CN110097011A - A kind of signal recognition method and device - Google Patents

Abstract

Embodiments of the present invention provide a signal identification method and device. The scheme is as follows: the signal to be identified in the preset frequency band can be obtained, the target time-frequency image of the signal to be identified can be determined, the target feature data of the target time-frequency image can be extracted, and the target feature data can be input into the pre-trained signal recognition model to determine the signal to be identified The type of signal, the signal recognition model is a model obtained by training the preset training set, which includes the sample feature data of the sample time-frequency image corresponding to multiple sample signals in the preset frequency band, and the sample type of each sample signal . Through the technical solution provided by the embodiment of the present invention, according to the difference in characteristic data in the time-frequency images corresponding to different signals, the trained signal recognition model is used to determine the type of the signal, so that different types of signals can be identified by using one device, It is no longer necessary to separately equip corresponding modules for each type of signal, which reduces the cost of equipment and improves the efficiency of signal identification.

Description

A signal identification method and device

technical field

The invention relates to the technical field of signal detection, in particular to a signal identification method and device.

Background technique

In the traditional navigation and positioning process, the positioning is often performed by identifying a Global Navigation Satellite System (Global Navigation Satellite System, GNSS) signal. With the continuous development of technology, more positioning methods have emerged, such as all-source navigation, signal-of-opportunity navigation, and multi-source fusion positioning. Using these positioning methods, positioning is often performed by identifying all radio frequency signals that can be used for positioning in the airspace. The radio frequency signal can include various non-navigation dedicated signals, such as digital audio broadcasting, digital TV broadcasting signal, AM and FM broadcasting signal, cellular base station signal, Bluetooth (Bluetooth) signal, ZigBee (ZigBee) signal, wireless network (Wi- Fi) signal, etc. Different from GNSS signals, these radio frequency signals are distributed in a wide frequency band, and the modulation methods adopted by each signal are different, which brings greater difficulty to the signal identification process.

At present, when identifying the above radio frequency signals, different communication modules are used to identify the radio frequency signals corresponding to each communication module, which leads to high equipment cost and low signal identification efficiency.

Contents of the invention

The purpose of the embodiments of the present invention is to provide a signal recognition method and device, so as to reduce equipment cost and improve signal recognition efficiency. The specific technical scheme is as follows:

An embodiment of the present invention provides a signal identification method, including:

Obtain the signal to be identified within the preset frequency band;

determining a target time-frequency image of the signal to be identified;

Extracting target feature data of the target time-frequency image;

Inputting the target feature data into a pre-trained signal recognition model to determine the type of the signal to be recognized, wherein the signal recognition model is a model obtained by training a preset training set, and the preset training set includes the The sample characteristic data of the sample time-frequency images corresponding to the multiple sample signals in the preset frequency band, and the sample type of each sample signal.

Optionally, the step of determining the target time-frequency image of the signal to be identified includes:

Using the following short-time Fourier transform formula, the target time-frequency image of the signal to be identified is obtained:

G _f (w,u)=∫f(t)g(tu)e ^-jwt dt

Wherein, w is the angular frequency of the signal to be identified, f is the frequency, t is the time t, u is the preset time window length u, the value of the function G _f (w, u) is the amplitude of the frequency component, ∫. dt is an integral operation on t, the function f(t) is the signal to be identified, the function g(tu) is a preset window function, the function e ^-jwt is a complex variable function, e is a natural constant, and j is an imaginary number unit.

Optionally, the step of extracting the target feature data in the target time-frequency image includes:

performing feature extraction on the target time-frequency image to obtain a plurality of feature points of the target time-frequency image;

Using the K-means clustering (K-means) algorithm, the plurality of feature points are clustered to obtain K classes;

Determine the number of feature points included in each of the K classes;

According to the number of feature points included in each class, target feature data in the target time-frequency image is determined.

Optionally, the step of performing feature extraction on the target time-frequency image to obtain a plurality of feature points of the target time-frequency image includes:

A Speeded-Up Robust Features (SURF) algorithm is used to extract multiple feature descriptors in the target time-frequency image to obtain multiple feature points of the target time-frequency image.

Optionally, the step of determining the target feature data in the target time-frequency image according to the number of feature points included in each class includes:

Constructing a first feature point distribution map of the target time-frequency image according to the number of feature points included in each class; determining the first feature point distribution map as target feature data in the target time-frequency image; or ,

According to the number of feature points included in each class, the probability of the feature points included in each class appearing in the target time-frequency image is counted, and according to the probability corresponding to each class, the second part of the target time-frequency image is constructed. A feature point distribution map: determining the second feature point distribution map as target feature data in the target time-frequency image.

Optionally, the signal recognition model is trained through the following steps, including:

Obtain the preset training set;

Input multiple sample feature data into the preset machine learning model to determine the type of each sample signal;

calculating a loss value according to the determined type of each sample signal and the sample type of each sample signal;

Determine whether the loss value is less than a preset threshold;

If not, then adjust the parameters of the preset machine learning model, return to perform the step of inputting a plurality of sample characteristic data into the preset machine learning model respectively, and determine the type of each sample signal;

If yes, the preset machine learning model is determined as the signal recognition model.

The embodiment of the present invention also provides a signal identification device, including:

The first acquisition module is used to acquire the signal to be identified in the preset frequency band;

A first determination module, configured to determine the target time-frequency image of the signal to be identified;

An extraction module, configured to extract target feature data of the target time-frequency image;

The second determination module is configured to input the target feature data into a pre-trained signal recognition model to determine the type of the signal to be recognized, wherein the signal recognition model is a model obtained by training a preset training set, so The preset training set includes sample feature data of sample time-frequency images corresponding to a plurality of sample signals in the preset frequency band, and a sample type of each sample signal.

Optionally, the first determining module is specifically configured to use the following short-time Fourier transform formula to obtain the target time-frequency image of the signal to be identified:

G _f (w,u)=∫f(t)g(tu)e ^-jwt dt

Wherein, w is the angular frequency of the signal to be identified, f is the frequency, t is the time t, u is the preset time window length u, the value of the function G _f (w, u) is the amplitude of the frequency component, ∫. dt is an integral operation on t, the function f(t) is the signal to be identified, the function g(tu) is a preset window function, the function e ^-jwt is a complex variable function, e is a natural constant, and j is an imaginary number unit.

Optionally, the extraction module includes:

The extraction submodule is used to perform feature extraction on the target time-frequency image to obtain a plurality of feature points of the target time-frequency image;

The clustering submodule is used to adopt the K-means clustering algorithm to perform clustering processing on the plurality of feature points to obtain K classes;

The first determination submodule is used to determine the quantity of feature points included in each class in the K classes;

The second determination submodule is configured to determine the target feature data in the target time-frequency image according to the number of feature points included in each class.

