Signal identification method and device, electronic equipment and storage medium

Abstract

The invention provides a signal identification method and device, electronic equipment and a storage medium, and the method comprises the steps: obtaining and inputting an initial signal into a first deep learning model for filtering processing, and obtaining target time-frequency feature information and a target to-be-identified signal corresponding to the target time-frequency feature information; inputting the target to-be-recognized signal into a signal type recognition model for type recognition to obtain a target signal type; determining a target modulation mode identification model in the second deep learning model based on the target signal type; inputting the target to-be-identified signal into a target modulation mode identification model for modulation mode identification to obtain a target modulation mode; and inputting the target modulation mode and the target time-frequency characteristic information into a third deep learning model for signal identification to obtain a signal identification result, the interference of background noise or other radio equipment signals can be reduced, the dependence on the comprehensiveness of signal feature information during signal identification is reduced, the generalization ability is improved, and the identification cost of signal identification is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of signal recognition, and in particular, to a signal recognition method, apparatus, electronic device, and storage medium. Background Art

[0002] Currently, the methods for identifying drone signals mainly include the identification method based on spectral features, modulation and protocol identification methods, signal recognition methods based on machine learning, and signal recognition methods for array signal processing. However, the above methods have the following disadvantages:

[0003] 1. The identification method based on spectral features is easily interfered by the background noise of drone signals or the signals of other radio devices, resulting in a low signal recognition accuracy; 2. The modulation and protocol identification methods rely strongly on the comprehensiveness of the characteristic information of drone signals. For drone signals with poor comprehensiveness of characteristic information, the recognition accuracy is low and the universality is poor; 3. The signal recognition method based on machine learning has limited generalization ability for drone signals with poor data quality (for example, drone signals obtained in complex scenarios such as multipath effects). Therefore, the signal recognition efficiency and accuracy are low; 4. The signal recognition method for array signal processing requires complex antenna arrays and signal processing equipment to improve robustness and accuracy, that is, it has high requirements for hardware and algorithm implementation, and the resulting signal recognition cost is high. Therefore, how to reduce the interference of background noise or the signals of other radio devices, reduce the dependence on the comprehensiveness of signal characteristic information during signal recognition, improve the generalization ability, and reduce the recognition cost of signal recognition are the technical problems that need to be solved urgently at present. Summary of the Invention

[0004] The present invention provides a signal recognition method, apparatus, electronic device, and storage medium, which can reduce the interference of background noise or the signals of other radio devices, reduce the dependence on the comprehensiveness of signal characteristic information during signal recognition, improve the generalization ability, and reduce the recognition cost of signal recognition.

[0005] According to a first aspect of the present invention, there is provided a signal recognition method, the method comprising:

[0006] Obtain an initial signal, and input the initial signal into a first deep learning model for filtering processing to obtain target time-frequency characteristic information and a target signal to be recognized corresponding to the target time-frequency characteristic information;

[0007] Input the target signal to be recognized into a signal type recognition model for type recognition to obtain a target signal type of the target signal to be recognized;

[0008] Based on the target signal type, determine a target modulation mode recognition model corresponding to the target signal type in a second deep learning model;

[0009] Input the target signal to be recognized into the target modulation mode recognition model corresponding to the target signal type for modulation mode recognition to obtain the target modulation mode;

[0010] Input the target modulation mode and the target time-frequency feature information into a third deep learning model for signal recognition to obtain the signal recognition result of the target signal to be recognized.

[0011] According to a second aspect of the present invention, there is provided a signal recognition device, the device comprising:

[0012] A signal processing module, configured to obtain an initial signal and input the initial signal into a first deep learning model for filtering processing to obtain target time-frequency feature information and a target signal to be recognized corresponding to the target time-frequency feature information;

[0013] A signal type recognition module, configured to input the target signal to be recognized into a signal type recognition model for type recognition to obtain the target signal type of the target signal to be recognized;

[0014] A model selection module, configured to determine, based on the target signal type, a target modulation mode recognition model corresponding to the target signal type in a second deep learning model;

[0015] A modulation mode recognition module, configured to input the target signal to be recognized into the target modulation mode recognition model corresponding to the target signal type for modulation mode recognition to obtain the target modulation mode;

[0016] A signal recognition module, configured to input the target modulation mode and the target time-frequency feature information into a third deep learning model for signal recognition to obtain the signal recognition result of the target signal to be recognized.

[0017] According to a third aspect of the present invention, there is provided an electronic device, comprising a processor and a memory,

[0018] The memory is configured to store codes and related data;

[0019] The processor is configured to execute the codes in the memory to implement the signal recognition method according to any one of the embodiments of the present invention.

[0020] According to a fourth aspect of the present invention, there is provided a storage medium, on which a computer program is stored, and when the program is executed by a processor, the signal recognition method according to any one of the embodiments of the present invention is implemented.

[0021] The signal recognition method provided by the present invention first filters the initial signal through a first deep learning model, reducing the interference of background noise or signals of other radio devices in the signal to be recognized, and improving the signal quality of the signal to be recognized.

[0022] Secondly, the signal recognition method provided by the present invention only needs to perform signal recognition based on the target modulation mode and target time-frequency feature information of the signal to be recognized to obtain the recognition result, without the need for more signal feature information. This reduces the types of feature information required for signal recognition, reduces the dependence on the comprehensiveness of feature information during signal recognition, and improves the universality of the signal recognition method.

