect.

According to a fifth aspect of embodiments of the present application, there is provided an electronic device, including: a processor, a memory, and executable instructions stored on the memory and executable on the processor, the processor executing the executable instructions implementing the method of training the radar signal identification model according to any one of the first aspects or the method of radar signal identification according to the second aspect.

According to a sixth aspect of embodiments of the present application, there is provided a computer-readable storage medium, in which at least one executable instruction is stored, which when executed on an electronic device, causes the electronic device to perform the operations of the method for training a radar signal identification model as set forth in any one of the first aspects, or to perform the operations of the method for radar signal identification as set forth in any one of the second aspects.

According to a seventh aspect of embodiments of the present application, there is provided a computer program product having stored therein at least one executable instruction for causing an electronic device to perform the operations of the method of training a radar signal identification model as described in any of the first aspects above, or to perform the operations of the method of radar signal identification as described in any of the second aspects above.

According to the embodiment of the application, in the process of training the radar signal identification model, a sample set comprising at least two types of sample radar signals can be obtained, according to the sample radar signals in the sample set and the initial radar signal identification model, a signal type prediction result of the sample radar signals in the sample set is obtained, then a loss value is determined according to the signal type prediction result of the sample radar signals in the sample set and a sample tag of the sample radar signals in the sample set, and further the initial radar signal identification model is iteratively updated according to the loss value, so that the target radar signal identification model is obtained. The loss value is determined according to a target loss function, and the target loss function can adjust the loss value corresponding to the sample radar signal according to the identification difficulty of the sample radar signal. That is, the objective loss function according to the embodiment of the present application can automatically adjust the loss value of the radar signal of different samples without adjusting the concentration factor γ. In this way, the loss value is calculated based on the target loss function, and then the initial radar signal recognition model is iteratively trained through the loss value, so that the trained target radar signal recognition model pays more attention to the learning and classification of difficult samples, the performance of the target radar signal recognition model is improved, and the recognition rate of radar signals is improved.

Furthermore, the initial radar signal identification model in the embodiment of the present application may be a convolutional neural network model built with a pyramid structure. Therefore, compared with the traditional standard convolution, the depth separable convolution can be adopted to combine the information of different channels, so that the parameter quantity and the calculated quantity can be effectively reduced, the model training efficiency is improved, and meanwhile, the expression capacity of the model can be maintained.

In addition, the K-fold verification method is adopted to train the initial radar signal recognition model, and the trained radar signal recognition model is verified, so that the method can be suitable for scenes with smaller sample data size, the evaluation model performance is more objective and reliable, the performance of the target radar signal recognition model is ensured, and the recognition capability of a small sample of the target radar signal recognition model is improved.

The foregoing description is only an overview of the technical solutions of the embodiments of the present application, and may be implemented according to the content of the specification, so that the technical means of the embodiments of the present application can be more clearly understood, and the following detailed description of the present application will be presented in order to make the foregoing and other objects, features and advantages of the embodiments of the present application more understandable.

Drawings

The drawings are only for purposes of illustrating embodiments and are not to be construed as limiting the application. Also, like reference numerals are used to designate like parts throughout the figures. In the drawings:

fig. 1 is a schematic flow chart of a training method of a radar signal recognition model according to an embodiment of the present application;

fig. 2 shows a schematic diagram of a chirp signal provided in an embodiment of the present application;

FIG. 3 is a schematic diagram of a two-phase encoded signal provided by an embodiment of the present application;

fig. 4 shows a schematic diagram of a two-frequency encoded signal according to an embodiment of the present application;

FIG. 5 shows a schematic diagram of an initial radar signal identification model provided by an embodiment of the present application;

FIG. 6 is a flowchart illustrating another method for training a radar signal identification model according to an embodiment of the present application;

FIG. 7 is a schematic diagram of a training and verification process for a radar signal identification model provided by an embodiment of the present application;

FIG. 8 is a schematic diagram of a training algebra and recognition accuracy change curve according to an embodiment of the present application;

fig. 9 shows a schematic flow chart of a radar signal identification method according to an embodiment of the present application;

fig. 10 shows a schematic hardware structure of a training device of a radar signal recognition model according to an embodiment of the present application;

Fig. 11 shows a schematic hardware structure of a radar signal identifying apparatus according to an embodiment of the present application;

fig. 12 shows a schematic hardware structure of an electronic device according to an embodiment of the present application.

Detailed Description

Exemplary embodiments of the present application will be described in more detail below with reference to the accompanying drawings. While exemplary embodiments of the present application are shown in the drawings, it should be understood that the present application may be embodied in various forms and should not be limited to the embodiments set forth herein.

Fig. 1 is a schematic flow chart of a training method of a radar signal recognition model according to an embodiment of the present application. The method may be performed by a server. As shown in FIG. 1, the method may include steps 110-140.

Step 110, a sample set is obtained, wherein the sample set comprises at least two types of sample radar signals, and a sample tag of each sample radar signal in the at least two types of sample radar signals.

In this embodiment, the sample set may be a sample set for model training. More specifically, the sample set may be a sample set used to train an initial radar signal identification model.

The sample set may include multiple types of sample radar signals (which may also be referred to as sample radar radiation source signals), as well as sample tags for each sample radar signal. It should be noted that different types of sample radar signals may refer to different phases of the sample radar signals, or may refer to different carrier frequencies of the sample radar signals.

In some examples, the sample radar signal may be a single pulse signal, such as a non-coherent pulse signal, a coherent varying periodic pulse signal. The sample radar signal may also be a frequency modulated pulse signal, such as a chirp signal, a non-chirp signal, a step frequency pulse signal. The sample radar signal may also be a phase encoded pulse signal, such as a two-phase encoded pulse signal, a multi-phase encoded pulse signal (e.g., a four-phase encoded pulse signal).

By way of example, the sample set may comprise seven types of sample radar signals, namely a chirp signal, a non-chirp signal, a fixed frequency signal, a two-phase encoded signal, a four-phase encoded signal, a two-frequency encoded signal, a four-frequency encoded signal.

The carrier frequency of the linear frequency modulation signal is randomly selected from 100-400 MHz, and the bandwidth of the linear frequency modulation signal can be 20MHz, 25MHz, 30MHz or 40MHz. Exemplary, a schematic of a chirp signal is shown in fig. 2.

The nonlinear frequency modulation signal can adopt cosine modulation, the carrier frequency of the nonlinear frequency modulation signal is selected randomly at 50-500 MHz, and the window function of the nonlinear frequency modulation signal can be a Hanning window function or a Blackman window function.

The carrier frequency of the fixed frequency signal may be 100MHz, 110MHz, 120MHz, or 130MHz.

The carrier frequency of the two-phase encoded signal is 200 MHz. The two-phase coded signal can be coded by 5-bit, 7-bit, 11-bit and 13-bit barker codes respectively. The pulse width of the two-phase encoded signal is 1 mus. An exemplary diagram of a two-phase encoded signal is shown in fig. 3.