Optionally, the extracting submodule is specifically configured to extract multiple feature descriptors in the target time-frequency image by using the SURF algorithm to obtain multiple feature points of the target time-frequency image.

Optionally, the second determining submodule is specifically configured to construct a first feature point distribution map of the target time-frequency image according to the number of feature points included in each class; Determined as the target feature data in the target time-frequency image; or according to the number of feature points included in each class, count the probability of the feature points included in each class appearing in the target time-frequency image, and according to each Construct a second feature point distribution map of the target time-frequency image according to the probability corresponding to the class; determine the second feature point distribution map as the target feature data in the target time-frequency image.

Optionally, the device also includes:

A second acquisition module, configured to acquire the preset training set;

The third determining module is used to respectively input a plurality of sample feature data into a preset machine learning model to determine the type of each sample signal;

A calculation module, configured to calculate a loss value according to the determined type of each sample signal and the sample type of each sample signal;

A judging module, configured to judge whether the loss value is less than a preset threshold;

The adjustment module is used to adjust the parameters of the preset machine learning model when the judgment result of the judgment module is No, return to execute the input of multiple sample feature data into the preset machine learning model, and determine each sample the step of the type of signal;

The fourth determination module is configured to determine the preset machine learning model as the signal recognition model when the judgment result of the judgment module is yes.

The embodiment of the present invention also provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory complete communication with each other through the communication bus;

memory for storing computer programs;

The processor is used to implement the steps of any one of the signal recognition methods described above when executing the program stored in the memory.

An embodiment of the present invention also provides a computer-readable storage medium, where a computer program is stored in the computer-readable storage medium, and when the computer program is executed by a processor, the steps of any one of the signal identification methods described above are implemented.

An embodiment of the present invention also provides a computer program product containing instructions, which when run on a computer, causes the computer to execute any one of the signal recognition methods described above.

A signal identification method and device provided by an embodiment of the present invention can obtain a signal to be identified in a preset frequency band, determine a target time-frequency image of the signal to be identified, extract target feature data of the target time-frequency image, and input the target feature data The pre-trained signal recognition model determines the type of the signal to be recognized, wherein the signal recognition model is a model obtained by training a preset training set, and the preset training set includes sample time-frequency images corresponding to multiple sample signals in a preset frequency band The sample characteristic data of , and the sample type of each sample signal. Through the technical solution provided by the embodiment of the present invention, according to the difference in characteristic data in the time-frequency images corresponding to different signals, the trained signal recognition model is used to determine the type of the signal, so that different types of signals can be identified by using one device, It is no longer necessary to separately equip corresponding modules for each type of signal, which reduces the cost of equipment and improves the efficiency of signal recognition.

Of course, implementing any product or method of the present invention does not necessarily need to achieve all the above-mentioned advantages at the same time.

Description of drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings that need to be used in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only These are some embodiments of the present invention. Those skilled in the art can also obtain other drawings based on these drawings without creative work.

FIG. 1 is a schematic flow chart of a signal identification method provided by an embodiment of the present invention;

Figure 2-a is a time-frequency image after short-time Fourier transform of a Wi-Fi signal provided by an embodiment of the present invention;

Fig. 2-b is a time-frequency image after short-time Fourier transform of the Bluetooth signal provided by the embodiment of the present invention;

Fig. 2-c is a kind of time-frequency image after carrying out short-time Fourier transform to ZigBee signal provided by the embodiment of the present invention;

Fig. 3-a is a schematic diagram of the scale space of the target time-frequency image provided by the embodiment of the present invention;

Fig. 3-b is a schematic diagram of determining the main direction of the feature points provided by the embodiment of the present invention;

FIG. 4 is a schematic flowchart of a training method for a signal recognition model provided by an embodiment of the present invention;

Figure 5-a shows that the sample signal provided by the embodiment of the present invention is one of the sample time-frequency images corresponding to the Bluetooth signal;

Figure 5-b shows that the sample signal provided by the embodiment of the present invention is one of the sample time-frequency images corresponding to the Bluetooth signal;

Figure 5-c shows that the sample signal provided by the embodiment of the present invention is one of the sample time-frequency images corresponding to the Bluetooth signal;

FIG. 6 is a schematic structural diagram of a signal identification device provided by an embodiment of the present invention;

FIG. 7 is a schematic structural diagram of an electronic device provided by an embodiment of the present invention.

Detailed ways

The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some, not all, embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

At present, when identifying radio frequency signals, different communication modules are used to identify corresponding types of radio frequency signals according to different communication protocols adopted by different types of radio frequency signals. Since different communication models are independent of each other, it is not conducive to the integration and deep integration of devices, which makes the cost of devices that identify signal types relatively high. In addition, each communication module can only identify signals of the type corresponding to the communication module, resulting in low signal identification efficiency.

In order to solve the above-mentioned problems of high equipment cost and low signal recognition efficiency, an embodiment of the present invention provides a signal recognition method. The method applies to any electronic device including a navigation system or a positioning system. In the method provided by the embodiment of the present invention, the signal to be identified in the preset frequency band can be obtained, the target time-frequency image of the signal to be identified can be determined, the target feature data of the target time-frequency image can be extracted, and the target feature data can be input into the pre-trained The signal recognition model determines the type of the signal to be recognized, wherein the signal recognition model is a model obtained by training a preset training set, and the preset training set includes sample feature data of sample time-frequency images corresponding to multiple sample signals in a preset frequency band , and the sample type of each sample signal.

Through the technical solution provided by the embodiment of the present invention, according to the difference in characteristic data in the time-frequency images corresponding to different signals, the trained signal recognition model is used to determine the type of the signal, so that different types of signals can be identified by using one device, It is no longer necessary to separately equip corresponding modules for each type of signal, which reduces the cost of equipment and improves the efficiency of signal identification.

The embodiments of the present invention will be described below through specific embodiments.

As shown in FIG. 1 , FIG. 1 is a schematic flowchart of a signal identification method provided by an embodiment of the present invention. The method includes the following steps.

Step S101, acquiring signals to be identified within a preset frequency band.

In this step, the electronic device acquires a radio frequency signal within a preset frequency band as the signal to be identified.

The above-mentioned radio frequency signal may include a Bluetooth signal, a Wi-Fi signal, and a ZigBee signal, etc., and the above-mentioned preset frequency band may be a 2.4 gigahertz (GHz) Industrial Scientific Medical (ISM) frequency band, or a 5.0 GHz frequency band. For example, the preset frequency band may be a 2.4GHz frequency band. At a certain moment, the electronic device can obtain a radio frequency signal in the 2.4GHz frequency band as a signal to be identified.

