Radar emitter individual identification method based on deep learning

Abstract

The present invention provides a method for identifying individual radar radiation sources based on deep learning, which relates to the field of communication identification technology. The present invention forms a complex signal and a residual signal according to the electromagnetic signal obtained by the radar, inputs the complex signal into a CNN network to train a recognition model and obtains the confidence in the recognition model. The residual signal is processed in the same way as the complex signal to obtain an energy spectrum of the residual signal, and the energy index in the energy spectrum is matched with the confidence in the recognition model. The sub-region in which the energy index has a definite relationship with the confidence and the corresponding influencing factor are obtained, and the confidence in the recognition model is corrected by the influencing factor, so as to improve the judgment ability of the recognition model by utilizing the influence of the residual signal on the confidence.

Description

Technical Field

[0001] The present invention relates to the field of communication identification technology, and in particular to a method for identifying individual radar emitters based on deep learning. Background Art

[0002] Identity recognition between radars is an important link in information-based combat scenarios. The current traditional radar identity recognition scheme has long relied on expert experience for modular judgment. This method is time-consuming, cumbersome and has low robustness. Especially in strong interference environments, identity recognition is extremely difficult. With the development of big data technology, radar emitter identification methods based on deep learning can skip expert experience and realize identity recognition by extracting electromagnetic signal features through self-supervised learning, which greatly improves the efficiency of radar emitter identity recognition. However, the interference problem in the environment still brings interference to radar emitter identity recognition based on deep learning, affecting the recognition success rate.

[0003] In the prior art, the technical document with publication number CN116165611A discloses a method, system, device and terminal for accurate and intelligent identification of radar emitter models. The radar repetition frequency PRI and radar pulse width PW of the received radar emitter intermediate frequency signal are estimated, and the radio frequency RF is estimated by obtaining its frequency domain signal based on the fast Fourier transform to form a feature vector; then the minimum covariance determinant MCD is used to calculate the outlier value to separate the abnormal signal; the PDW word of the signal is used as a feature, and the feature label is used to introduce the long short-term memory network training of the attention mechanism to obtain a trained model; when the number of abnormal signal samples reaches a certain number, the K-means algorithm is used to cluster the unknown abnormal signals, and then the MCD is used to separate the points with larger outliers in the new class, and the newly divided class is incrementally learned based on the original model to identify the unknown signal.

[0004] The technical document with publication number CN111913156A discloses a radar emitter individual identification method based on a deep learning model and feature combination. The method includes: 1) collecting intermediate frequency AD signal data emitted by different radars, intercepting intra-pulse signal data to generate a radar emitter individual identification sample set; 2) normalizing the radar emitter individual identification samples and dividing them into a training sample set, a verification sample set, and a test sample set; 3) constructing a radar emitter individual identification model based on a deep learning model and feature combination; 4) training the radar emitter individual identification model based on a deep learning model and feature combination; 5) using the test sample set to obtain the radar emitter individual identification model results and feature judgment results; 6) using the radar emitter individual identification model results and feature judgment results to jointly calculate the final identification result and calculate the recognition accuracy rate.

[0005] Although the two public technical documents have achieved deep learning-based radar radiation source identification in an interference environment, they did not explore and utilize the relationship between the recognition effect and the presence of interference signals during the recognition model training process. They only focus on reducing the dimension of the interference effect and are unable to enhance the judgment effect of the recognition model for the main components of the interference signal.

[0006] The above information disclosed in this Background section is only for enhancement of understanding of the background of the present disclosure and therefore it may contain information that does not form the prior art that is already known to a person of ordinary skill in the art. Summary of the Invention

[0007] The purpose of the present invention is to provide a method for identifying individual radar emitters based on deep learning to solve the problems raised in the above background technology.