[0023] Furthermore, the signal to be recognized, the target time-frequency feature information, and the target modulation mode are all data results obtained based on the training of the deep learning model. That is, the signal recognition method provided by the present invention is implemented based on the deep learning model. The feature extraction ability, complex data processing ability, and generalization ability of the deep learning model are superior to those of the machine learning model. Therefore, for the signal recognition of UAV signals obtained in complex scenarios such as multipath effects, compared with the signal recognition method based on the machine learning model, the generalization ability and signal recognition efficiency of the signal recognition method are improved.

[0024] Finally, the signal recognition method provided by the present invention does not require complex antenna arrays and signal processing devices, and only requires basic hardware devices to achieve the purpose of signal recognition, reducing the recognition cost of signal recognition. Description of the Drawings

[0025] 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 required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0026] Figure 1 It is a schematic flowchart of the signal recognition method provided by an embodiment of the present invention;

[0027] Figure 2 It is another schematic flowchart of the signal recognition method provided by an embodiment of the present invention;

[0028] Figure 3 It is a schematic diagram of the signal recognition method provided by an embodiment of the present invention;

[0029] Figure 4 It is a schematic structural diagram of the signal recognition device provided by an embodiment of the present invention;

[0030] Figure 5It is a schematic structural diagram of an electronic device provided by an embodiment of the present invention. Detailed implementation manners

[0031] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0032] The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present invention and the above accompanying drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged under appropriate circumstances so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0033] Next, the technical solutions of the present invention will be described in detail with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0034] Figure 1 It is a schematic flowchart of a signal recognition method provided by an embodiment of the present invention. This method can be executed by a signal recognition device, and the device can be implemented in software and / or hardware. In a specific embodiment, the device can be integrated in an electronic device, such as a computer, a server, etc. The following embodiments will be described by taking the device integrated in the electronic device as an example. Refer to Figure 1 , and the method can specifically include the following steps:

[0035] Step 101: Obtain an initial signal, and input the initial signal into a first deep learning model for filtering processing to obtain target time-frequency feature information and a target signal to be recognized corresponding to the target time-frequency feature information.

[0036] Among them, the initial signal can be understood as the received unknown signal. The first deep learning model can be understood as a deep learning model for signal processing and filtering of the initial signal. The first deep learning model can include a time-frequency feature map extraction network, a signal detection network, and a filtering network. The time-frequency feature map extraction network can be a Short-Time Fourier Transform (STFT) network. The network structure of the signal detection network can be a network structure based on a Convolutional Neural Network (CNN). The filtering network can be a two-dimensional time-frequency filter. Specifically, a corresponding two-dimensional time-frequency filter can be designed in advance based on the time-frequency feature information of the signal. The time-frequency feature information can be understood as the feature representation information of the signal in the two dimensions of time and frequency; the time-frequency feature information can include, but is not limited to, one or more of the signal bandwidth, time slot information, and frequency hopping pattern information, etc. The time slot information can include the start time, end time, and time slot of the signal, and the time slot information can be determined according to the start time and end time of the signal. The frequency hopping pattern information can include the start frequency, end frequency, and frequency hopping pattern of the signal, and the frequency hopping pattern information can be determined according to the start frequency and end frequency of the signal. The target time-frequency feature information can be understood as the time-frequency feature information of the signal obtained through the filtering process of the first deep learning model. The target signal to be recognized can be understood as the signal corresponding to the target time-frequency feature information. Specifically, in the embodiments of the present invention, the initial signal and the target signal to be recognized can be drone signals. In other embodiments, the initial signal and the target signal to be recognized can be other signals that also have time-frequency feature information, not limited to drone signals only.

[0037] In the embodiments of the present invention, since the network structure of the signal detection network can be a network structure based on CNN, and CNN has strong feature extraction ability, it can extract the time-frequency feature information of the signal more comprehensively, enhancing the information extraction ability of the time-frequency feature information in the signal recognition method and increasing the comprehensiveness of the time-frequency feature information. Moreover, by filtering the time-frequency feature information extracted by the signal detection network through the two-dimensional time-frequency filter, the interference of background noise or signals of other radio devices in the target signal to be recognized is reduced, thus improving the signal quality of the target signal to be recognized.

[0038] In an alternative embodiment, the initial signal can be input into the time-frequency feature map extraction network for time-frequency feature map extraction to obtain the time-frequency feature map of the initial signal; the time-frequency feature map can be input into the signal detection network for time-frequency feature extraction to obtain the time-frequency feature information of the initial signal; and the time-frequency feature information can be input into the filtering network for filtering to obtain the target time-frequency feature information and the target signal to be recognized corresponding to the target time-frequency feature information.

[0039] Specifically, the time-frequency feature map is input into the signal detection network for time-frequency feature extraction to obtain the time-frequency feature information of the initial signal. It can be understood that the time-frequency feature map is input into the signal detection network, and the signal detection network marks the signals that may exist in the time-frequency space and outputs the time-frequency feature information of the marked signals. The signal marking method can be marking with a bounding box or other marking methods, and no specific limitation is made here.

[0040] Filtering the time-frequency feature information through the filtering network may include: screening out the target time-frequency feature information that meets the preset conditions from the time-frequency feature information through the filtering network according to the bandwidth, time slot, and frequency hopping pattern in the time-frequency feature information. Among them, the preset conditions may include: 1. The bandwidth belongs to the preset bandwidth interval; 2. The time slot belongs to the preset time slot interval; 3. The frequency hopping pattern conforms to the preset frequency hopping pattern.

[0041] It should be noted that if the time-frequency feature information includes other information in addition to the bandwidth, time slot, and frequency hopping pattern, conditions corresponding to other information in addition to the bandwidth, time slot, and frequency hopping pattern can be added to the preset conditions to improve the accuracy of signal filtering.