The aforementioned preset frequency bands may be set according to actual application scenarios and user requirements. For example, when broadcasting signals are used for navigation or positioning, the broadcasting signals may include digital audio broadcasting signals, digital television broadcasting signals, FM broadcasting signals, and AM broadcasting signals, etc., and these broadcasting signals are distributed in different frequency bands, then the above preset frequency bands can be is the frequency band corresponding to each broadcast signal. In the embodiment of the present invention, there is no specific limitation on the aforementioned preset frequency bands and the number of preset frequency bands.

Step S102, determining the target time-frequency image of the signal to be identified.

In this step, the signal to be identified obtained in the above step S101 is a time domain signal, and the electronic device can perform a certain conversion process on the signal to be identified, that is, convert the time domain signal into a frequency domain signal, thereby obtaining the signal to be identified The time-frequency image is used as the target recognition image.

In an optional embodiment, the above-mentioned step of determining the target time-frequency image of the signal to be identified can be expressed as:

Use the following short-time Fourier transform formula to obtain the target time-frequency image of the signal to be identified, that is, use the following formula to perform short-time Fourier transform on the signal to be identified to obtain the target time-frequency image of the signal to be identified. Among them, the short-time Fourier formula can be expressed as:

G _f (w,u)=∫f(t)g(tu)e ^-jwt dt

w is the angular frequency of the signal to be identified, f is the frequency, t is the time t, u is the preset time window length u, the value of the function G _f (w, u) is the amplitude of the frequency component, ∫ dt is the The integral operation of , the function f(t) is the signal to be identified, the function g(tu) is the preset window function, the function e ^-jwt is a complex variable function, e is a natural constant, and j is an imaginary number unit.

In the embodiment of the present invention, since the signal to be identified is often a non-stationary random signal, if the signal to be identified is directly subjected to Fourier transform, the result after Fourier transform will have certain errors. In order to improve the accuracy of the target time-frequency image above, short-time Fourier transform is performed on the signal to be identified by using a preset window function. That is, when the signal to be identified is converted from a time domain signal to a frequency domain signal, the signal within the time window is intercepted from the signal to be identified through a smaller time window, and the intercepted signal is Fourier transformed. Since the length of the preset time window is small, the signal intercepted in this time window can be considered as a stationary random signal, and the time window is moved on the signal to be identified to complete the short-time Fourier transform process of the signal to be identified. Then the target time-frequency image can be obtained.

Take Wi-Fi signal, Bluetooth signal and ZigBee signal as examples. After performing short-time Fourier transform on the Wi-Fi signal, Bluetooth signal and ZigBee signal, the time-frequency image shown in Figure 2 can be obtained. Among them, FIG. 2-a is a time-frequency image after short-time Fourier transform of a Wi-Fi signal provided by an embodiment of the present invention. Fig. 2-b is a time-frequency image after short-time Fourier transform of the Bluetooth signal provided by the embodiment of the present invention. Fig. 2-c is a time-frequency image after short-time Fourier transform of the ZigBee signal provided by the embodiment of the present invention. In Figure 2-a, Figure 2-b, and Figure 2-c, the horizontal direction represents the frequency of the signal, and the vertical direction represents time. The black area in the time-frequency image is composed of multiple sampling points, and each sampling point is expressed as The magnitude of the frequency component at the corresponding sampling point.

In an optional embodiment, the above-mentioned preset window function may select different window functions according to actual conditions and user requirements, such as Hanning window, Hamming window, Blackman window and so on.

In an optional embodiment, the preset time window length of the preset window function may be determined according to the time resolution and frequency resolution in the time-frequency diagram. For example, when the length of the preset time window is longer, the intercepted signal to be identified within the preset time window is longer, and the frequency resolution after the short-time Fourier transform is higher and the time resolution is lower. When the length of the preset time window is shorter, the intercepted signal to be identified within the preset time window is also shorter, and the frequency resolution after the short-time Fourier transform is lower and the time resolution is higher.

In the embodiment of the present invention, no specific limitation is set on the above-mentioned preset window function and preset time window.

In an optional embodiment, before determining the target time-frequency image of the signal to be identified, that is, before performing short-time Fourier transform on the signal to be identified, the electronic device may perform certain preprocessing on the signal to be identified. For example, filter processing is performed on the signal to be identified, and noise signals in the signal to be identified are filtered out. In the embodiment of the present invention, the preprocessing process and manner are not specifically limited.

Step S103, extracting target feature data of the target time-frequency image.

In this step, the electronic device may perform feature extraction on the target time-frequency image of the signal to be identified to obtain target feature data.

In an optional embodiment, a preset algorithm or a preset network model may be used to perform feature extraction on the signal to be identified to obtain target feature data of the target time-frequency image. Here, the preset algorithm and the preset network model are not described in detail.

Step S104, inputting target feature data into a pre-trained signal recognition model to determine the type of signal to be recognized.

In this step, the extracted target feature data of the target time-frequency image is input into a pre-trained signal recognition model. According to the target characteristic data, and the structure and parameters of the signal recognition model, the type of the signal to be recognized can be obtained. Wherein, the signal recognition model is a model trained through a preset training set. The preset training set may include sample feature data of sample time-frequency images corresponding to a plurality of sample signals in a preset frequency band, and a sample type of each sample signal.

In one embodiment, when the above-mentioned signal recognition model determines the type of the signal to be recognized, it can determine the difference between the target characteristic data and different samples according to the extracted target characteristic data of the signal to be recognized and the sample characteristic data of the sample signal in the preset training set. The degree of coincidence between the sample feature data of the signal. According to the determined coincidence degree, the sample category of the sample signal corresponding to the sample feature data with the highest coincidence degree with the target feature data is determined as the type of the signal to be identified. Wherein, the coincidence degree can be represented by Mahalanobis distance, Euclidean distance, etc., taking Euclidean distance as an example, if the Euclidean distance is smaller, the above coincidence degree is greater. The greater the Euclidean distance, the smaller the degree of coincidence.

The above-mentioned Wi-Fi signal, Bluetooth signal and ZigBee signal are still taken as examples for illustration. If the above-mentioned signal recognition model is a recognition model for these three signals, then the above-mentioned sample signals are Wi-Fi signals, Bluetooth signals and ZigBee signals, and the above-mentioned sample characteristic data are samples corresponding to Wi-Fi signals, Bluetooth signals and ZigBee signals Feature data in time-frequency images.

In an optional embodiment, for the above step S103, extracting the target feature data of the target time-frequency image may specifically include the following steps.

Step S1031, performing feature extraction on the target time-frequency image to obtain multiple feature points of the target time-frequency image.