[0008] To achieve the above object, the present invention provides the following technical solutions:

[0009] The radar emitter individual identification method based on deep learning includes the following steps:

[0010] Step 1: The historical electromagnetic signals acquired by the radar and their corresponding source identities are aggregated to form a database. The electromagnetic signals are filtered to obtain complex signals and residual signals. The complex signals and residual signals are then divided into the same frames and sampled at the same points to form the energy spectra of the residual signals and the energy spectra of the complex signals.

[0011] Step 2: Using the identity of the electromagnetic signal corresponding to the energy spectrum of each complex signal as a label, the energy spectrum of the complex signal is input into the CNN network for training, the recognition model is obtained, and the confidence level of the completed identity recognition is obtained based on the output of the fully connected layer;

[0012] Step 3: Divide the energy spectrum of the residual signal into subregions, obtain the energy index of each subregion, summarize the energy indexes of the same subregion of all electromagnetic signals to form an energy index sequence, and form a scatter plot of each subregion based on the confidence level corresponding to each element in each energy index sequence. Based on the relationship between the confidence level and the energy index in the scatter plot, obtain the subregions where the energy index and the confidence level have a definite relationship and their corresponding influencing factors;

[0013] Step 4: Obtain the electromagnetic signal received by this radar and generate a filtered complex signal and a residual signal. Input the filtered complex signal into the recognition model to obtain the identity. When the recognition model cannot obtain the identity, correct the confidence in the recognition model by the influence factor of the residual signal, set the confidence threshold, and identify the identity again based on the relationship between the corrected confidence and the confidence threshold.

[0014] Furthermore, the electromagnetic signals and identities of the radiation sources obtained by the local radar are aggregated into a database. The electromagnetic signals in the database are converted into complex signals through IQ frequency modulation, and then clipped so that each complex signal has the same length. The complex signals are denoised using the wavelet denoising method, and the residual signal is obtained from the filtered complex signals. The formula is as follows:

[0015] ;

[0016] in, is the residual signal, is the unfiltered complex signal, is the filtered complex signal, Retrieve variables for time, , the length of time to be retained for clipping;

[0017] At the same time, sampling points are set at equal time intervals in each complex signal and residual signal, and the sampling points are framed using a Hamming window. The frame shift of the Hamming window has a 50% overlap. Short-time Fourier transform is performed on the complex signal and residual signal of each frame. The sampling point setting and framing of the complex signal and residual signal are exactly the same, based on the following formula:

[0018] ;

[0019] in, Indicates in Frame, The time-frequency components at frequencies, is the frame number index variable, , , is the total number of Hamming windows, is the frequency number index variable, , , is the number of sampling points in the Hamming window, Indicates in The first frame in the Hamming window The complex signal of sampling points, Retrieve variables for sampling points, , , is the number of sampling points in the Hamming window, Indicates the first The weight of each sampling point is determined by the radar manufacturer. is the Fourier basis.

[0020] Furthermore, the energy spectrum of the residual signal and complex signal of each frame in the Hamming window is obtained according to the following formula:

[0021] ;

[0022] in, Indicates in Frame, The energy value at a frequency, Indicates in Frame, The time-frequency components at frequencies, is the frame number index variable, , , is the total number of Hamming windows, is the frequency number index variable, , , is the number of sampling points in the Hamming window;

[0023] The number of rows of each energy spectrum is the number of frames, and the number of columns is the number of sampling points in the Hamming window. The energy spectra of the residual signal and the complex signal are normalized respectively. The energy spectrum of the residual signal corresponds one-to-one to the energy spectrum of the complex signal. The logic of the one-to-one correspondence is: each residual signal is generated by an independent complex signal, and the residual signal uses exactly the same processing method as the complex signal, so the energy spectrum of each residual signal is generated by and only by the energy spectrum of one corresponding complex signal.

[0024] Furthermore, the energy spectrum of each complex signal is labeled, where the label is the source identity of the electromagnetic signal of the energy spectrum. The energy spectrum of the complex signal is divided into a training set and a validation set in a ratio of 8:2. The training set of the energy spectrum of the complex signal is input into the CNN network for training to obtain a recognition model.