The carrier frequency of the four-phase coded signal is selected randomly at 200-350 MHz. The four-phase coded signal can adopt a coding mode of 16-bit French four-phase codes. The pulse width of the four-phase encoded signal is 1 mus.

The two-frequency coded signal may have a plurality of carrier frequencies, for example, the carrier frequencies of the two-frequency coded signal are 50MHz and 100MHz; the carrier frequency of the two-frequency coding signal is 50MHz and 75MHz; the carrier frequency of the two-frequency coding signal is 250MHz and 75MHz; the carrier frequencies of the two-frequency encoded signals are 250MHz and 50MHz. The two-frequency coding signal can adopt a coding mode of 13-bit barker codes. The pulse width of the two-frequency encoded signal is 1 mus. The symbol width of the two-frequency encoded signal is 0.77 mus. Exemplary, a schematic diagram of a two-frequency encoded signal is shown in fig. 4.

The four-frequency encoded signal may have four carrier frequencies, such as 100MHz, 150 MHz, 200MHz, and 250MHz. The four-frequency coding signal can adopt a coding mode of 13-bit barker codes. The pulse width of the four-frequency encoded signal is 1 mus. The symbol width of the four-frequency encoded signal is 0.77 mus.

In some examples, the number of sample radar signals in the sample set may be set according to the complexity of identification of the radar signal being targeted. The number of sample radar signals of each type in the sample set may be the same or different. In addition, the signal-to-noise ratio of each sample radar signal in the sample set is smaller than a preset signal-to-noise ratio. For example, the preset signal-to-noise ratio is 5dB.

Taking the example that the sample set includes the above seven types of sample radar signals, in the case that the number of the various types of sample radar signals in the sample set is the same, the number of each type of sample radar signals may be 50, that is, the sample set includes 350 sample radar signals. In the implementation, the sampling frequency of each sample radar signal is 10000Hz, and the number of sampling points is 10000.

The sample tag of the sample radar signal is used for identifying the signal type of the sample radar signal. That is, the sample tags corresponding to different types of sample radar signals are different. The difference (also called as distance) between the signal type prediction result output by the radar signal identification model and the sample label can be used for measuring the classification effect of the radar signal identification model.

For example, the sample tag of the sample radar signal may be a coded sequence obtained by using one hot coding mode. For example, the sample tag of a first type of radar signal (e.g., a chirp signal) is 0000001; sample tags of the second type of radar signal (e.g., non-chirp signal) are 0000010; sample tags of a third type of radar signal (e.g., fixed frequency signal) are 0000100; sample tags for a fourth type of radar signal (e.g., two-phase encoded signal) are 0001000; the sample tag of the fifth type of radar signal (e.g., a four-phase encoded signal) is 0010000; the sample label of the sixth type radar signal (such as a two-frequency coded signal) is 0100000; the sample tag of the seventh type of radar signal (e.g., a four-frequency encoded signal) is 1000000.

In the implementation, when each sample radar signal in the sample set is collected, a manual labeling mode can be adopted to set sample labels for each sample radar signal.

And 120, obtaining a signal type prediction result of the sample radar signals in the sample set according to the sample radar signals in the sample set and the initial radar signal identification model.

The initial radar signal recognition model is provided with initial model parameters, and a signal radar prediction result of a sample radar signal output by the initial radar signal recognition model is related to the initial model parameters. It should be noted that, the design of the initial radar signal identification model may be any feasible model architecture, which is not specifically limited in the embodiments of the present application.

In some embodiments, the initial radar signal identification model may be a convolutional neural network model built using a pyramid structure. For example, the initial radar signal identification model may include at least two convolution layers and a classification layer, wherein the at least two convolution layers may output radar signal feature maps of different scales, which may in turn capture features of the radar signal from different scales and resolutions. Moreover, each convolution layer can adopt depth separable convolution, so that compared with the traditional standard convolution, the depth separable convolution can be adopted to combine information of different channels, further, the parameter quantity and the calculated quantity can be effectively reduced, the model training efficiency is improved, and meanwhile, the expression capacity of the model can be maintained.

In this embodiment, the step of obtaining a signal type prediction result of the sample radar signal in the sample set according to the sample radar signal in the sample set and the initial radar signal identification model may further include: processing the sample radar signals in the sample set according to at least two convolution layers to obtain a target radar signal characteristic diagram; and processing the target radar signal feature map according to the classification layer to obtain a signal type prediction result of the sample radar signal in the sample set.

That is, the sample radar signals in the sample set are used as the input of the initial radar signal identification model, and are sequentially processed through at least two convolution layers to obtain the target radar signal feature map. And then, the classification layer performs classification processing on the target radar signal feature map to obtain a signal type prediction result of the sample radar signal.

In some embodiments, before obtaining the signal type prediction result of the sample radar signal in the sample set according to the sample radar signal in the sample set and the initial radar signal identification model, the method may further include: each sample radar signal in the sample set is preprocessed. Illustratively, each sample radar signal in the sample set is normalized to obtain a one-dimensional sample radar signal sequence. That is, after the sample set is obtained, each sample radar signal in the sample set is normalized to obtain a one-dimensional sample radar signal sequence, and then the one-dimensional sample radar signal sequence is used as the input of the initial radar signal identification model, i.e. the one-dimensional sample radar signal sequence is sequentially processed by at least two convolution layers. In this embodiment, after the sample set is obtained, preprocessing is performed on each sample radar signal in the sample set, so that diversity and number of sample data can be increased, and generalization performance and robustness of the model can be improved.

In some embodiments, referring to fig. 5, the at least two convolution layers include three first convolution layers, three second convolution layers, three third convolution layers, and three fourth convolution layers. That is, the initial radar signal identification model may include twelve convolutional layers (i.e., twelve one-dimensional convolutional neural networks). The size of the convolution kernels of the second convolution layer is the same as that of the convolution kernels of the first convolution layer, the number of the convolution kernels of the second convolution layer is larger than that of the convolution kernels of the first convolution layer, the size of the convolution kernels of the third convolution layer is larger than that of the convolution kernels of the second convolution layer, the number of the convolution kernels of the third convolution layer is equal to that of the convolution kernels of the second convolution layer, the size of the convolution kernels of the fourth convolution layer is the same as that of the convolution kernels of the third convolution layer, and the number of the convolution kernels of the fourth convolution layer is larger than that of the convolution kernels of the third convolution layer.

In this embodiment, the step of processing the sample radar signals in the sample set according to at least two convolution layers to obtain the target radar signal feature map may further include: processing the sample radar signals in the sample set according to the three first convolution layers to obtain a first radar signal characteristic diagram; processing the first radar signal feature map according to the three second convolution layers to obtain a second radar signal feature map; processing the second radar signal feature map according to the three third convolution layers to obtain a third radar signal feature map; and processing the third radar signal feature map according to the three fourth convolution layers to obtain a target radar signal feature map.