[0042] Exemplarily, the initial signal is r(t), the time-frequency feature map extraction network is the STFT network, and the filtering network is a two-dimensional time-frequency filter. The initial signal r(t) can be obtained, and the initial signal r(t) is input into the STFT for time-frequency feature map extraction to obtain the time-frequency feature map of the initial signal; the time-frequency feature map is input into the signal detection network for time-frequency feature extraction to obtain the time-frequency feature information of the initial signal (start time end time start frequency end frequency time slot, bandwidth, and frequency hopping pattern). Among them, n represents the nth segment of the signal, 1 ≤ n ≤ N, and N is a positive integer. Finally, the time-frequency feature information is input into the filtering network for filtering to obtain the target time-frequency feature information and the target signal to be recognized s n (t).

[0043] Step 102, input the target signal to be recognized into the signal type recognition model for type recognition to obtain the target signal type of the target signal to be recognized.

[0044] Among them, the network structure of the signal type recognition model can be a network structure based on deep learning. The target signal type can be understood as the signal type of the target signal to be recognized. The target signal type may include linear signals and non-linear signals.

[0045] Specifically, since the linear signal is an envelope type and the non-linear signal is a constant envelope signal, the signal type recognition model can determine the target signal type of the target signal to be recognized according to the envelope category of the target signal to be recognized.

[0046] In other embodiments, the signal type recognition model can adopt frequency domain and high-order spectrum methods to determine the target signal type of the target signal to be recognized according to the time-frequency characteristic information of the target signal to be recognized. The specific method for determining the target signal type of the target signal to be recognized by the signal type recognition model is not specifically limited herein.

[0047] Step 103: Based on the target signal type, determine the target modulation mode recognition model corresponding to the target signal type in the second deep learning model.

[0048] Among them, the second deep learning model can be understood as a deep learning model for recognizing the modulation mode of a signal. The second deep learning model can include a first modulation mode recognition model and a second modulation mode recognition model. The first modulation mode recognition model can be understood as a deep learning model for recognizing the modulation mode of a linear signal. The second modulation mode recognition model can be understood as a deep learning model for recognizing the modulation mode of a non-linear signal. The target modulation mode recognition model can be understood as the second deep learning model corresponding to the target signal type.

[0049] In an alternative embodiment, it can be determined whether the target signal type is a linear signal; when the target signal type is a linear signal, the first modulation mode recognition model is determined as the target modulation mode recognition model; when the target signal type is a non-linear signal, the second modulation mode recognition model is determined as the target modulation mode recognition model, so as to input different types of target signals to be recognized into the corresponding modulation mode recognition model for modulation mode recognition, improving the recognition efficiency and accuracy of the target modulation mode.

[0050] Exemplarily, the target signal type of the target signal to be recognized s n (t) is a linear signal. When the target signal type of the target signal to be recognized s n (t) is a linear signal, the first modulation mode recognition model is determined as the target modulation mode recognition model.

[0051] Step 104: Input the target signal to be recognized into the target modulation mode recognition model corresponding to the target signal type for modulation mode recognition to obtain the target modulation mode.

[0052] Among them, the modulation methods of the signal can include linear modulation and non-linear modulation; the modulation methods corresponding to linear modulation can include various linear modulation methods such as Quadrature Phase Shift Keying (QPSK), 4-Quadrature Amplitude Modulation (QAM), 16QAM, etc., and the modulation methods corresponding to non-linear modulation can include various non-linear modulation methods such as Frequency-Shift Keying (FSK) and Gaussian Filtered Minimum Shift Keying (GMSK). In other embodiments, linear modulation can also include other linear modulation methods in addition to the various linear modulation methods mentioned above, and non-linear modulation can also include other non-linear modulation methods in addition to the various non-linear modulation methods mentioned above. Since different modulation methods require different processing operations to achieve accurate identification of the target modulation method of the target signal to be identified, therefore, in an alternative embodiment, when the target signal type is a linear signal, the target signal to be identified is input into the first modulation method identification model for modulation method identification to obtain the target modulation method of the target signal to be identified; when the target signal type is a non-linear signal, the target signal to be identified is input into the second modulation method identification model for modulation method identification to obtain the target modulation method of the target signal to be identified.

[0053] In one embodiment, the first modulation method identification model can include an orthogonal demodulation network, a constellation diagram generation network, and a constellation diagram identification network. The orthogonal demodulation network can be understood as a network for demodulating the target signal to be identified to obtain the demodulated signal corresponding to the target signal to be identified. The constellation diagram can be understood as a distribution diagram of the endpoints of the signal vectors. The constellation diagram can be used to determine the modulation method of the signal. The constellation diagram generation network can be understood as a network for generating a constellation diagram based on the demodulated signal. The constellation diagram identification network can be understood as a network for identifying the modulation method of the signal based on the constellation diagram.

[0054] Specifically, inputting the target signal to be identified into the first modulation method identification model for modulation method identification to obtain the target modulation method of the target signal to be identified can include: inputting the target signal to be identified into the orthogonal demodulation network for signal demodulation to obtain the first demodulated signal; inputting the first demodulated signal into the constellation diagram generation network for image generation to obtain the constellation diagram; inputting the constellation diagram into the constellation diagram identification network for modulation method identification to obtain the target modulation method of the target signal to be identified. Among them, the first demodulated signal can be understood as the signal obtained by inputting the target signal to be identified into the orthogonal demodulation network for signal demodulation.