[0025] Obtain the output of the fully connected layer in the recognition model and identify the energy spectrum of the complex signal to complete the confidence level of identity recognition. The formula is as follows:

[0026] ;

[0027] in, represents the confidence level of the identity recognition completed by the energy spectrum of the complex signal, The energy spectrum of the complex signal output by the fully connected layer in the recognition model is The confidence level of an identity, Retrieve variables for identities, , , is the total number of identities, Select a function for Maximum.

[0028] Furthermore, the energy spectrum of the residual signal is divided, and the division logic is as follows:

[0029] The energy spectrum is divided into 4x4 sub-regions according to the number of rows and columns, that is, 16 sub-regions of equal size. The number of rows of the energy spectrum is the same as the number of Hamming windows, and the number of columns is the same as the number of sampling points in the Hamming window. The energy index of all sub-regions is obtained respectively according to the following formula:

[0030] ;

[0031] in, For the The energy index of each sub-region, Retrieve variable for subregion number, , , is the total number of Hamming windows, is the number of sampling points in the Hamming window, Indicates the The sum of the normalized energy values ​​of the regions.

[0032] Furthermore, the energy spectrum of the residual signal generated by each electromagnetic signal is obtained and summarized to obtain the energy index sequence of each sub-region respectively, according to the following formula:

[0033] ;

[0034] in, Indicates the The energy index sequence of the sub-regions, Retrieve variable for subregion number, , , Indicates the The electromagnetic signal The energy index of each sub-region, Retrieve variable for electromagnetic signal number, , , is the total number of electromagnetic signals.

[0035] Furthermore, the energy index sequence of each sub-region is plotted into a scatter plot. The scatter plot is a two-dimensional plot based on a plane rectangular coordinate system. The horizontal axis is the energy index in the energy index sequence and increases to the right, and the vertical axis is the confidence level corresponding to the energy index and increases upward. The confidence level point corresponding to each energy index is marked in the scatter plot. The scatter plot is fitted by linear regression to obtain the influencing factor of the energy index leading to a decrease in confidence. The logic is as follows:

[0036] Fit the 16 scatter plots respectively to obtain the corresponding formulas of confidence and energy index. The formulas are as follows:

[0037] ;

[0038] in, is the confidence level, is the energy indicator, is the corresponding formula of confidence and energy index;

[0039] Obtain the Pearson relationship coefficient of each fitting curve separately and set the coefficient threshold. When the Pearson relationship coefficient of the fitting curve is less than or equal to the coefficient threshold, it means that there is a definite relationship between the energy index and the confidence level of the sub-region. The corresponding formula at this time is the influence factor of the sub-region. When the Pearson relationship coefficients of all fitting curves are greater than the coefficient threshold, it is determined that only the energy index and the confidence level of the sub-region with the smallest Pearson relationship coefficient have a definite relationship. The corresponding formula at this time is the influence factor of the sub-region.

[0040] Summarize the sub-regions where there is a definite relationship between energy index and confidence and their corresponding impact factors.

[0041] Furthermore, the electromagnetic signal actually received by the radar station is processed to obtain a complex signal, and the complex signal is filtered using the same filtering method as used in the recognition model training to obtain a filtered complex signal and a residual signal. The filtered complex signal is input into the recognition model to obtain the identity of the electromagnetic signal. When the recognition model cannot identify the identity of the electromagnetic signal, the energy spectrum of the residual signal and the energy index of each sub-region are obtained through the residual signal according to step 3. The modified confidence is obtained through the influence factor of the sub-region where there is a definite relationship between the energy index and the confidence. The logic is as follows:

[0042] The maximum confidence is obtained through the fully connected layer of the recognition model. The maximum confidence is the maximum value of all confidences output by the fully connected layer for different identity matches. The corrected confidence is obtained according to the following formula:

[0043] ;

[0044] in, To correct the confidence level, is the maximum confidence level, There is a definite relationship between energy index and confidence level. The energy index of each sub-region, There is a definite relationship between energy index and confidence level. The impact factor of each sub-region, Retrieve variables for sub-regions where there is a definite relationship between energy index and confidence level, , , is the total number of sub-regions for which there is a definite relationship between energy index and confidence.