That is, the sample radar signals in the sample set are used as the input of the first convolution layers in the initial radar signal identification model, and are sequentially processed through the three first convolution layers to obtain the first radar signal feature map. And then, the first radar signal characteristic map is used as the input of a second convolution layer in the initial radar signal identification model, and is sequentially processed by three second convolution layers to obtain a second radar signal characteristic map. And then, the second radar signal characteristic map is used as the input of a third convolution layer in the initial radar signal identification model, and is sequentially processed by the three third convolution layers to obtain a third radar signal characteristic map. And finally, the third radar signal characteristic diagram is used as the input of a fourth convolution layer in the initial radar signal identification model, and is processed through the three fourth convolution layers in sequence to obtain the target radar signal characteristic diagram.

The classification and identification process of the initial radar signal identification model and the sample radar signal will be described below by taking an example that the initial radar signal identification model may include twelve convolution layers.

As shown in fig. 5, the initial radar signal identification model may include twelve convolutional layers (i.e., twelve one-dimensional convolutional neural networks). Specifically, the size of the convolution kernel of the first three layers (may also be referred to as the size of the convolution kernel) is 1×3, and the number of convolution kernels is 32, that is, the size of the convolution kernel of the first convolution layer is 1×3, and the number of convolution kernels of the first convolution layer is 32. The convolution kernels of the fourth layer to the sixth layer have a size of 1×3, and the number of convolution kernels is 64, that is, the convolution kernels of the second convolution layer have a size of 1×3, and the number of convolution kernels of the second convolution layer is 64. The size of the convolution kernels of the seventh layer to the ninth layer is 1×5, the number of convolution kernels is 64, that is, the size of the convolution kernels of the third convolution layer is 1×5, and the number of the convolution kernels of the third convolution layer is 64. The sizes of the convolution kernels of the tenth layer to the twelfth layer are 1×5, and the number of convolution kernels is 128, namely, the size of the convolution kernel of the fourth convolution layer is 1×5, and the number of convolution kernels of the fourth convolution layer is 128.

In the specific implementation, a sample radar signal is used as the input of an initial radar signal identification model, and sequentially passes through the first three first convolution layers to obtain a first radar signal characteristic diagram; then, the first radar feature map sequentially passes through three second convolution layers to output a second radar signal feature map; then, the second radar feature map sequentially passes through three third convolution layers to output a third radar signal feature map; and finally, the third radar feature map sequentially passes through three fourth convolution layers to output a target radar signal feature map.

The following is a convolution kernel of 1×3 (i.e., first convolution layer)For example, a process of generating the first radar signal characteristic map will be described.

Specifically, the first radar signal characteristic map may be generated by the following formulas (1) - (4).

In this embodiment, the sample radar signal is sequentially processed by each convolution layer in the initial radar signal identification model, so that radar signal feature maps with different scales and resolutions can be extracted, and further, the initial radar signal identification model is trained based on the radar signal feature maps with different scales and resolutions, so that the number of model parameters and the calculated amount of the model can be effectively reduced, the training efficiency of the model can be improved, and the expression capability of the model can be ensured. Furthermore, the target radar signal recognition model obtained through training can be used for classifying and recognizing the sample radar signals based on radar signal feature maps with different scales and resolutions, and the classification accuracy of the model can be improved.

In some embodiments, the initial radar signal identification model adopts a leakage ReLU activation function, so that the problem of neuronal death can be solved, and in the back propagation process, the leakage ReLU can be calculated to obtain the gradient for the part with the input of the ReLU activation function smaller than zero, so that the problem of gradient direction saw teeth is avoided.

Step 130, determining a loss value according to a signal type prediction result of the sample radar signal in the sample set and a sample tag of the sample radar signal in the sample set; the loss value is determined according to a target loss function, and the target loss function can adjust the loss value corresponding to the sample radar signal according to the identification difficulty of the sample radar signal.

The signal type prediction result of the sample radar signal is the radar signal classification result output by the initial radar signal identification model. For example, the signal type prediction result output by the initial radar signal identification model may be a signal type probability sequence including probabilities for each sample radar signal. For example, the signal type prediction result output by the initial radar signal identification model is: 0.1,0.03,0.6,0.2,0,0.03,0.04, the third ranked probability of the signal type prediction results (i.e. the signal type probability sequence) is the maximum value, which indicates that the signal type prediction result output by the initial radar signal identification model is the fourth radar signal (i.e. the signal type corresponding to the third ranked probability). The sample tag of the sample radar signal may represent the true signal type of the sample radar signal.

In this embodiment, after a signal type prediction result of a sample radar signal in a sample set is obtained according to the sample radar signal in the sample set and an initial radar signal identification model, the signal type prediction result of the sample radar signal is used as initial training output of the initial radar signal identification model, and a sample tag of the sample radar signal is used as supervision information to perform iterative training on the initial radar signal identification model to obtain a target radar signal identification model.

In some embodiments, taking a signal type prediction result of a sample radar signal as an initial training output of an initial radar signal recognition model, taking a sample tag of the sample radar signal as supervision information, and performing iterative training on the initial radar signal recognition model to obtain a target radar signal recognition model, may include: determining a loss value according to a signal type prediction result of the sample radar signal in the sample set and a sample tag of the sample radar signal in the sample set; and iteratively updating the initial radar signal identification model according to the loss value to obtain the target radar signal identification model.

In this embodiment, the loss value may be determined according to a target loss function, and the target loss function may adjust the loss value corresponding to the sample radar signal according to the difficulty in identifying the sample radar signal. More specifically, the more the signal type prediction result of the sample radar signal in the sample set is close to 1, the easier the sample radar signal is classified. The more the signal type prediction result of the sample radar signal in the sample set is close to 0, the more difficult the sample radar signal is to classify.

In some embodiments, the loss value is obtained by the following equation (5):wherein (1)>For loss value, +_>The result is predicted for the signal type of the sample radar signal in the sample set.

In the present embodiment, sinceI.e. +.>Thus, the objective loss function of the embodiments of the present application (i.e., equation (5) above) employs a cosine function (i.e., +.>) Instead of focusing on the coefficient gamma in the existing cross entropy loss function, the target loss function of the embodiment of the application can automatically adjust the loss value of different sample radar signals based on the characteristic of second order minimality of the cosine function without adjusting the focusing on the coefficient gamma. That is, when the signal type prediction result of the sample radar signal (i.e.) The closer to 1, i.e., the easier the sample radar signal is to classify, the closer to 0 the loss value. When the signal type of the sample radar signal predicts (i.e.)>) The closer to 0, i.e. the more difficult it is to classify the sample radar signal, the greater the loss value. In this way, a loss value is calculated based on the target loss function, and then the initial radar signal identification model is iteratively trained through the loss value, The trained target radar signal recognition model is focused on the learning and classification of difficult samples, so that the performance of the target radar signal recognition model is improved, and the recognition rate of radar signals is improved.

In some embodiments, the loss value is determined from a target loss function and a attention weighting coefficient, the attention weighting coefficients corresponding to different sample radar signals being different.

Illustratively, the loss value is obtained by the following equation (6) and equation (7):（6）

wherein the method comprises the steps ofFor loss value, +_>Attention weight coefficient corresponding to sample radar signal in sample set, < ->The result is predicted for the signal type of the sample radar signal in the sample set.