In this step, the electronic device may use the aforementioned preset algorithm or preset network model to extract multiple feature points in the target time-frequency image of the signal to be identified, and obtain multiple feature points in the target time-frequency image.

In step S1032, the K-means algorithm is used to perform clustering processing on multiple feature points to obtain K clusters.

In this step, the electronic device may use the K-means algorithm to cluster the extracted multiple feature points of the target time-frequency image to obtain K classes.

Step S1033, determining the number of feature points included in each of the K classes.

In this step, for each of the above K classes, the number of feature points included in the class may be counted, so as to determine the number of feature points included in each of the K classes.

Step S1034, according to the number of feature points included in each class, determine the target feature data in the target time-frequency image.

In an optional embodiment, according to the number of feature points included in each class, the first feature point distribution map of the target time-frequency image is constructed; the first feature point distribution map is determined as the target feature data in the target time-frequency image .

Specifically, the electronic device can construct a distribution map of feature points as the first feature point of the target time-frequency image according to the number of feature points included in each class and the position of each feature point in the above-mentioned target time-frequency image Distribution. According to the first feature point distribution map, the electronic device can determine the target feature data of the target time-frequency image, such as the distribution of feature points in the target time-frequency image, the distribution of each class, and the number of feature points contained in each class. quantity.

In another optional embodiment, according to the number of feature points included in each class, the probability of the feature points included in each class appearing in the target time-frequency image is counted, and according to the probability corresponding to each class, the target time-frequency The second feature point distribution map of the frequency image; the second feature point distribution map is determined as the target feature data in the target time-frequency image.

Specifically, the electronic device can count the probability of feature points included in each class appearing in the target time-frequency image according to the number of feature points included in each class, and construct a feature point distribution map according to the probability corresponding to each class , as the second feature point distribution map of the target time-frequency image. The electronic device can determine the target feature data of the target time-frequency image, such as the distribution of feature points in the target time-frequency image, the distribution of each class and the number of feature points contained in each class, and the feature points contained in each class. The probability of the point appearing, etc.

By extracting the features of the target time-frequency image and clustering the extracted multiple feature points, the target feature data is obtained, which improves the accuracy of the target feature data, thereby improving the signal recognition model to determine the category of the signal to be recognized The accuracy improves the signal recognition efficiency.

In an optional embodiment, in the above step S1031, in the process of performing feature extraction on the target time-frequency image to obtain multiple feature points of the target time-frequency image, the electronic device can use the SURF algorithm to extract the target time-frequency image. Multiple feature descriptors to obtain multiple feature points of the target time-frequency image. Specifically, the following steps may be included:

Step S1031A, determining the integral image corresponding to the target time-frequency image.

In this step, the electronic device may perform integral processing on the target time-frequency image to obtain an integral image corresponding to the target time-frequency image.

In the integral image, if the pixel point at the upper left corner of the integral image is taken as the coordinate origin, the horizontal direction to the right is the X-axis direction, and the vertical downward direction is the Y-axis direction, then the coordinate value is (x, y) The integral pixel value of the pixel at the point is the sum of the pixel values of all the pixel points in the rectangular area from the pixel point to the upper left corner of the integral image, that is, the sum of the pixel values of all the pixel points in the rectangular area formed by the pixel point and the coordinate origin . Specifically, it can be expressed as:

Wherein, I(x,y) is the integral pixel value of the pixel point (x,y), and I( _xi ,y _j ) is the pixel value of the pixel point ( _xi ,y _j ).

Step S1031B, performing convolution processing on the integral image to obtain the scale space corresponding to the target time-frequency image. Among them, the scale space is the representation form of the image at different resolutions.

In an optional embodiment, the electronic device can use the box filter to process the above integral image, that is, the size of the box filter is constantly changed, and the changed box filter is convolved with the integral image to obtain the target The Gaussian pyramid scale space of the time-frequency image is used as the scale space of the above target time-frequency image.

When constructing the above-mentioned Gaussian pyramid scale space, that is, when constructing the scale space of the target time-frequency image, the SURF algorithm uses the Hessian matrix determinant to approximate the image, then the Hessian matrix of the integral pixel value of the pixel point (x, y) H(I(x,y)) can be expressed as:

in, and is the result of the second-order partial derivative of the integrated pixel value I at the pixel point (x, y), where I is the integrated pixel value at the pixel point (x, y).

The SURF algorithm uses the second-order standard Gaussian function g(x, y, σ) to convolve the above-mentioned Hessian matrix determinant approximation image to construct a Gaussian pyramid scale space, that is, to construct the scale space of the target time-frequency image. At this time, for the above-mentioned pixel point (x, y), in the case of a scale of σ, the above-mentioned Hessian matrix can be expressed as:

Among them, L _xx (x,y,σ), L _xy (x,y,σ), L _yy (x,y,σ) are the second-order standard function g(x,y, σ) is the result of convolution of the second-order partial derivative at the pixel point (x, y).

Step S1031C, determining the positions of multiple feature points in the scale space to obtain multiple feature points.

In this step, the position of the pixel corresponding to the local maximum value of the above H(x, y, σ) is selected in the scale space as the position of the feature point to obtain multiple feature points.

Taking FIG. 3-a as an example for illustration, FIG. 3-a is a schematic diagram of a scale space of a target time-frequency image provided by an embodiment of the present invention. Wherein, the image 301 is an image obtained by convolving the target time-frequency image with the above-mentioned second-order standard Gaussian function when the scale is σ _i , and the pixel 302 is a pixel in the image 301 . Above and below the image 301 there are convolved images of the target time-frequency image corresponding to the scales σ _i+1 and σ _i-1 and the second-order standard Gaussian function (not shown in FIG. 3-a ). When determining the feature points in the image 302, such as determining whether the pixel point 302 is a feature point, it is necessary to determine whether the H (x, y, σ) at the pixel point 302 is its adjacent 26 pixel points (in the image 301 At most 8 pixels adjacent to the pixel 302 , at most 18 pixels adjacent to the pixel 302 in the image adjacent to the image 301 ) is the largest. If the value of H(x, y, σ) of the pixel point 302 is the largest, then the pixel point 302 can be determined as a feature point. If the value of the pixel point H(x, y, σ) is not the largest, it can be determined that the pixel point 302 is not a feature point.

Step S1031D, determine the main direction of each feature point.

In this step, in order to improve the robustness of the SURF algorithm to rotation changes, that is, to improve the robustness of the feature points extracted after the target time-frequency image is rotated, the main direction of each feature point can be determined.