[0055] Among them, the constellation diagram generation network may include a frequency offset estimation module, a frequency offset compensation module, and a constellation diagram generation module. Inputting the first demodulated signal into the constellation diagram generation network for image generation to obtain a constellation diagram may include: using the frequency offset estimation module to estimate the frequency offset of the first demodulated signal to obtain a first offset; using the frequency offset compensation module to compensate the first demodulated signal according to the first offset to obtain a compensated signal; using the frequency offset estimation module to estimate the frequency offset of the compensated signal to obtain a second offset; when the second offset does not belong to the preset offset interval, performing signal compensation and frequency offset estimation on the compensated signal again until the second offset belongs to the preset offset interval; when the second offset belongs to the preset offset interval, using the constellation diagram generation module to generate a constellation diagram according to the compensated signal. In this way, the frequency offset estimation module and the frequency offset compensation module can perform iterative frequency offset estimation and frequency offset compensation on the first demodulated signal after quadrature demodulation to achieve frequency alignment of the compensated signal, thereby improving the accuracy of the constellation diagram generated according to the frequency-aligned compensated signal, and further improving the accuracy of the target modulation method obtained according to the constellation diagram. Among them, the frequency offset estimation module can be understood as a module for estimating the offset of the frequency of a signal. The frequency offset compensation module can be understood as a module for compensating a signal according to the offset. The constellation diagram generation module can be understood as a module for generating a constellation diagram according to a signal. The first offset can be understood as the offset of the frequency of the first demodulated signal. The second offset can be understood as the offset of the frequency of the compensated signal.

[0056] In a specific embodiment of the present invention, the constellation diagram recognition network may include a first fully connected layer and a first decision maker. The first fully connected layer can be used to identify the modulation method according to the constellation diagram and output all possible first recognition results. The first decision maker can be used to determine the target modulation method according to all possible first recognition results.

[0057] In one of the embodiments, the second modulation method recognition model may include a frequency modulation demodulation network, an eye diagram generation network, and an eye diagram recognition network. The frequency modulation demodulation network can be understood as a network for restoring the original signal of the target signal to be recognized. The eye diagram can be understood as a graph displayed by accumulating a series of digital signals on an oscilloscope. The eye diagram generation network can be understood as a network for generating a corresponding eye diagram according to the demodulated signal. The eye diagram recognition network can be understood as a network for identifying the modulation method of a signal according to the eye diagram.

[0058] Specifically, inputting the target signal to be recognized into the second modulation method recognition model for modulation method recognition to obtain the target modulation method of the target signal to be recognized may include: inputting the target signal to be recognized into a frequency modulation demodulation network for signal demodulation to obtain a second demodulation signal; inputting the second demodulation signal into an eye diagram generation network for eye diagram imaging to obtain an eye diagram; inputting the eye diagram into an eye diagram recognition network for modulation method recognition to obtain the target modulation method of the target signal to be recognized. Among them, the second demodulation signal can be understood as the signal obtained by inputting the target signal to be recognized into the frequency modulation demodulation network for signal demodulation. The signal demodulation method of the frequency modulation demodulation network can be any one of the slope borrowing method, the phase discriminator method, and the phase-locked loop demodulation method, and the signal demodulation method of the frequency modulation demodulation network is not specifically limited herein.

[0059] Among them, the eye diagram recognition network may include a second fully connected layer and a second decision maker. The second fully connected layer can be used to recognize the modulation method according to the eye diagram and output all possible second recognition results. The second decision maker can be used to determine the target modulation method according to all possible second recognition results.

[0060] In the embodiment of the present invention, inputting the target signal to be recognized into the frequency modulation demodulation network for signal demodulation to obtain a second demodulation signal, that is, solving the problem that the frequencies of the second demodulation signals cannot be aligned by means of frequency modulation demodulation.

[0061] Exemplarily, the target signal type of the target signal s n (t) is a linear signal. The preset offset range is [-0.1, 0.1]. When the target signal type of the target signal s n (t) is a linear signal, input the target signal to be recognized into the quadrature demodulation network for signal demodulation to obtain a first demodulation signal Among them, represents the nth segment of the first demodulation signal, s n (t) is the nth segment of the target signal to be recognized, f n is the first offset of the nth segment of the target signal to be recognized, and j represents a preset parameter. Then, use the frequency offset estimation module to estimate the frequency offset of the first demodulation signal to obtain the first offset f n ; use the frequency offset compensation module to compensate the first demodulation signal according to the first offset to obtain a compensation signal; use the frequency offset estimation module to estimate the frequency offset of the compensation signal to obtain a second offset When the second offset does not belong to the preset offset range [-0.1, 0.1], perform signal compensation and frequency offset estimation on the compensation signal again until the second offset It belongs to the preset offset range [-0.1, 0.1], and then the constellation diagram generation module generates a constellation diagram according to the compensation signal. Finally, the constellation diagram is input into the constellation diagram recognition network for modulation mode recognition to obtain the target modulation mode of the target signal to be recognized.

[0062] Step 105, input the target modulation mode and the target time-frequency feature information into the third deep learning model for signal recognition to obtain the signal recognition result of the target signal to be recognized.

[0063] Among them, the third deep learning model can be understood as a deep learning model for signal recognition according to the target modulation mode and the target time-frequency feature information of the target signal to be recognized. The signal recognition result may include the signal type of the target signal to be recognized; in other embodiments, the signal recognition result may include, but is not limited to, one or more pieces of information such as the signal type, signal name, and signal type of the target signal to be recognized. The third deep learning model may include a third fully connected layer and a third decision maker. The third fully connected layer can be used to perform information fusion according to the target modulation mode and the target time-frequency feature information and output all possible third recognition results. The third decision maker can be used to determine the signal recognition result according to all possible third recognition results.