（7）

Wherein,attention weight coefficients corresponding to the sample radar signals in the sample set; q represents query, K represents key, V represents value, Q, K, V is composed of input matrix X and three trainable parameter matrices +.>Is derived from a linear transformation of (a). />Is a similar matrix +.>The dimension of k.

From the above formulas (6) and (7), the target loss function can be obtained as follows formula (8):

（8）

in the present embodiment, whenWhen larger, the similarity matrix ∈>The variance of (2) is increased, the difference of different dimensions can be reduced by performing scaling processing according to the formula (8), so that gradient update is more stable in model training, and further, after normalization processing is performed through a softmax function, an attention weight coefficient larger than 0 and smaller than 1 can be generated for each sample radar signal. That is, the attention mechanism is added to the target Loss function to assist network training, and meanwhile, the traditional cross entropy Loss function is replaced by improvement on the basis of the Focal Loss function, so that the embodiment of the application does not need to rely on a large number of training samples, and under the condition that a sample set comprises less sample data, the weight occupied by difficult samples can be redistributed, so that the trained target radar signal recognition model focuses more on the learning and classification of the difficult samples, the performance of the target radar signal recognition model is improved, the problem that the conventional radar signal recognition model is easy to be subjected to fitting when the training samples are insufficient is solved, the classification precision of the target radar signal recognition model is improved, and the recognition rate of the target radar signal recognition model is improved.

And 140, iteratively updating the initial radar signal identification model according to the loss value to obtain a target radar signal identification model.

The target radar signal recognition model can be obtained by training an initial radar signal recognition model, and the target radar signal recognition model can be used for classifying and recognizing the radar signal to be recognized.

In some embodiments, the step of iteratively updating the initial radar signal identification model according to the loss value to obtain the target radar signal identification model may further include: iteratively updating the initial radar signal recognition model according to the loss value to obtain a trained radar signal recognition model; and under the condition that the first preset condition is met, determining the trained radar signal recognition model as a target radar signal recognition model.

The first preset condition may be an end condition for judging whether the model completes the iterative training. The first preset condition may be determined by whether the initial radar signal identification model converges.

In some examples, the first preset condition may be that the loss value is less than or equal to a preset threshold. And under the condition that the loss value is smaller than or equal to a preset threshold value, the model is converged, namely the initial radar signal identification model completes iterative training. It should be noted that, the preset threshold may be set according to the complexity of the radar signal to be identified, and the specific numerical value of the preset threshold is not limited in the embodiment of the present application.

In some examples, the first preset condition may be that the number of iterations of the initial radar signal identification model reaches a preset number. And under the condition that the iteration times of the initial radar signal recognition model reach the preset times, the model is proved to be converged, namely the initial radar signal recognition model completes the iteration training. It should be noted that the preset number of times may be set according to the complexity of the radar signal to be identified, for example, the preset number of times is 300. Specific numerical values of the preset times are not limited in the embodiment of the present application.

In some embodiments, iteratively updating the initial radar signal identification model based on the loss value may further comprise: step 210-step 220.

Step 210, in the case that the training period of the initial radar signal identification model does not reach the second preset period, iteratively updating the initial radar signal identification model according to the loss value based on the first learning rate.

That is, the initial radar signal identification model is iteratively updated based on the first learning rate when the training period of the initial radar signal identification model does not reach the second preset period, that is, the preceding second preset period. The second preset period may be 10 periods, in this embodiment of the present application, the second preset period may be set according to an actual situation, and specific numerical values of the second preset period are not limited in this embodiment of the present application.

Step 220, under the condition that the training period of the initial radar signal recognition model reaches a second preset period, iteratively updating the initial radar signal recognition model according to the loss value based on a second learning rate, wherein the second learning rate is smaller than the first learning rate.

That is, the learning rate is gradually reduced when the training period of the initial radar signal identification model reaches the second preset period, that is, the subsequent training period, and the radar signal identification model is continuously iteratively updated at the adjusted learning rate.

In some examples, the second learning rate may be obtained by the following equation (9):

（9）

wherein,for the second learning rate, n is the difference between the current training period and the second preset period,/>Is the first learning rate.

For example, the second preset period is taken as 10 training periods, and the learning rate of model training is the first learning rate in the first 10 training periodsThe method comprises the steps of carrying out a first treatment on the surface of the In the 11 th training period, the learning rate of model training is +.>The method comprises the steps of carrying out a first treatment on the surface of the In the 12 th training period, the learning rate of model training is +.>The method comprises the steps of carrying out a first treatment on the surface of the In the 13 th training period, the learning rate of model training is +.>. That is, in the latter period of training, the learning rate of each training period is attenuated by 10% with respect to the learning rate of the preceding training period.

In this embodiment, when training a model, the learning rate may be dynamically adjusted, a larger learning rate is adopted at the initial stage of model training, and the learning rate may be gradually reduced at the later stage of model training, so that the optimal solution of the model may be quickly found in the model training process, and simultaneously, with deepening of model training, the learning rate may be gradually reduced, so that model training efficiency may be improved, and the model training may be prevented from being fitted.

In some embodiments, to further ensure performance of the model, after iteratively updating the initial radar signal identification model according to the loss value, the trained radar signal identification model may be validated using a validation set, and when the trained radar signal identification model meets the requirements, the trained radar signal identification model is determined to be the target radar signal identification model.

In some examples, to solve the problem that the existing model training method needs to rely on a large number of samples, a K-fold verification method may be used to verify the trained radar signal identification model. In this embodiment, the sample set may be divided into N sample subsets, wherein each sample subset of the N sample subsets includes at least two types of sample radar signals, and sample tags of each sample radar signal.

Taking the sample set including N sample subsets as an example, the process of training an initial radar signal identification model and verifying the trained radar signal identification model by adopting a K-fold verification method is described below. Referring to fig. 6, the process includes: step 210-step 270.

Step 210, using any sample subset of the N sample subsets as a test sample subset, and using sample subsets other than the test sample subset of the N sample subsets as training sample subsets.

Step 220, obtaining a signal type prediction result of the sample radar signals in the training sample subset according to the sample radar signals in the training sample subset and the initial radar signal identification model.

Step 230, determining a loss value according to the signal type prediction result of the sample radar signal in the training sample subset and the sample label of the sample radar signal in the training sample subset.

And step 240, iteratively updating the initial radar signal recognition model according to the loss value to obtain a trained radar signal recognition model.

That is, the sample set is divided into N sample subsets, each sample subset of the N sample subsets is sequentially used as a test sample subset, the initial radar signal recognition model is iteratively trained by adopting other sample subsets except the test sample subset of the N sample subsets, a trained radar signal recognition model is obtained, the trained radar signal recognition model is verified by adopting the test sample subset, and whether the radar signal recognition model finishes training is judged according to the verification score.

It should be noted that, for a specific implementation manner of performing iterative training on the initial radar signal identification model by using the sample subsets other than the test sample subset among the N sample subsets, reference may be made to the specific implementation manners of the foregoing steps 120 to 140, and in order to avoid repetition, a description thereof is omitted here.