For the convenience of understanding, FIG. 3-b is taken as an example for illustration. FIG. 3-b is a schematic diagram of determining the main direction of a feature point provided by an embodiment of the present invention. In the circular area with the feature point as the center, that is, in the area 303, according to the preset rotation angle, such as α, the fan-shaped area 304 is rotated in the area 303, and the horizontal and vertical directions of all the feature points in the fan-shaped area 304 are calculated. The response value of the Haar (Haar) wavelet on , and the fan-shaped direction with the largest Haar wavelet cumulative value in the fan-shaped area 304 is determined as the main direction of the feature point. Wherein, the angle of the fan-shaped area is a preset angle, such as 60°. In the embodiment of the present invention, the method for calculating the response value of the above-mentioned Haar wavelet is not specifically described.

Step S1031E, for each feature point, according to the main direction of the feature point, determine the feature vector corresponding to the feature point as a feature descriptor.

In this step, for each feature point, a rectangular area with a preset size is selected according to the main direction of the feature point, and the rectangular area is divided into multiple sub-areas. Count the response values of the Haar wavelet in the horizontal and vertical directions of the preset number of pixels in each sub-region. According to the sum of the response values of the Haar wavelet in the horizontal direction, the sum of the absolute values of the response values of the Haar wavelet in the horizontal direction, the sum of the response values of the Haar wavelet in the vertical direction, and the sum of the absolute values of the response values of the Haar wavelet in the vertical direction, generate features Vector, as the feature descriptor of the feature point. Wherein, the horizontal direction and the vertical direction are the horizontal direction and the vertical direction relative to the main direction of the feature point.

Feature descriptors for the above feature points It can be expressed as:

Among them, X is the horizontal direction relative to the main direction of the feature point, Y is the vertical direction relative to the main direction of the feature point, d _X is the response value of the Haar wavelet in the horizontal direction, d _Y is the response value of the Haar wavelet in the vertical direction , |·| represents the absolute value.

Taking feature point 1 among the multiple feature points as an example, the corresponding scale of feature point 1 in the above scale space is σ ₁ , and a square area with a side length of 20σ ₁ in the main direction of feature point 1 can be selected as the above preset Set the size of the rectangular area. Divide the square area into 4*4=16 sub-areas, each sub-area has a side length of 20σ ₁ /4=5σ ₁ , then each sub-area can include 5*5=25 pixels, that is, the above-mentioned predetermined Let the quantity be 25. For the 25 pixels included in each sub-region, calculate the response value of each pixel in the horizontal direction and vertical direction of the main direction of feature point 1 to determine the Haar wavelet of the pixel point, and then generate the feature description of feature point 1 symbol.

The feature points in the target time-frequency image are extracted by using the SURF algorithm. In addition, the scale-invariant feature transform (Scale-invariant feature transform, SIFT) algorithm can also be used for fast and rotated brief introduction (oriented FAST and rotated Brief, ORB) algorithm and other preset algorithms, as well as the preset network model of the convolutional neural network, to extract the feature points in the target time-frequency image. In the embodiment of the present invention, the algorithm for extracting the above feature points is not specifically limited.

In an optional embodiment, for the above step S1032, the K-means algorithm is used to perform clustering processing on multiple feature points to obtain K clusters, which may specifically include the following steps.

Step S1032A, selecting K feature points from multiple feature points as target feature points.

In this step, according to certain rules, such as random selection or equidistant selection, K feature points may be selected from the extracted feature points as target feature points.

In the embodiment of the present invention, each target feature point corresponds to a feature class, that is, K target feature points belong to K classes.

Step S1032B, for each other feature point, determine the distance between the other feature point and each target centroid, and add the other feature point to the class where the closest target feature point is located.

In this step, for each other feature point except the above-mentioned target feature point, the distance between the other feature point and each of the K target feature points is calculated. Add the other feature points to the class of the target feature point closest to the other feature points.

In the embodiment of the present invention, the distance between the above-mentioned other feature points and the target feature point represents the similarity between the other feature points and the features corresponding to the target feature point. The smaller the distance, the greater the similarity; The larger, the smaller the similarity. In addition, the above-mentioned distance may be represented by Euclidean distance, Minkowski distance, Manhattan distance, etc., and the above-mentioned distance is not specifically limited here.

Step S1032C, according to the clustering result, determine the clustering indexes of the K classes. Among them, the clustering index is better than measuring the clustering effect of K classes.

In an optional embodiment, the following formula is used to calculate the sum of squared errors (Sum of SquaredError, SSE) of the K classes as the clustering index of the K classes. Among them, the clustering index of K classes can be expressed as:

SSE is the clustering index of K classes, that is, the sum of square errors of the above K classes, k is the number of classes, i is the i-th, j is the j-th, A _i is the i-th class, a _j is The jth feature point in A _i , dist(a _j , A _i ) is the distance from the feature point a _j to the target feature point in A _i .

Step S1032D, judging whether the clustering index is smaller than a preset index threshold. If not, execute step S1032E. If yes, execute step S1032F.

In this step, the determined clustering index is compared with a preset index threshold to determine whether the clustering index is smaller than the preset index threshold. If the clustering index is less than the preset index threshold, it is determined that the clustering process on the multiple feature points is completed. If the clustering index is not less than the preset index threshold, it is determined that the clustering process on the multiple feature points is completed.

Step S1032E, for each of the K classes, redetermine the target feature points of the class, and return to step S1032B.

In this step, when the above-mentioned clustering index is not less than the preset index threshold, that is, when the clustering index is greater than or equal to the preset index threshold, the electronic device can re-determine the class of each of the K classes. target feature point, and return to perform the above steps for each other feature point, determine the distance between the other feature point and each target centroid, and add the other feature point to the class of the nearest target feature point .

In step S1032F, the clustering process is ended, and the target feature data of the target time-frequency image is obtained.

In this step, when the above-mentioned clustering index is less than the preset index threshold, the electronic device can determine that the clustering process has been completed, end the clustering process, and determine that the clustered K classes are the target feature data of the target time-frequency image .

The K-means algorithm is used when clustering the extracted multiple feature points. In addition, other clustering algorithms can also be used, such as the Density-Based SpatialClustering of Applications with Noise (DBSCAN) algorithm based on the density of the noise application space clustering (Density-Based SpatialClustering of Applications with Noise, DBSCAN) algorithm, etc. The clustering algorithm is not specifically limited.

By clustering the feature points in the extracted target time-frequency image, the target feature data of the target time-frequency image is obtained, so that the classification of the target feature data is clear, accurate, and more representative, and the accuracy of the target feature data is improved. , thereby improving the accuracy of the signal identification model in identifying the category of the signal to be identified, and improving the efficiency of signal identification.