[0045] Furthermore, a confidence threshold is set. When the corrected confidence is greater than or equal to the confidence threshold, the identity of the electromagnetic signal with the maximum confidence is determined to be matched. When the corrected confidence is less than the confidence threshold, the identity of the electromagnetic signal is determined to be unrecognizable and an alarm is issued.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] The present invention forms a complex signal and a residual signal based on the electromagnetic signal obtained by the radar, trains the recognition model with the complex signal, obtains the energy spectrum of the residual signal by subjecting the residual signal to the same processing as the complex signal, corresponds the energy index in the energy spectrum with the confidence in the recognition model, obtains the sub-region where the energy index and the confidence have a definite relationship and the corresponding influencing factor, and corrects the confidence in the recognition model through the influencing factor, so as to improve the judgment ability of the recognition model by utilizing the influence of the residual signal on the confidence. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 Schematic diagram of the overall method of the present invention. DETAILED DESCRIPTION

[0049] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to specific embodiments.

[0050] It should be noted that, unless otherwise defined, the technical or scientific terms used in the present invention should have the usual meanings understood by people with ordinary skills in the field to which the present invention belongs. The "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative position relationships. When the absolute position of the object being described changes, the relative position relationship may also change accordingly.

[0051] Example:

[0052] See also Figure 1 , the present invention provides a technical solution:

[0053] The radar emitter individual identification method based on deep learning includes the following steps:

[0054] Step 1: The historical electromagnetic signals acquired by the radar and their corresponding source identities are aggregated to form a database. The electromagnetic signals are filtered to obtain complex signals and residual signals. The complex signals and residual signals are then divided into the same frames and sampled at the same points to form the energy spectra of the residual signals and the energy spectra of the complex signals.

[0055] The step 1 includes the following:

[0056] Step 101: The electromagnetic signals and identities of the radiation sources acquired by the local radar are aggregated into a database. The electromagnetic signals in the database are converted into complex signals through IQ frequency modulation, and are clipped so that each complex signal has the same length. The complex signals are de-noised using the wavelet denoising method, and the residual signal is obtained from the filtered complex signals. The formula is as follows:

[0057] ;

[0058] in, is the residual signal, is the unfiltered complex signal, is the filtered complex signal, Retrieve variables for time, , the length of time to be retained for clipping;

[0059] At the same time, sampling points are set at equal time intervals in each complex signal and residual signal, and the sampling points are framed using a Hamming window. The frame shift of the Hamming window has a 50% overlap. Short-time Fourier transform is performed on the complex signal and residual signal of each frame. The sampling point setting and framing of the complex signal and residual signal are exactly the same, based on the following formula:

[0060] ;

[0061] in, Indicates in Frame, The time-frequency components at frequencies, is the frame number index variable, , , is the total number of Hamming windows, is the frequency number index variable, , , is the number of sampling points in the Hamming window, Indicates in The first frame in the Hamming window The complex signal of sampling points, Retrieve variables for sampling points, , , is the number of sampling points in the Hamming window, Indicates the first The weight of each sampling point is determined by the radar manufacturer. is the Fourier basis.

[0062] It reflects the energy distribution and phase information of the signal in the frequency domain within the frame, that is, the characteristics of the radar signal in the corresponding time period and frequency band in the real environment. Reflects the instantaneous amplitude and phase of the signal in the time domain; Used to smooth the window edge, reduce sidelobe leakage, and ensure more accurate spectrum estimation; Reflects the effect of each sampling point on the The contribution of each frequency component.