Step 250, obtaining a signal type prediction result of the sample radar signals in the test sample subset according to the sample radar signals in the test sample subset and the trained radar signal identification model.

Step 260, determining the verification score according to the signal type prediction result of the sample radar signal in the test sample subset and the sample label of the sample radar signal in the test sample subset.

After the initial radar signal recognition model is iteratively trained by adopting other sample subsets except the test sample subset in the N sample subsets to obtain a trained radar signal recognition model, the test sample subset can be adopted to verify the trained radar signal recognition model, and then whether the radar signal recognition model finishes training is judged according to the verification score.

In some examples, the verification score may be a recognition accuracy rate of the trained radar signal recognition model, or may be a recognition error rate of the trained radar signal recognition model.

Taking verification score as recognition accuracy of the trained radar signal recognition model as an example, the calculation process of the verification score is described. Assume that the sample set up includes a total of 350 sample radar signals, wherein each sample subset includes 50 sample radar signals, i.e., the test sample subset includes 50 sample radar signals. And (3) verifying the trained radar signal identification model by adopting the test sample subset, wherein if 49 sample radar signals are correctly identified, the verification score is the ratio of the number of the correctly identified sample radar signals to the total sample radar signals of the sample set, namely, the verification score is 49/350.

Step 270, determining the trained radar signal recognition model as the target radar signal recognition model under the condition that the verification score meets the second preset condition.

In some embodiments, in the case that the verification score is an identification accuracy of the trained radar signal identification model, the verification score satisfies a second preset condition, that is, the verification score is greater than or equal to the first threshold. It should be noted that, the first threshold may be set according to actual situations, and specific values of the first threshold are not limited in this embodiment of the present application.

In some embodiments, in the case that the verification score is an identification error rate of the trained radar signal identification model, the verification score satisfies a second preset condition, i.e. the verification score is less than or equal to a second threshold value. It should be noted that, the second threshold may be set according to actual situations, and specific values of the second threshold are not limited in this embodiment of the present application.

In some embodiments, to prevent the model training from overfitting, the step of determining the verification score based on the signal type predictions of the sample radar signals in the subset of test samples and the sample tags of the sample radar signals in the subset of test samples may further comprise: under the condition that the training period of the initial radar signal identification model reaches a first preset period, determining the current verification score according to the signal type prediction result of the sample radar signals in the test sample subset and the sample labels of the sample radar signals in the test sample subset. The step of determining the trained radar signal identification model as the target radar signal identification model may further include: and under the condition that the current verification score is smaller than or equal to the last verification score, determining the trained radar signal recognition model as a target radar signal recognition model.

In this embodiment, when the training period of the initial radar signal identification model reaches the first preset period, it is explained that the performance of the model has reached a certain degree of optimization, at this time, the current verification score is obtained, the current verification score is compared with the previous verification score, if the current verification score is smaller than or equal to the previous verification score, it is explained that the performance of the model is not further improved, that is, the model meets the requirement, at this time, the model training is stopped, and the trained radar signal identification model obtained at this time is determined as the target radar signal identification model.

The current verification score may be a verification score obtained by testing the trained radar signal identification model after the current training period. The last verification score may be a verification score obtained by testing the trained radar signal identification model after the last training period. The first preset period may be 10, that is, after 10 training periods, the model training is stopped by comparing the front and rear verification scores to stop the model training under the condition that the model performance is no longer improved, that is, the trained radar signal recognition model obtained at this time is determined as the target radar signal recognition model.

Taking the initial radar signal recognition model including 4 sub-models and the sample set including 5 sample subsets (i.e., sample subset a, sample subset b, sample subset c, sample subset d, and sample subset e) as an example, the process of training the initial radar signal recognition model and verifying the trained radar signal recognition model by using the K-fold verification method will be described.

Referring to fig. 7, in the first stage (i.e., the first fold): sample subset a is used as a test sample subset, and sample subset b, sample subset c, sample subset d and sample subset e are used as training sample subsets. And training the initial radar signal recognition model by adopting each training sample subset respectively to obtain the radar signal recognition model after the first training. And then, testing the radar signal identification model after the first fold training by adopting the sample subset a to obtain a first verification score.

Second stage (i.e. second fold): sample subset b was used as the test sample subset, and sample subset a, sample subset c, sample subset d, and sample subset e were used as the training sample subsets. And respectively carrying out iterative training on the radar signal recognition model after the first training by adopting each training sample subset to obtain the radar signal recognition model after the second training. And then, testing the radar signal recognition model after the second training by adopting the sample subset b to obtain a second verification score.

Third stage (i.e. third fold): sample subset c is taken as a test sample subset, and sample subset a, sample subset b, sample subset d and sample subset e are taken as training sample subsets. And respectively carrying out iterative training on the radar signal recognition model after the second training by adopting each training sample subset to obtain the radar signal recognition model after the third training. And then, testing the radar signal recognition model after the third training by adopting the sample subset c to obtain a third verification score.

Fourth stage (i.e. fourth fold): sample subset d was used as the test sample subset, and sample subset a, sample subset b, sample subset c, and sample subset e were used as the training sample subsets. And respectively adopting each training sample subset to carry out iterative training on the radar signal recognition model after the third training to obtain the radar signal recognition model after the fourth training. And then, testing the radar signal recognition model after the fourth training by adopting the sample subset d to obtain a fourth verification score.

Fifth stage (i.e. fifth fold): sample subset e was used as the test sample subset, and sample subset a, sample subset b, sample subset c, and sample subset d were used as the training sample subsets. And respectively carrying out iterative training on the radar signal recognition model after the fourth training by adopting each training sample subset to obtain the radar signal recognition model after the fifth training. And then, testing the radar signal recognition model after the fifth folding training by adopting the sample subset e to obtain a fifth verification score.

And then determining the verification score according to the first verification score, the second verification score, the third verification score, the fourth verification score and the fifth verification score, and determining the radar signal recognition model after the fifth training as a target radar signal recognition model under the condition that the verification score meets a second preset condition.

Illustratively, the verification score may be obtained by the following equation (10):（10）

wherein,to verify score, ++>For the first verification score, ++>For the second verification score, ++>For the third verification score,>for the fourth verification score, ++>Is the fifth verification score.

In the embodiment of the application, the K-fold verification method is adopted to train the initial radar signal recognition model, and the trained radar signal recognition model is verified, so that the method can be applied to scenes with smaller sample data size, the evaluation model performance is more objective and reliable, the performance of the target radar signal recognition model is ensured, and the recognition capability of a small sample of the target radar signal recognition model is improved.

The following describes a specific example of a training method of the radar signal identification model according to the embodiment of the present application.

Taking three types of sample radar signals including a linear frequency modulation signal (i.e. LFM), a binary phase coding signal (BPSK) and a binary frequency coding signal (BFSK) as examples, designing a plurality of sample sets, wherein the value range of the number of samples of each type of sample radar signal in the sample sets is 50-140, and the value range of the signal-to-noise ratio of the sample radar signal is-5 dB. And respectively carrying out iterative training on the initial radar signal recognition model by adopting each sample set to obtain a target radar signal recognition model. And testing the target radar signal identification model, wherein the test results are shown in the following table 1.