To sum up, through the method provided by the embodiment of the present invention, according to the difference in the feature data in the time-frequency images corresponding to different signals, the trained signal recognition model is used to determine the type of the signal, so that one device can be used to identify different different types of signals, it is no longer necessary to separately equip corresponding modules for each type of signal, which reduces the cost of equipment and improves the efficiency of signal recognition.

In an optional embodiment, for the above-mentioned signal recognition model, the embodiment of the present invention further provides a method for training the signal recognition model. Specifically, as shown in FIG. 4 , FIG. 4 is a schematic flowchart of a method for training a signal recognition model provided by an embodiment of the present invention. The method includes the following steps.

Step S401, obtaining a preset training set.

In this step, the above-mentioned preset training set is obtained, that is, the sample feature data of the sample time-frequency images corresponding to the multiple sample signals in the preset frequency band, and the sample type of each sample signal.

Specifically, a plurality of sample signals and a sample type of each sample signal are acquired within a preset frequency band. The sample time-frequency images corresponding to the acquired multiple sample signals are determined, and sample characteristic data in the sample time-frequency images are extracted.

When the above-mentioned multiple sample signals and the sample type of each sample signal are acquired, the sample signal of each sample type may be respectively sent in a preset frequency band, and the sample signal is received. For example, the sample signal is the above-mentioned Bluetooth signal, Wi-Fi signal and ZigBee signal, and the Bluetooth signal, Wi-Fi signal and ZigBee signal can be sent and received sequentially in the preset frequency band, and the received signal is used as the sample signal, and each The sample type of a sample signal. In addition, for the determination of the sample time-frequency image of the sample signal and the method for extracting the sample feature data, reference may be made to the above-mentioned method for determining the target time-frequency image and extracting the target feature data, which will not be described in detail here.

The above preset training set may include sample time-frequency images of different categories of sample signals. In addition, the preset training set may also include multiple sample time-frequency images of the same type of sample signal. Take Figure 5 as an example for illustration. Fig. 5-a, Fig. 5-b and Fig. 5-c are all sample time-frequency images corresponding to the sample signal provided by the embodiment of the present invention being a Bluetooth signal. Although Figure 5-a, Figure 5-b, and Figure 5-c are all sample time-frequency images corresponding to Bluetooth signals, there are obvious differences among the three. Compared with Figure 5-a and Figure 5-b, Bluetooth The distribution position of the amplitude of the frequency component of the signal in the sample time-frequency image is obviously different. Comparing Figure 5-a and Figure 5-b with Figure 5-c, the frequency resolution of the sample time-frequency image in Figure 5-c is the same as Time resolution The frequency resolution of the sample time-frequency images corresponding to Figure 5-a and Figure 5-b is significantly different from the time resolution.

In the embodiment of the present invention, there is no specific limitation on the data in the above-mentioned preset training set.

In step S402, a plurality of sample feature data are respectively input into a preset machine learning model to determine the type of each sample signal.

In this step, a plurality of sample feature data in the above-mentioned preset training set are respectively input into a preset machine learning model to obtain the corresponding type of each sample signal.

In the embodiment of the present invention, the above machine learning model may be a neural network model or a classifier, such as a convolutional neural network and a support vector machine (Support Vector Machine, SVM). Here, the above machine learning model is not specifically limited.

Step S403, calculating a loss value according to the determined type of each sample signal and the sample type of each sample signal.

In this step, the loss value is calculated according to the type of each sample signal output by the above machine learning model and the sample type of each sample signal in the preset training set.

In one embodiment, for each sample signal, the output type of the machine learning model corresponding to the sample signal may be compared with the sample type of the sample signal to determine whether the output type of the machine learning model is correct. The error rate of the type recognition of the sample signal by the statistical machine learning model.

Other numerical values are also possible with respect to the above-mentioned loss value. For example, the loss value may be the reciprocal of the correct rate of the machine learning model identifying the type of the sample signal. For another example, the loss value can also be expressed as the above SSE. If the category output by the machine learning model is the same as the sample category of the sample signal, it is recorded as 0; if the category output by the machine learning model is different from the sample category of the sample signal, record is 1. The SSE corresponding to multiple sample signals is used as the loss value. In the embodiment of the present invention, the above loss value is not specifically limited.

Step S404, judging whether the loss value is smaller than a preset threshold. If not, execute step S405. If yes, execute step S406.

In this step, according to the above loss value, it is determined whether the machine learning model is converged. Specifically, the loss value calculated above is compared with a preset threshold to determine whether the loss value is smaller than the preset threshold, and then determine whether the machine learning model converges. When the loss value is less than the preset threshold, the electronic device can determine that the machine learning model is converged. When the loss value is not less than the preset threshold, that is, greater than or equal to the preset threshold, the electronic device may determine that the machine learning model has not converged.

Taking the above error rate as an example for illustration, if the preset threshold of the error rate is 2%. During a certain training, if the calculated loss value is 2.1%, if 2.1%>2%, it can be determined that the machine learning model has not converged. During another training, the calculated loss value is 1.9%, and if 1.9%<2%, it can be determined that the machine learning model is converged.

In this embodiment of the present invention, the foregoing preset threshold is not specifically limited.

Step S405, adjust the parameters of the preset machine learning model, and return to execute the above step S402.

In this step, when the above-mentioned loss value is not less than the preset threshold, that is, when the machine learning model has not converged, the parameters of the above-mentioned machine learning model can be adjusted, and return to the above-mentioned step S402, that is, return to the above-mentioned processing of multiple sample features The data are respectively input into a preset machine learning model to determine the type of each sample signal.

The methods for adjusting the parameters of the machine learning model include but are not limited to the gradient descent method, the reverse adjustment method, etc. In the embodiment of the present invention, the method for adjusting the parameters of the machine learning model is not specifically limited.

Step S406, determining a preset machine learning model as a signal recognition model.

In this step, when the aforementioned loss value is less than a preset threshold, that is, when the machine learning model converges, the preset machine learning model may be determined as the aforementioned signal recognition model. At this time, the machine learning model is the pre-trained signal recognition model in the above step S104.

Through the training of the machine learning model, a trained signal recognition model can be obtained. Using the trained signal recognition model, the type of the signal to be recognized can be accurately recognized, and different types of signals can be recognized by using one device. Each type of signal is individually equipped with a corresponding module, which reduces the cost of the equipment and improves the efficiency of signal identification.