[0063] In fact, it is the superposition effect of the weighted complex signals of all sampling points at a specific frequency. In other words, The amplitude and phase of the By emphasizing or de-emphasizing certain sampling points, the distribution of the overall contribution is changed, and the exponential function determines how these contributions are mapped to specific frequencies in the frequency domain.

[0064] When the amplitude of some sampling points increases, the The contribution of When the corresponding exponential function gets a strong resonance, The magnitude of will increase; similarly, The design makes the contribution of the middle sampling points greater than that of the edge parts, ensuring that the main signal components are prominently reflected; the exponential term superimposes the sampling points according to the frequency domain basis by rotating the phase. If the sampling points are coherent at the same frequency, their contributions will be constructively superimposed to form a local energy peak.

[0065] By aggregating electromagnetic signals acquired by local radars and their source identities into a database and preprocessing the signals using methods such as IQ frequency modulation and wavelet denoising, the original signal is converted into a complex signal and the residual signal is simultaneously extracted, ensuring that both signals can be processed using exactly the same framing and sampling settings. Synchronous processing helps ensure data consistency and comparability during subsequent energy spectrum extraction. Filtering and residual calculations effectively reduce noise interference and improve the fidelity of signal features, laying the foundation for providing high-quality input features for deep learning models. It also enables seamless integration between upstream and downstream processing steps, creating a stable and reliable signal preprocessing pipeline.

[0066] Step 102: Obtain the energy spectrum of the residual signal and the complex signal of each frame in the Hamming window according to the following formula:

[0067] ;

[0068] in, Indicates in Frame, The energy value at a frequency, Indicates in Frame, The time-frequency components at frequencies, is the frame number index variable, , , is the total number of Hamming windows, is the frequency number index variable, , , is the number of sampling points in the Hamming window;

[0069] It reflects the energy of each component of the actual signal in the time-frequency plane, that is, in radar signal processing, it describes the signal at the first Frame specific frequency The power or energy on. It is the complex time-frequency component obtained by short-time Fourier transform, which contains amplitude and phase information, and the modulus square The phase information is discarded and only the energy information of the signal is retained, thus more intuitively reflecting the strength characteristics of the signal at each time-frequency point.

[0070] When the amplitude of is large, its square is also large, and vice versa. This demonstrates a non-negative property that enhances energy differences, allowing the energy spectrum to more clearly distinguish between strong and weak components in the signal, thereby helping to achieve more accurate and robust feature extraction in actual radar signal target recognition.

[0071] With the input signal Amplitude and window function at the sampling point The weighted effect on the signal, if at a certain frequency The contribution of The amplitude of will increase, resulting in The square relationship shows a more significant improvement; this change relationship enables the energy spectrum to more intuitively reflect the power distribution of the signal in practical applications, which is beneficial to subsequent signal feature analysis and input processing of deep learning models.

[0072] The number of rows of each energy spectrum is the number of frames, and the number of columns is the number of sampling points in the Hamming window. The energy spectra of the residual signal and the complex signal are normalized respectively. The energy spectrum of the residual signal corresponds one-to-one to the energy spectrum of the complex signal. The logic of the one-to-one correspondence is: each residual signal is generated by an independent complex signal, and the residual signal uses exactly the same processing method as the complex signal, so the energy spectrum of each residual signal is generated by and only by the energy spectrum of one corresponding complex signal.

[0073] By performing a short-time Fourier transform on the complex signal and residual signal after each frame is framed by a Hamming window and calculating the energy spectrum of each time-frequency component, a two-dimensional energy spectrum is generated, with the number of rows corresponding to the number of frames and the number of columns corresponding to the number of sampling points. This two-dimensional energy spectrum not only fully captures the local characteristics of the signal from a time-frequency perspective, but also ensures comparability between different signals through normalization, providing stable and discriminative feature input for subsequent CNN network training. It also ensures consistency and scientificity in data transmission between upstream and downstream steps, making the feature extraction process of the entire recognition process more accurate and reliable.