TABLE 1 verification scores (recognition accuracy) for different signal-to-noise ratios and different sample numbers

As can be seen from Table 1, the recognition accuracy of the target radar signal recognition model obtained by training based on the embodiment of the application is not affected by the number of samples and the signal-to-noise ratio of the sample radar signal, and the recognition accuracy of the target radar signal recognition model obtained by training in the embodiment of the application in a noise-free and high-noise environment is high, and the classification accuracy can exceed 90%. In addition, referring to fig. 8, a graph of the training algebra and the recognition accuracy of the model is shown, and as can be seen from fig. 8, the recognition accuracy of the model tends to be stable and can reach more than 80% after the training algebra (i.e. training period) reaches 10 periods.

That is, the target radar signal recognition model obtained through training in the embodiment of the application has good performance and high classification precision, and can be suitable for small sample scenes.

Fig. 9 is a schematic flow chart of a radar signal identification method according to an embodiment of the present application. The method may be performed by a server. As shown in FIG. 9, the method may include steps 310-320.

Step 310, a radar signal to be identified is acquired.

Step 320, determining the signal type of the radar signal to be identified according to the radar signal to be identified and the target radar signal identification model.

The target radar signal recognition model may be iteratively trained by the training method of the radar signal recognition model provided in the foregoing embodiment.

In some embodiments, the step of determining the signal type of the radar signal to be identified according to the radar signal to be identified and the target radar signal identification model may further comprise: processing the radar signal to be identified according to at least two convolution layers to obtain a target radar signal characteristic diagram; and processing the target radar signal feature map according to the classification layer to obtain the signal type of the radar signal to be identified.

That is, the radar signal to be identified is used as the input of the target radar signal identification model, and is sequentially processed by at least two convolution layers, so as to obtain the target radar signal feature map. And then, the classifying layer classifies the target radar signal feature map to obtain the signal type of the radar signal to be identified.

In some embodiments, the at least two convolution layers include three first convolution layers, three second convolution layers, three third convolution layers, and three fourth convolution layers. That is, the target radar signal identification model may include twelve convolutional layers (i.e., twelve one-dimensional convolutional neural networks). The size of the convolution kernels of the second convolution layer is the same as that of the convolution kernels of the first convolution layer, the number of the convolution kernels of the second convolution layer is larger than that of the convolution kernels of the first convolution layer, the size of the convolution kernels of the third convolution layer is larger than that of the convolution kernels of the second convolution layer, the number of the convolution kernels of the third convolution layer is equal to that of the convolution kernels of the second convolution layer, the size of the convolution kernels of the fourth convolution layer is the same as that of the convolution kernels of the third convolution layer, and the number of the convolution kernels of the fourth convolution layer is larger than that of the convolution kernels of the third convolution layer.

In this embodiment, the step of processing the radar signal to be identified according to at least two convolution layers to obtain the target radar signal feature map may further include: processing radar signals to be identified according to the three first convolution layers to obtain a first radar signal characteristic diagram; processing the first radar signal feature map according to the three second convolution layers to obtain a second radar signal feature map; processing the second radar signal feature map according to the three third convolution layers to obtain a third radar signal feature map; and processing the third radar signal feature map according to the three fourth convolution layers to obtain a target radar signal feature map.

That is, the radar signal to be identified is used as the input of the first convolution layers in the target radar signal identification model, and is sequentially processed through the three first convolution layers, so as to obtain a first radar signal characteristic diagram. And then, the first radar signal characteristic map is used as the input of a second convolution layer in the target radar signal identification model, and is sequentially processed by the three second convolution layers to obtain a second radar signal characteristic map. And then, the second radar signal characteristic map is used as the input of a third convolution layer in the target radar signal identification model, and is sequentially processed by the three third convolution layers to obtain the third radar signal characteristic map. And finally, the third radar signal characteristic diagram is used as the input of a fourth convolution layer in the target radar signal identification model, and is processed through the three fourth convolution layers in sequence to obtain the target radar signal characteristic diagram.

In some embodiments, before determining the signal type of the radar signal to be identified from the radar signal to be identified and the target radar signal identification model, the method may further include: preprocessing the radar signal to be identified. Illustratively, the radar signal to be identified is normalized to obtain a one-dimensional radar signal sequence. That is, after the radar signal to be identified is obtained, the radar signal to be identified is normalized to obtain a one-dimensional radar signal sequence, and then the one-dimensional radar signal sequence is used as the input of the target radar signal identification model, i.e. the one-dimensional radar signal sequence is sequentially processed by at least two convolution layers. In this embodiment, after the radar signal to be identified is obtained, the radar signal to be identified is preprocessed, which is conducive to improving generalization performance and robustness of the model, and further improving accuracy of radar signal classification.

According to the embodiment of the application, the target loss function of the embodiment of the application can automatically adjust the loss value of different sample radar signals without adjusting the concentration coefficient gamma. Therefore, the loss value is calculated based on the target loss function, and further the initial radar signal recognition model is subjected to iterative training through the loss value, so that the trained target radar signal recognition model is more focused on the learning and classification of difficult samples, and further the performance of the target radar signal recognition model is improved. Furthermore, in the process of identifying radar signals, the target radar signal identification model obtained through training in the embodiment of the application is adopted to carry out classification identification on the radar signals to be identified, so that classification accuracy can be improved.

Referring to fig. 10, the embodiment of the application further provides a training device 1000 for a radar signal recognition model, as shown in fig. 10, the training device 1000 includes: an acquisition module 1001, a prediction module 1002, a determination module 1003, and a training module 1004.

Wherein the obtaining module 1001 may be configured to obtain a sample set, where the sample set includes at least two types of sample radar signals, and a sample tag of each of the at least two types of sample radar signals; the prediction module 1002 may be configured to obtain a signal type prediction result of the sample radar signal in the sample set according to the sample radar signal in the sample set and the initial radar signal identification model; the determining module 1003 may be configured to determine a loss value according to a signal type prediction result of the sample radar signal in the sample set and a sample tag of the sample radar signal in the sample set; the loss value is determined according to a target loss function, and the target loss function can adjust the loss value corresponding to the sample radar signal according to the identification difficulty of the sample radar signal; training module 1004 may be configured to iteratively update the initial radar signal identification model according to the loss value to obtain the target radar signal identification model.

In some embodiments, the above-described loss value is obtained by the following formula:wherein (1)>For loss value, +_>The result is predicted for the signal type of the sample radar signal in the sample set.

In some embodiments, the loss value is determined according to a target loss function and attention weighting coefficients, and the attention weighting coefficients corresponding to different sample radar signals are different.

In some embodiments, the above-described loss value is obtained by the following formula:wherein (1)>For loss value, +_>Attention weight coefficient corresponding to sample radar signal in sample set, < ->The result is predicted for the signal type of the sample radar signal in the sample set.