In addition, in the existing identification of radio frequency signals, in addition to using different communication modules to identify radio frequency signals according to different communication protocols, electronic equipment can also perform energy detection on different types of radio frequency signals. Through the set energy threshold, Identify each RF signal. However, because it is difficult to determine the decision threshold, that is, the energy threshold is difficult to determine, the signal-to-noise ratio has a greater impact on the detection results of energy detection, and different types of radio frequency signals may overlap in the same frequency band, so energy detection is used. The identification effect of the method for identifying radio frequency signals is poor. Compared with the existing energy detection method for identifying radio frequency signals, the embodiment of the present application obtains a signal identification model by training a machine learning model, and there is no It is difficult to determine the decision threshold, and the signal-to-noise ratio and signal overlap have a great influence on the signal recognition results, which makes the signal recognition effect relatively good.

Based on the same inventive concept, according to the signal identification method provided by the above-mentioned embodiment of the present invention, the embodiment of the present invention also provides a signal identification device. Specifically, as shown in FIG. 6 , FIG. 6 is a schematic structural diagram of a signal identification device provided by an embodiment of the present invention. The device includes the following modules.

The first acquiring module 601 is configured to acquire signals to be identified within a preset frequency band.

The first determining module 602 is configured to determine a target time-frequency image of the signal to be identified.

An extraction module 603, configured to extract target feature data of the target time-frequency image.

The second determining module 604 is configured to input target feature data into a pre-trained signal recognition model to determine the type of the signal to be recognized, wherein the signal recognition model is a model obtained through training a preset training set, and the preset training set includes a preset The sample feature data of the sample time-frequency images corresponding to the multiple sample signals in the frequency band, and the sample type of each sample signal are set.

Optionally, the above-mentioned first determination module 602 can specifically be used to obtain the target time-frequency image of the signal to be identified by using the following short-time Fourier transform formula:

G _f (w,u)=∫f(t)g(tu)e ^-jwt dt

Among them, w is the angular frequency of the signal to be identified, f is the frequency, t is the time t, u is the preset time window length u, the value of the function G _f (w, u) is the amplitude of the frequency component, ∫ dt is For the integral operation of t, the function f(t) is the signal to be identified, the function g(tu) is the preset window function, the function e ^-jwt is a complex variable function, e is a natural constant, and j is an imaginary unit.

Optionally, the above extraction module 603 may include:

The extraction sub-module is used to perform feature extraction on the target time-frequency image to obtain multiple feature points of the target time-frequency image.

The clustering sub-module is used to use the K-means clustering algorithm to cluster multiple feature points to obtain K clusters.

The first determining submodule is used to determine the number of feature points included in each of the K classes.

The second determination sub-module is configured to determine the target feature data in the target time-frequency image according to the number of feature points included in each class.

Optionally, the above extraction sub-module may specifically be used to extract multiple feature descriptors in the target time-frequency image by using the SURF algorithm to obtain multiple feature points of the target time-frequency image.

Optionally, the above-mentioned second determining submodule can be specifically used to construct the first feature point distribution map of the target time-frequency image according to the number of feature points included in each class; when determining the first feature point distribution map as the target The target feature data in the frequency image; or according to the number of feature points included in each class, the probability of the feature points included in each class appearing in the target time-frequency image is counted, and according to the corresponding probability of each class, the target time The second feature point distribution map of the frequency image; the second feature point distribution map is determined as the target feature data in the target time-frequency image.

Optionally, the above-mentioned signal identification device may also include:

The second obtaining module is used to obtain a preset training set.

The third determination module is used to respectively input a plurality of sample feature data into a preset machine learning model to determine the type of each sample signal.

The calculation module is used to calculate the loss value according to the determined type of each sample signal and the sample type of each sample signal.

A judging module, configured to judge whether the loss value is smaller than a preset threshold.

The adjustment module is used to adjust the parameters of the preset machine learning model when the judgment result of the judging module is No, and return to perform the process of inputting a plurality of sample feature data into the preset machine learning model respectively to determine the type of each sample signal step.

The fourth determination module is configured to determine the preset machine learning model as the signal recognition model when the judgment result of the judging module is yes.

Through the device provided by the embodiment of the present invention, according to the difference in the feature data in the time-frequency images corresponding to different signals, the trained signal recognition model is used to determine the type of the signal, so that different types of signals can be recognized by one device, without Furthermore, it is necessary to separately equip corresponding modules for each type of signal, which reduces the cost of equipment and improves the efficiency of signal recognition.

Based on the same inventive concept, according to the signal identification method provided by the above-mentioned embodiments of the present invention, the embodiments of the present invention also provide an electronic device, as shown in FIG. 7 , which is a kind of electronic device provided by the embodiments of the present invention. Schematic. The electronic device includes a processor 701, a communication interface 702, a memory 703, and a communication bus 704, wherein the processor 701, the communication interface 702, and the memory 703 complete mutual communication through the communication bus 704;

Memory 703, used to store computer programs;

When the processor 701 is used to execute the program stored on the memory 703, the following steps are implemented:

Obtain the signal to be identified within the preset frequency band;

Determine the target time-frequency image of the signal to be identified;

Extracting target feature data of the target time-frequency image;

Input the target feature data into the pre-trained signal recognition model to determine the type of the signal to be recognized, wherein the signal recognition model is a model obtained through the training of the preset training set, and the preset training set includes multiple sample signals corresponding to the preset frequency band The sample feature data of the sample time-frequency image of , and the sample type of each sample signal.

The electronic device provided by the embodiment of the present invention uses the trained signal recognition model to determine the type of the signal according to the difference in the characteristic data in the time-frequency image corresponding to different signals, so that different types of signals can be recognized by one device. It is no longer necessary to separately equip a corresponding communication module for each type of signal, which reduces the cost of equipment and improves the efficiency of signal identification.

The communication bus mentioned in the above electronic device may be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus or the like. The communication bus can be divided into an address bus, a data bus, a control bus, and the like. For ease of representation, only one thick line is used in the figure, but it does not mean that there is only one bus or one type of bus.

The communication interface is used for communication between the electronic device and other devices.

The memory may include a random access memory (Random Access Memory, RAM), and may also include a non-volatile memory (Non-Volatile Memory, NVM), such as at least one disk memory. Optionally, the memory may also be at least one storage device located far away from the aforementioned processor.

Above-mentioned processor can be general-purpose processor, comprises central processing unit (Central Processing Unit, CPU), network processor (Network Processor, NP) etc.; Can also be Digital Signal Processor (Digital Signal Processing, DSP), ASIC (Application Specific Integrated Circuit, ASIC), Field-Programmable Gate Array (Field-Programmable Gate Array, FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

Based on the same inventive concept, according to the signal identification method provided by the above-mentioned embodiments of the present invention, the embodiments of the present invention also provide a computer-readable storage medium, in which a computer program is stored in the computer-readable storage medium, and the computer program is The processor implements the steps of any one of the above-mentioned signal identification methods when executed.