[0064] Exemplarily, the target time-frequency feature information of the target signal to be recognized s n (t) is Information, the target signal type of the target signal to be recognized s n (t) is a linear signal, the target modulation mode of the target signal to be recognized s n (t) is linear modulation, and the target signal to be recognized s n (t) is a 5GNR signal. When the target signal type of the target signal to be recognized s n (t) is a linear signal, the first modulation mode recognition model can be determined as the target modulation mode recognition model. Then the target signal to be recognized s n (t) is input into the quadrature demodulation network for signal demodulation to obtain the first demodulated signal The first demodulated signal is input into the constellation diagram generation network for image generation to obtain a constellation diagram; the constellation diagram is input into the constellation diagram recognition network for modulation mode recognition to obtain the target modulation mode (linear modulation) of the target signal to be recognized s n (t). Finally, the target time-frequency feature information (Information) and the target modulation mode (linear modulation) of the target signal to be recognized s n (t) are input into the third deep learning model for signal recognition to obtain the signal recognition result of the target signal to be recognized s n (t) as a 5GNR signal.

[0065] In the embodiments of the present invention, first, the initial signal is filtered by the first deep learning model, reducing the interference of background noise or signals of other radio devices in the signal to be recognized, and improving the signal quality of the signal to be recognized. Secondly, the signal recognition method provided by the present invention only needs to perform signal recognition based on the target modulation method and target time-frequency feature information of the signal to be recognized to obtain the recognition result, without the need for more signal feature information, reducing the types of information of the feature information required for signal recognition, thus reducing the dependence on the comprehensiveness of the feature information during signal recognition and improving the universality of the signal recognition method. Moreover, the signal to be recognized, the target time-frequency feature information, and the target modulation method are all data results obtained based on the training of the deep learning model, that is, the signal recognition method provided by the present invention is implemented based on the deep learning model, and the feature extraction ability, complex data processing ability, and generalization ability of the deep learning model are superior to those of the machine learning model. Therefore, for the signal recognition of UAV signals obtained in complex scenarios such as multipath effects, compared with the signal recognition method based on the machine learning model, the generalization ability and signal recognition efficiency of the signal recognition method are improved. Finally, the signal recognition method provided by the present invention does not require complex antenna arrays and signal processing devices, and only basic hardware devices are needed to achieve the purpose of signal recognition, reducing the recognition cost of signal recognition.

[0066] The signal recognition method provided by the embodiments of the present invention will be further described below. As Figure 2 shown, Figure 2 is another schematic flowchart of the signal recognition method provided by the embodiments of the present invention, which may specifically include the following steps:

[0067] Step 201, obtain the initial signal.

[0068] Step 202, input the initial signal into the time-frequency feature map extraction network to extract the time-frequency feature map, and obtain the time-frequency feature map of the initial signal.

[0069] Step 203, input the time-frequency feature map into the signal detection network to extract the time-frequency features, and obtain the time-frequency feature information of the initial signal.

[0070] Step 204, input the time-frequency feature information into the filtering network for filtering, and obtain the target time-frequency feature information and the signal to be recognized corresponding to the target time-frequency feature information.

[0071] Exemplarily, as Figure 3As shown, the initial signal r(t) sent by the drone is obtained through a signal acquisition device, and the initial signal r(t) is input into the STFT network for time-frequency feature map extraction to obtain the time-frequency feature map of the initial signal; then the time-frequency feature map is input into the signal detection network to obtain the time-frequency feature information of the initial signal; finally, the time-frequency feature information is input into the time-frequency two-dimensional filter for filtering to obtain the target time-frequency feature information and the target signal to be recognized s n (t).

[0072] Step 205, determine whether the target signal type is a linear signal; if so, execute steps 206 to 209, if not, execute steps 210 to 213.

[0073] Step 206, determine the first modulation mode recognition model as the target modulation mode recognition model.

[0074] Step 207, input the target signal to be recognized into the quadrature demodulation network for signal demodulation to obtain the first demodulated signal.

[0075] Step 208, input the first demodulated signal into the constellation diagram generation network for image generation to obtain the constellation diagram.

[0076] Step 209, input the constellation diagram into the constellation diagram recognition network for modulation mode recognition to obtain the target modulation mode of the target signal to be recognized.

[0077] Exemplarily, as Figure 3 shown, when it is determined that the target signal to be recognized s n (t) is a linear signal, input the target signal to be recognized s n (t) into the first modulation mode recognition model for modulation mode recognition to obtain the target modulation mode of the target signal to be recognized.

[0078] Specifically, as Figure 3 shown, input the target signal to be recognized s n (t) into the first modulation mode recognition model for modulation mode recognition to obtain the target modulation mode of the target signal to be recognized, which may include: input the target signal to be recognized s n (t) into the quadrature demodulation network for signal demodulation to obtain the first demodulated signal; use the frequency offset estimation module to estimate the frequency offset of the first demodulated signal to obtain the first offset f n ; use the frequency offset compensation module to compensate the first demodulated signal according to the first offset to obtain the compensated signal; use the frequency offset estimation module to estimate the frequency offset of the compensated signal to obtain the second offset. When the second offset does not belong to the preset offset range [-0.1, 0.1], perform signal compensation and frequency offset estimation on the compensated signal again until the second offset belongs to the preset offset range [-0.1, 0.1]; at the second offset when it belongs to the preset offset range [-0.1, 0.1], use the constellation diagram generation module to generate a constellation diagram according to the compensation signal; use the constellation diagram recognition network to input the constellation diagram into the first fully connected layer for modulation mode recognition and output all possible first recognition results; then use the constellation diagram recognition network to input all possible first recognition results into the first decision maker, so that the first decision maker determines and outputs the target modulation mode according to all possible first recognition results.