In some embodiments, training module 1004 may include: the training unit can be used for iteratively updating the initial radar signal identification model according to the loss value to obtain a trained radar signal identification model; the determining unit can be used for determining the trained radar signal recognition model as a target radar signal recognition model under the condition that the first preset condition is met; the first preset condition comprises: the loss value is less than or equal to a preset threshold; or the iteration times of the initial radar signal identification model reach the preset times.

In some embodiments, the sample set comprises N sample subsets, each sample subset of the N sample subsets comprising at least two types of sample radar signals, and a sample tag for each sample radar signal; the training apparatus 1000 may further include: the dividing module can be used for taking any sample subset of the N sample subsets as a test sample subset and taking sample subsets except the test sample subset of the N sample subsets as training sample subsets; the prediction module 1002 may be specifically configured to obtain a signal type prediction result of the sample radar signal in the training sample subset according to the sample radar signal in the training sample subset and the initial radar signal identification model; the determining module 1003 may be specifically configured to determine the loss value according to a signal type prediction result of the sample radar signal in the training sample subset and a sample tag of the sample radar signal in the training sample subset.

In some embodiments, the determining unit may be specifically configured to: obtaining a signal type prediction result of the sample radar signals in the test sample subset according to the sample radar signals in the test sample subset and the trained radar signal identification model; determining a verification score according to a signal type prediction result of the sample radar signals in the test sample subset and sample labels of the sample radar signals in the test sample subset; and under the condition that the verification score meets the second preset condition, determining the trained radar signal identification model as a target radar signal identification model.

In some embodiments, the determining unit may be specifically configured to: under the condition that the training period of the initial radar signal identification model reaches a first preset period, determining a current verification score according to a signal type prediction result of the sample radar signals in the test sample subset and sample labels of the sample radar signals in the test sample subset; and under the condition that the current verification score is smaller than or equal to the last verification score, determining the trained radar signal recognition model as a target radar signal recognition model.

In some embodiments, the initial radar signal identification model includes at least two convolution layers and a classification layer, and the prediction module 1002 may be specifically configured to: processing the sample radar signals in the sample set according to at least two convolution layers to obtain a target radar signal characteristic diagram; and processing the target radar signal feature map according to the classification layer to obtain a signal type prediction result of the sample radar signal in the sample set.

In some embodiments, the at least two convolution layers include three first convolution layers, three second convolution layers, three third convolution layers, and three fourth convolution layers; the size of the convolution kernels of the second convolution layer is the same as that of the convolution kernels of the first convolution layer, the number of the convolution kernels of the second convolution layer is larger than that of the convolution kernels of the first convolution layer, the size of the convolution kernels of the third convolution layer is larger than that of the second convolution layer, the number of the convolution kernels of the third convolution layer is equal to that of the second convolution layer, the size of the convolution kernels of the fourth convolution layer is the same as that of the convolution kernels of the third convolution layer, and the number of the convolution kernels of the fourth convolution layer is larger than that of the third convolution layer; processing the sample radar signals in the sample set according to at least two convolution layers to obtain a target radar signal feature map, wherein the processing comprises the following steps: processing the sample radar signals in the sample set according to the three first convolution layers to obtain a first radar signal characteristic diagram; processing the first radar signal feature map according to the three second convolution layers to obtain a second radar signal feature map; processing the second radar signal feature map according to the three third convolution layers to obtain a third radar signal feature map; and processing the third radar signal feature map according to the three fourth convolution layers to obtain a target radar signal feature map.

In some embodiments, training module 1004 may be specifically configured to: under the condition that the training period of the initial radar signal identification model does not reach the second preset period, iteratively updating the initial radar signal identification model according to the loss value based on the first learning rate; and under the condition that the training period of the initial radar signal recognition model reaches a second preset period, iteratively updating the initial radar signal recognition model according to the loss value based on a second learning rate, wherein the second learning rate is smaller than the first learning rate.

Referring to fig. 11, the embodiment of the application further provides a radar signal identifying apparatus 1100, as shown in fig. 11, the radar signal identifying apparatus 1100 includes: an acquisition module 1101 and an identification module 1102. The obtaining module 1101 may be configured to obtain a radar signal to be identified; the recognition module 1102 may be configured to determine a signal type of the radar signal to be recognized according to the radar signal to be recognized and a target radar signal recognition model, where the target radar signal recognition model is obtained by training by the training device of the radar signal recognition model provided according to the foregoing embodiment.

Referring to fig. 12, an embodiment of the present application further provides an electronic device, including: the processor 1201, the memory 1202, and executable instructions stored on the memory 1202 and executable on the processor 1201, the processor 1201 when executing the executable instructions implements the method of training the radar signal recognition model provided by any one of the above embodiments, or the method of radar signal recognition provided by the above embodiments.

The embodiment of the application also provides electronic equipment, which comprises the training device of the radar signal identification model.

The embodiment of the application also provides electronic equipment, which comprises the radar signal identifying device.

In some embodiments, the electronic device may be a server.

The embodiment of the application also provides a computer readable storage medium, wherein at least one executable instruction is stored in the storage medium, and when the executable instruction runs on the electronic device, the electronic device is caused to execute the operation of the training method of the radar signal identification model or execute the operation of the radar signal identification method.

The algorithms or displays presented herein are not inherently related to any particular computer, virtual system, or other apparatus. In addition, embodiments of the present application are not directed to any particular programming language.

In the description provided herein, numerous specific details are set forth. It is understood, however, that embodiments of the present application may be practiced without these specific details. Similarly, in the above description of exemplary embodiments of the application, various features of embodiments of the application are sometimes grouped together in a single embodiment, figure, or description thereof for the purpose of streamlining the application and aiding in the understanding of one or more of the various inventive aspects. Wherein the claims following the detailed description are hereby expressly incorporated into this detailed description, with each claim standing on its own as a separate embodiment of this application.

Those skilled in the art will appreciate that the modules in the apparatus of the embodiments may be adaptively changed and disposed in one or more apparatuses different from the embodiments. The modules or units or components of the embodiments may be combined into one module or unit or component and, furthermore, they may be divided into a plurality of sub-modules or sub-units or sub-components. Except that at least some of such features and/or processes or elements are mutually exclusive.

It should be noted that the above-mentioned embodiments illustrate rather than limit the application, and that those skilled in the art will be able to design alternative embodiments without departing from the scope of the appended claims. In the claims, any reference signs placed between parentheses shall not be construed as limiting the claim. The word "comprising" does not exclude the presence of elements or steps not listed in a claim. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. The application may be implemented by means of hardware comprising several distinct elements, and by means of a suitably programmed computer. In the unit claims enumerating several means, several of these means may be embodied by one and the same item of hardware. The use of the words first, second, third, etc. do not denote any order. These words may be interpreted as names. The steps in the above embodiments should not be construed as limiting the order of execution unless specifically stated.