Based on the same inventive concept, according to the signal recognition method provided by the above-mentioned embodiments of the present invention, the embodiments of the present invention also provide a computer program product containing instructions, which, when run on a computer, cause the computer to execute any of the above-mentioned embodiments. A signal identification method.

In the above embodiments, all or part of them may be implemented by software, hardware, firmware or any combination thereof. When implemented using software, it may be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on the computer, the processes or functions according to the embodiments of the present invention will be generated in whole or in part. The computer can be a general purpose computer, a special purpose computer, a computer network, or other programmable devices. The computer instructions may be stored in or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions may be transmitted from a website, computer, server or data center Transmission to another website site, computer, server, or data center by wired (eg, coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (eg, infrared, wireless, microwave, etc.). The computer-readable storage medium may be any available medium that can be accessed by a computer, or a data storage device such as a server or a data center integrated with one or more available media. The available medium may be a magnetic medium (for example, a floppy disk, a hard disk, or a magnetic tape), an optical medium (for example, DVD), or a semiconductor medium (for example, a Solid State Disk (SSD)).

It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply that there is a relationship between these entities or operations. There is no such actual relationship or order between them. Furthermore, the term "comprises", "comprises" or any other variation thereof is intended to cover a non-exclusive inclusion such that a process, method, article, or apparatus comprising a set of elements includes not only those elements, but also includes elements not expressly listed. other elements of or also include elements inherent in such a process, method, article, or device. Without further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article or apparatus comprising said element.

Each embodiment in this specification is described in a related manner, the same and similar parts of each embodiment can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for embodiments such as devices, electronic equipment, computer-readable storage media, and computer program products, since they are basically similar to the method embodiments, the description is relatively simple. For relevant information, please refer to the part of the description of the method embodiments. .

The above descriptions are only preferred embodiments of the present invention, and are not intended to limit the protection scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present invention are included in the protection scope of the present invention.

Claims (10)

1. A signal identification method, characterized in that, comprising: Obtain the signal to be identified within the preset frequency band; determining a target time-frequency image of the signal to be identified; Extracting target feature data of the target time-frequency image; Inputting the target feature data into a pre-trained signal recognition model to determine the type of the signal to be recognized, wherein the signal recognition model is a model obtained by training a preset training set, and the preset training set includes the The sample characteristic data of the sample time-frequency images corresponding to the multiple sample signals in the preset frequency band, and the sample type of each sample signal. 2. The method according to claim 1, wherein the step of determining the target time-frequency image of the signal to be identified comprises: Using the following short-time Fourier transform formula, the target time-frequency image of the signal to be identified is obtained: G _f (w,u)=∫f(t)g(tu)e ^-jwt dt Wherein, w is the angular frequency of the signal to be identified, f is the frequency, t is the time t, u is the preset time window length u, the value of the function G _f (w, u) is the amplitude of the frequency component, ∫. dt is an integral operation on t, the function f(t) is the signal to be identified, the function g(tu) is a preset window function, the function e ^-jwt is a complex variable function, e is a natural constant, and j is an imaginary number unit. 3. The method according to claim 1, wherein the step of extracting the target feature data in the target time-frequency image comprises: performing feature extraction on the target time-frequency image to obtain a plurality of feature points of the target time-frequency image; Using the K-means clustering algorithm, the plurality of feature points are clustered to obtain K classes; Determine the number of feature points included in each of the K classes; According to the number of feature points included in each class, target feature data in the target time-frequency image is determined. 4. The method according to claim 3, wherein the step of performing feature extraction on the target time-frequency image to obtain a plurality of feature points of the target time-frequency image comprises: The accelerated robust feature SURF algorithm is used to extract multiple feature descriptors in the target time-frequency image to obtain multiple feature points of the target time-frequency image. 5. method according to claim 3, is characterized in that, described according to the quantity of the characteristic point that each class comprises, the step of determining the target feature data in the target time-frequency image comprises: Constructing a first feature point distribution map of the target time-frequency image according to the number of feature points included in each class; determining the first feature point distribution map as target feature data in the target time-frequency image; or , According to the number of feature points included in each class, the probability of the feature points included in each class appearing in the target time-frequency image is counted, and according to the probability corresponding to each class, the second part of the target time-frequency image is constructed. A feature point distribution map: determining the second feature point distribution map as target feature data in the target time-frequency image. 6. The method according to claim 1, wherein the signal recognition model is trained through the following steps, comprising: Obtain the preset training set; Input multiple sample feature data into the preset machine learning model to determine the type of each sample signal; calculating a loss value according to the determined type of each sample signal and the sample type of each sample signal; Determine whether the loss value is less than a preset threshold; If not, then adjust the parameters of the preset machine learning model, return to perform the step of inputting a plurality of sample characteristic data into the preset machine learning model respectively, and determine the type of each sample signal; If yes, the preset machine learning model is determined as the signal recognition model. 7. A signal identification device, characterized in that it comprises: The first acquisition module is used to acquire the signal to be identified in the preset frequency band; A first determination module, configured to determine the target time-frequency image of the signal to be identified; An extraction module, configured to extract target feature data of the target time-frequency image; The second determination module is configured to input the target feature data into a pre-trained signal recognition model to determine the type of the signal to be recognized, wherein the signal recognition model is a model obtained by training a preset training set, so The preset training set includes sample feature data of sample time-frequency images corresponding to a plurality of sample signals in the preset frequency band, and a sample type of each sample signal. 8. The device according to claim 7, wherein the first determination module is specifically configured to use the following short-time Fourier transform formula to obtain the target time-frequency image of the signal to be identified: G _f (w,u)=∫f(t)g(tu)e ^-jwt dt Wherein, w is the angular frequency of the signal to be identified, f is the frequency, t is the time t, u is the preset time window length u, the value of the function G _f (w, u) is the amplitude of the frequency component, ∫. dt is an integral operation on t, the function f(t) is the signal to be identified, the function g(tu) is a preset window function, the function e ^-jwt is a complex variable function, e is a natural constant, and j is an imaginary number unit. 9. The device according to claim 7, wherein the extraction module comprises: The extraction submodule is used to perform feature extraction on the target time-frequency image to obtain a plurality of feature points of the target time-frequency image; The clustering submodule is used to adopt the K-means clustering algorithm to perform clustering processing on the plurality of feature points to obtain K classes; The first determination submodule is used to determine the quantity of feature points included in each class in the K classes; The second determination submodule is configured to determine the target feature data in the target time-frequency image according to the number of feature points included in each class. 10. The device according to claim 9, wherein the extraction sub-module is specifically configured to use the accelerated robust feature SURF algorithm to extract multiple feature descriptors in the target time-frequency image to obtain the target Multiple feature points of a time-frequency image.