[0079] After performing step 209, perform step 214.

[0080] Step 210, determine the second modulation mode recognition model as the target modulation mode recognition model.

[0081] Step 211, input the target signal to be recognized into the frequency modulation demodulation network for signal demodulation to obtain a second demodulated signal.

[0082] Step 212, input the second demodulated signal into the eye diagram generation network for eye diagram imaging to obtain an eye diagram.

[0083] Step 213, input the eye diagram into the eye diagram recognition network for modulation mode recognition to obtain the target modulation mode of the target signal to be recognized.

[0084] Exemplarily, as Figure 3 shown, when the target signal type is a non-linear signal, input the target signal to be recognized s n (t) into the second modulation mode recognition model for modulation mode recognition to obtain the target modulation mode of the target signal to be recognized.

[0085] Specifically, as Figure 3 shown, input the target signal to be recognized s n (t) into the second modulation mode recognition model for modulation mode recognition to obtain the target modulation mode of the target signal to be recognized, which may include: inputting the target signal to be recognized s n (t) into the frequency modulation demodulation network for signal demodulation to obtain a second demodulated signal; inputting the second demodulated signal into the eye diagram generation network for eye diagram imaging to obtain an eye diagram; using the eye diagram recognition network to input the eye diagram into the second fully connected layer for modulation mode recognition and output all possible second recognition results; then using the eye diagram recognition network to input all possible second recognition results into the second decision maker, so that the second decision maker determines and outputs the target modulation mode according to all possible second recognition results.

[0086] Step 214, input the target modulation mode and the target time-frequency feature information into the third deep learning model for signal recognition to obtain the signal recognition result of the target signal to be recognized.

[0087] Exemplarily, as Figure 3 shown, input the target modulation method and the target time-frequency characteristic information into the third deep learning model for signal recognition, and the target signal to be recognized s n (t) can be obtained, and the signal recognition result (5GNR signal).

[0088] Specifically, as Figure 3 shown, use the third deep learning model to input the target modulation method and the target time-frequency characteristic information into the third fully connected layer for information fusion and output all possible third recognition results; then use the third deep learning model to input all possible third recognition results into the third decision maker to determine and output the target signal to be recognized s n (t) signal recognition result (5GNR signal).

[0089] In the embodiment of the present invention, first, the initial signal is filtered by the first deep learning model, reducing the interference of background noise or other radio device signals in the target signal to be recognized, and improving the signal quality of the target signal to be recognized. Secondly, only based on the target modulation method and the target time-frequency characteristic information of the target signal to be recognized for signal recognition to obtain the recognition result, without other more signal characteristic information, reducing the types of information of the characteristic information required for signal recognition, then reducing the dependence on the comprehensiveness of the characteristic information during signal recognition, and improving the universality of the signal recognition method. Furthermore, the target signal to be recognized, the target time-frequency characteristic information, and the target modulation method are all data results obtained based on the training of the deep learning model, and the feature extraction ability, complex data processing ability, and generalization ability of the deep learning model are superior to those of the machine learning model. Therefore, for the signal recognition of UAV signals obtained in complex scenarios such as multipath effects, compared with the signal recognition method based on the machine learning model, the generalization ability and signal recognition efficiency of the signal recognition method are improved. Finally, without complex antenna arrays and signal processing devices, only basic hardware devices are required to achieve the purpose of signal recognition, reducing the recognition cost of signal recognition.

[0090] Figure 4 is a schematic structural diagram of a signal recognition device provided by an embodiment of the present invention, and this device is applicable to execute the signal recognition method provided by the embodiment of the present invention. As Figure 4 shown, this device may specifically include:

[0091] A signal processing module 401, configured to obtain an initial signal, and input the initial signal into the first deep learning model for filtering processing to obtain target time-frequency characteristic information and a target signal to be recognized corresponding to the target time-frequency characteristic information;

[0092] A signal type recognition module 402, configured to input the target signal to be recognized into a signal type recognition model for type recognition, so as to obtain a target signal type of the target signal to be recognized;

[0093] A model selection module 403, configured to determine a target modulation mode recognition model corresponding to the target signal type from a second deep learning model based on the target signal type;

[0094] A modulation mode recognition module 404, configured to input the target signal to be recognized into the target modulation mode recognition model corresponding to the target signal type for modulation mode recognition, so as to obtain a target modulation mode;

[0095] A signal recognition module 405, configured to input the target modulation mode and the target time-frequency feature information into a third deep learning model for signal recognition, so as to obtain a signal recognition result of the target signal to be recognized.

[0096] Optionally, the second deep learning model includes a first modulation mode recognition model and a second modulation mode recognition model, the target signal type includes a linear signal and a non-linear signal, and the model selection module 403 is specifically configured to:

[0097] Determine whether the target signal type is a linear signal;

[0098] When the target signal type is the linear signal, determine the first modulation mode recognition model as the target modulation mode recognition model;

[0099] When the target signal type is the non-linear signal, determine the second modulation mode recognition model as the target modulation mode recognition model.

[0100] Optionally, the modulation mode recognition module 404 is specifically configured to:

[0101] When the target signal type is the linear signal, input the target signal to be recognized into the first modulation mode recognition model for modulation mode recognition, so as to obtain a target modulation mode of the target signal to be recognized;

[0102] When the target signal type is the non-linear signal, input the target signal to be recognized into the second modulation mode recognition model for modulation mode recognition, so as to obtain a target modulation mode of the target signal to be recognized.