Claims

1. A method of training a radar signal recognition model, the method comprising:

acquiring a sample set, wherein the sample set comprises at least two types of sample radar signals and sample tags of each sample radar signal in the at least two types of sample radar signals;

obtaining a signal type prediction result of the sample radar signals in the sample set according to the sample radar signals in the sample set and the initial radar signal identification model;

determining a loss value according to a signal type prediction result of the sample radar signals in the sample set and a sample tag of the sample radar signals in the sample set; the loss value is determined according to a target loss function, and the target loss function can adjust the loss value corresponding to the sample radar signal according to the identification difficulty of the sample radar signal;

iteratively updating the initial radar signal identification model according to the loss value to obtain a target radar signal identification model;

wherein the loss value is obtained by the following formula:

wherein IFL (p _t ) For the loss value, p _t Predicting a result for a signal type of the sample radar signal in the sample set.

2. The method of claim 1, wherein the loss value is determined based on the target loss function and a attention weighting coefficient, the attention weighting coefficients corresponding to different ones of the sample radar signals being different.

3. The method of claim 2, wherein the loss value is obtained by the formula:

wherein IFL (p _t ) For the loss value, attention (Q, K, V) is the Attention weight coefficient corresponding to the sample radar signal in the sample set, p _t Predicting a result for a signal type of the sample radar signal in the sample set.

4. The method according to claim 1, wherein iteratively updating the initial radar signal identification model according to the loss value results in a target radar signal identification model, comprising:

iteratively updating the initial radar signal recognition model according to the loss value to obtain a trained radar signal recognition model;

under the condition that a first preset condition is met, determining the trained radar signal recognition model as the target radar signal recognition model;

wherein the first preset condition includes:

the loss value is smaller than or equal to a preset threshold value; or,

And the iteration times of the initial radar signal identification model reach preset times.

5. The method of claim 4, wherein the sample set comprises N sample subsets, each sample subset of the N sample subsets comprising at least two types of sample radar signals, and a sample tag for each sample radar signal; after the acquisition of the sample set, the method further comprises:

taking any sample subset in the N sample subsets as a test sample subset, and taking sample subsets except the test sample subset in the N sample subsets as training sample subsets;

the obtaining a signal type prediction result of the sample radar signals in the sample set according to the sample radar signals in the sample set and the initial radar signal identification model comprises the following steps:

obtaining a signal type prediction result of the sample radar signals in the training sample subset according to the sample radar signals in the training sample subset and the initial radar signal identification model;

the determining a loss value according to the signal type prediction result of the sample radar signal in the sample set and the sample label of the sample radar signal in the sample set includes:

And determining a loss value according to the signal type prediction result of the sample radar signals in the training sample subset and the sample labels of the sample radar signals in the training sample subset.

6. The method of claim 5, wherein said determining the trained radar signal identification model as the target radar signal identification model comprises:

obtaining a signal type prediction result of the sample radar signals in the test sample subset according to the sample radar signals in the test sample subset and the trained radar signal identification model;

determining a verification score according to a signal type prediction result of the sample radar signals in the test sample subset and sample labels of the sample radar signals in the test sample subset;

and under the condition that the verification score meets a second preset condition, determining the trained radar signal identification model as the target radar signal identification model.

7. The method of claim 6, wherein determining the verification score based on the signal type prediction result of the sample radar signals in the subset of test samples and the sample tags of the sample radar signals in the subset of test samples comprises:

Under the condition that the training period of the initial radar signal identification model reaches a first preset period, determining a current verification score according to a signal type prediction result of sample radar signals in the test sample subset and sample labels of the sample radar signals in the test sample subset;

and under the condition that the verification score meets a second preset condition, determining the trained radar signal identification model as the target radar signal identification model, wherein the method comprises the following steps of:

and under the condition that the current verification score is smaller than or equal to the last verification score, determining the trained radar signal identification model as the target radar signal identification model.

8. The method of claim 1, wherein the initial radar signal identification model includes at least two convolution layers and a classification layer, and obtaining a signal type prediction result of the sample radar signals in the sample set based on the sample radar signals in the sample set and the initial radar signal identification model includes:

processing the sample radar signals in the sample set according to the at least two convolution layers to obtain a target radar signal feature map;

And processing the target radar signal feature map according to the classification layer to obtain a signal type prediction result of the sample radar signals in the sample set.

9. The method of claim 8, wherein the at least two convolution layers comprise three first convolution layers, three second convolution layers, three third convolution layers, and three fourth convolution layers; wherein the size of the convolution kernels of the second convolution layer is the same as the size of the convolution kernels of the first convolution layer, the number of convolution kernels of the second convolution layer is greater than the number of convolution kernels of the first convolution layer, the size of the convolution kernels of the third convolution layer is greater than the size of the convolution kernels of the second convolution layer, the number of convolution kernels of the third convolution layer is equal to the number of convolution kernels of the second convolution layer, the size of the convolution kernels of the fourth convolution layer is the same as the size of the convolution kernels of the third convolution layer, and the number of convolution kernels of the fourth convolution layer is greater than the number of convolution kernels of the third convolution layer;

the processing the sample radar signals in the sample set according to the at least two convolution layers to obtain a target radar signal feature map, including:

Processing the sample radar signals in the sample set according to the three first convolution layers to obtain a first radar signal characteristic diagram;

processing the first radar signal feature map according to the three second convolution layers to obtain a second radar signal feature map;

processing the second radar signal feature map according to the three third convolution layers to obtain a third radar signal feature map;

and processing the third radar signal characteristic map according to the three fourth convolution layers to obtain a target radar signal characteristic map.

10. The method of claim 1, wherein iteratively updating the initial radar signal identification model based on the loss values comprises:

iteratively updating the initial radar signal recognition model according to the loss value based on a first learning rate under the condition that the training period of the initial radar signal recognition model does not reach a second preset period;

and under the condition that the training period of the initial radar signal identification model reaches the second preset period, iteratively updating the initial radar signal identification model according to the loss value based on a second learning rate, wherein the second learning rate is smaller than the first learning rate.

11. A method of radar signal identification, the method comprising:

acquiring a radar signal to be identified;

determining the signal type of the radar signal to be identified according to the radar signal to be identified and a target radar signal identification model, wherein the target radar signal identification model is trained according to the training method of the radar signal identification model of any one of claims 1-10.

12. A training device for a radar signal identification model, the device comprising:

an acquisition module configured to acquire a sample set, where the sample set includes at least two types of sample radar signals, and a sample tag of each of the at least two types of sample radar signals;

the prediction module is used for obtaining a signal type prediction result of the sample radar signals in the sample set according to the sample radar signals in the sample set and the initial radar signal identification model;

the determining module is used for determining a loss value according to a signal type prediction result of the sample radar signals in the sample set and a sample tag of the sample radar signals in the sample set; the loss value is determined according to a target loss function, and the target loss function can adjust the loss value corresponding to the sample radar signal according to the identification difficulty of the sample radar signal;

The training module is used for iteratively updating the initial radar signal identification model according to the loss value to obtain a target radar signal identification model;

wherein the loss value is obtained by the following formula:

13. A radar signal identification device, the device comprising:

the acquisition module is used for acquiring radar signals to be identified;

the recognition module is configured to determine a signal type of the radar signal to be recognized according to the radar signal to be recognized and a target radar signal recognition model, where the target radar signal recognition model is obtained by training the training device of the radar signal recognition model according to claim 12.