[0103] Optionally, the first modulation mode recognition model includes an orthogonal demodulation network, a constellation diagram generation network, and a constellation diagram recognition network. The modulation mode recognition module 404 inputs the target signal to be recognized into the first modulation mode recognition model for modulation mode recognition, and obtaining a target modulation mode of the target signal to be recognized includes:

[0104] Input the target signal to be recognized into the quadrature demodulation network for signal demodulation to obtain a first demodulated signal;

[0105] Input the first demodulated signal into the constellation diagram generation network for image generation to obtain a constellation diagram;

[0106] Input the constellation diagram into the constellation diagram recognition network for modulation method recognition to obtain the target modulation method of the target signal to be recognized.

[0107] Optionally, the constellation diagram generation network includes a frequency offset estimation module, a frequency offset compensation module, and a constellation diagram generation module, and a modulation method recognition module 404. Inputting the first demodulated signal into the constellation diagram generation network for image generation to obtain a constellation diagram includes:

[0108] Use the frequency offset estimation module to estimate the frequency offset of the first demodulated signal to obtain a first offset;

[0109] Use the frequency offset compensation module to compensate the first demodulated signal according to the first offset to obtain a compensated signal;

[0110] Use the frequency offset estimation module to estimate the frequency offset of the compensated signal to obtain a second offset;

[0111] When the second offset does not belong to the preset offset interval, perform signal compensation and frequency offset estimation on the compensated signal again until the second offset belongs to the preset offset interval;

[0112] When the second offset belongs to the preset offset interval, use the constellation diagram generation module to generate the constellation diagram according to the compensated signal.

[0113] Optionally, the second modulation method recognition model includes a frequency modulation demodulation network, an eye diagram generation network, and an eye diagram recognition network, and a modulation method recognition module 404. Inputting the target signal to be recognized into the second modulation method recognition model for modulation method recognition to obtain the target modulation method of the target signal to be recognized includes:

[0114] Input the target signal to be recognized into the frequency modulation demodulation network for signal demodulation to obtain a second demodulated signal;

[0115] Input the second demodulated signal into the eye diagram generation network for eye diagram imaging to obtain an eye diagram;

[0116] Input the eye diagram into the eye diagram recognition network for modulation method recognition to obtain the target modulation method of the target signal to be recognized.

[0117] Optionally, the first deep learning model includes a time-frequency feature map extraction network, a signal detection network, and a filtering network. The signal processing module 401 is specifically configured to:

[0118] Input the initial signal into the time-frequency feature map extraction network to extract the time-frequency feature map of the initial signal, obtaining the time-frequency feature map of the initial signal;

[0119] Input the time-frequency feature map into the signal detection network to perform time-frequency feature extraction, obtaining the time-frequency feature information of the initial signal;

[0120] Input the time-frequency feature information into the filtering network for filtering, obtaining the target time-frequency feature information and the target signal to be recognized corresponding to the target time-frequency feature information.

[0121] Those skilled in the art can clearly understand that for the convenience and simplicity of description, only the above division of each functional module is used as an example. In actual applications, the above functions can be allocated to different functional modules according to needs, that is, the internal structure of the device is divided into different functional modules to complete all or part of the functions described above. The specific working process of the above-described functional modules can refer to the corresponding process in the foregoing method embodiments and will not be elaborated here.

[0122] The signal recognition device provided by the embodiments of the present invention, firstly, filters the initial signal through the first deep learning model, reducing the interference of background noise or signals of other radio devices in the target signal to be recognized, and improving the signal quality of the target signal to be recognized. Secondly, only based on the target modulation method and the target time-frequency feature information of the target signal to be recognized for signal recognition to obtain the recognition result, without other more signal feature information, reducing the types of information of the feature information required for signal recognition, thus reducing the dependence on the comprehensiveness of the feature information during signal recognition and improving the universality of the signal recognition method. Furthermore, the target signal to be recognized, the target time-frequency feature information, and the target modulation method are all data results obtained based on the training of the deep learning model, and the feature extraction ability, complex data processing ability, and generalization ability of the deep learning model are superior to those of the machine learning model. Therefore, for the signal recognition of UAV signals obtained in complex scenarios such as multipath effects, compared with the signal recognition method based on the machine learning model, the generalization ability and signal recognition efficiency of the signal recognition method are improved. Finally, without complex antenna arrays and signal processing equipment, only basic hardware equipment is required to achieve the purpose of signal recognition, reducing the recognition cost of signal recognition.

[0123] Figure 5 It is a schematic diagram of the structure of an electronic device provided by an embodiment of the present invention.

[0124] Please refer to Figure 5, an electronic device 50 is provided, including:

[0125] a processor 51; and,

[0126] a memory 52 for storing executable instructions of the processor;

[0127] wherein, the processor 51 is configured to execute the methods involved above by executing the executable instructions.

[0128] The processor 51 can communicate with the memory 52 through a bus 53.

[0129] An embodiment of the present invention also provides a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the methods involved above are implemented.

[0130] Those of ordinary skill in the art can understand that all or part of the steps of implementing the above method embodiments can be completed by hardware related to program instructions. The foregoing program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps including the above method embodiments; and the foregoing storage medium includes: various media such as ROM, RAM, magnetic disk or optical disc that can store program codes.

[0131] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A signal recognition method, characterized in that: The method comprises: Acquire an initial signal, and input the initial signal into a first deep learning model for filtering processing to obtain target time-frequency feature information and a target signal to be identified corresponding to the target time-frequency feature information; Inputting the target signal to be identified into a signal type recognition model for type recognition to obtain a target signal type of the target signal to be identified; Based on the target signal type, determining a target modulation mode recognition model corresponding to the target signal type in a second deep learning model; Inputting the target signal to be identified into the target modulation mode identification model corresponding to the target signal type to perform modulation mode identification to obtain a target modulation mode; The target modulation mode and the target time-frequency feature information are input into a third deep learning model for signal recognition to obtain a signal recognition result of the target signal to be recognized.

2. The method according to claim 1, characterized in that The second deep learning model includes a first modulation mode recognition model and a second modulation mode recognition model, the target signal type includes a linear signal and a nonlinear signal, and based on the target signal type, determining a target modulation mode recognition model corresponding to the target signal type in the second deep learning model includes: Determining whether the target signal type is a linear signal; When the target signal type is the linear signal, determining the first modulation mode recognition model as the target modulation mode recognition model; When the target signal type is the nonlinear signal, the second modulation mode identification model is determined as the target modulation mode identification model.

3. The method according to claim 2, characterized in that The step of inputting the target signal to be identified into the target modulation mode identification model corresponding to the target signal type to perform modulation mode identification to obtain the target modulation mode includes: When the target signal type is the linear signal, inputting the target signal to be identified into the first modulation mode identification model to perform modulation mode identification to obtain a target modulation mode of the target signal to be identified; When the target signal type is the nonlinear signal, the target signal to be identified is input into the second modulation mode identification model to perform modulation mode identification to obtain a target modulation mode of the target signal to be identified.

4. The method according to claim 3, characterized in that The first modulation mode recognition model includes an orthogonal demodulation network, a constellation diagram generation network and a constellation diagram recognition network, and the step of inputting the target signal to be recognized into the first modulation mode recognition model for modulation mode recognition to obtain a target modulation mode of the target signal to be recognized includes: Inputting the target signal to be identified into the orthogonal demodulation network for signal demodulation to obtain a first demodulated signal; Inputting the first demodulated signal into the constellation diagram generation network to generate an image, thereby obtaining a constellation diagram; The constellation diagram is input into the constellation diagram recognition network to perform modulation mode recognition to obtain the target modulation mode of the target signal to be recognized.

5. The method according to claim 4, characterized in that The constellation diagram generation network includes a frequency offset estimation module, a frequency offset compensation module and a constellation diagram generation module. The first demodulated signal is input into the constellation diagram generation network for image generation to obtain a constellation diagram, including: Using the frequency offset estimation module to estimate the frequency offset of the first demodulated signal to obtain a first offset; Using the frequency offset compensation module to compensate the first demodulated signal according to the first offset to obtain a compensated signal; Using the frequency offset estimation module to estimate the frequency offset of the compensation signal to obtain a second offset; When the second offset does not belong to the preset offset interval, performing signal compensation and frequency offset estimation on the compensation signal again until the second offset belongs to the preset offset interval; When the second offset belongs to the preset offset interval, the constellation diagram generation module is used to generate the constellation diagram according to the compensation signal.

6. The method according to claim 3, characterized in that The second modulation mode recognition model includes a frequency modulation demodulation network, an eye diagram generation network and an eye diagram recognition network, and the step of inputting the target signal to be recognized into the second modulation mode recognition model for modulation mode recognition to obtain a target modulation mode of the target signal to be recognized includes: Inputting the target signal to be identified into the FM demodulation network for signal demodulation to obtain a second demodulated signal; Inputting the second demodulated signal into the eye diagram generation network to perform eye diagram imaging to obtain an eye diagram; The eye diagram is input into the eye diagram recognition network to perform modulation mode recognition, so as to obtain the target modulation mode of the target signal to be recognized.

7. The method according to claim 1, characterized in that The first deep learning model includes a time-frequency feature map extraction network, a signal detection network and a filtering network. The initial signal is input into the first deep learning model for filtering to obtain target time-frequency feature information and a target signal to be identified corresponding to the target time-frequency feature information, including: Inputting the initial signal into the time-frequency feature map extraction network to extract the time-frequency feature map, thereby obtaining the time-frequency feature map of the initial signal; Inputting the time-frequency feature graph into the signal detection network to extract time-frequency features, thereby obtaining time-frequency feature information of the initial signal; The time-frequency characteristic information is input into the filtering network for filtering to obtain the target time-frequency characteristic information and the target signal to be identified corresponding to the target time-frequency characteristic information.

8. A signal recognition device, characterized in that: The device comprises: A signal processing module is used to obtain an initial signal, and input the initial signal into a first deep learning model for filtering processing to obtain target time-frequency feature information and a target signal to be identified corresponding to the target time-frequency feature information; A signal type identification module, used for inputting the target signal to be identified into a signal type identification model for type identification to obtain a target signal type of the target signal to be identified; A model selection module, configured to determine, based on the target signal type, a target modulation mode recognition model corresponding to the target signal type in a second deep learning model; A modulation mode identification module, used for inputting the target signal to be identified into the target modulation mode identification model corresponding to the target signal type to perform modulation mode identification to obtain a target modulation mode; The signal recognition module is used to input the target modulation mode and the target time-frequency feature information into a third deep learning model for signal recognition to obtain a signal recognition result of the target signal to be recognized.

9. An electronic device, characterized in that: Including processor and memory, The memory is used to store codes and related data; The processor is used to execute the code in the memory to implement the signal recognition method according to any one of claims 1 to 7.

10. A storage medium having a computer program stored thereon, wherein when the program is executed by a processor, the signal recognition method according to any one of claims 1 to 7 is implemented.

Images