[0074] Step 2: Using the identity of the electromagnetic signal corresponding to the energy spectrum of each complex signal as a label, the energy spectrum of the complex signal is input into the CNN network for training, the recognition model is obtained, and the confidence level of the completed identity recognition is obtained based on the output of the fully connected layer;

[0075] The step 2 includes the following:

[0076] Label the energy spectrum of each complex signal, where the label indicates the source identity of the electromagnetic signal of the energy spectrum. Divide the energy spectrum of the complex signal into a training set and a validation set in a ratio of 8:2. Input the training set of the energy spectrum of the complex signal into a CNN network for training to obtain a recognition model.

[0077] The input layer of the recognition model is the energy spectrum of the complex signal. The energy spectrum is a two-dimensional matrix. The number of vector columns is the number of frames, and the vector dimension is the number of frequencies. The formula is as follows:

[0078] ;

[0079] in, is the two-dimensional matrix of energy spectrum, Retrieve variables for frame numbers, , , is the total number of Hamming windows, is the frequency number index variable, , , is the number of sampling points in the Hamming window, Represents the operation of adding a channel;

[0080] Assume that the first convolutional layer has a single-channel input, and the calculation formula is:

[0081] ;

[0082] in, is the output, size is , is the two-dimensional matrix of energy spectrum, is the convolution kernel, size is 3x3, is the bias term, is a 32-dimensional vector, is a two-dimensional convolution operation, , Select the function for the maximum value, Represents the unknown;

[0083] The number of filters in the first convolutional layer is 32;

[0084] Set the pooling kernel size of the first pooling layer to 2x2, the stride to 2, and the output size to ;

[0085] In the second convolutional layer, the pooled output is convolved and extracted according to the following formula:

[0086] ;

[0087] in, is the output, size is , is the output of the first pooling layer, is the convolution kernel, size is 3x3, is the bias term, is a 64-dimensional vector, is a two-dimensional convolution operation, , Select the function for the maximum value, Represents the unknown;

[0088] The second pooling layer has the same size as the first pooling layer, and the output size is ;

[0089] The output of the second pooling layer is globally averaged and each channel outputs a mean value. The formula is as follows:

[0090] ;

[0091] in, Indicates the The channel mean, Represents the output of the second pooling layer Channel No. Vector The characteristics of the dimension, Retrieve variables for the number of vectors, , , is the total number of Hamming windows ， Retrieve variables for vector dimensions, , , is the number of sampling points in the Hamming window, Retrieve variables for a channel, , ;

[0092] The output of the global average pooling is input to the fully connected classification layer to obtain the original score of each component in each label. The output of the fully connected classification layer is then processed by Softmax to obtain the final prediction probability of each category. The label corresponding to the maximum prediction probability is the recognition result.

[0093] Obtain the output of the fully connected layer in the recognition model and identify the energy spectrum of the complex signal to complete the confidence level of identity recognition. The formula is as follows:

[0094] ;

[0095] in, represents the confidence level of the identity recognition completed by the energy spectrum of the complex signal, The energy spectrum of the complex signal output by the fully connected layer in the recognition model is The confidence level of an identity, Retrieve variables for identities, , , is the total number of identities, Select a function for Maximum.

[0096] represents the confidence level of the complex signal energy spectrum in completing the identity recognition, and It represents the predicted probability of each identity after the fully connected layer of the recognition model is processed by Softmax. In fact, it reflects the model's confidence in the input signal belonging to different identities. They are all related to the feature extraction of the input signal and the weight distribution of the parameters within the model, reflecting the degree of match between the features captured in the signal energy spectrum and the specific identity template; therefore, As the calculation result of max(Pg), the highest value among all predictions is taken, which intuitively reflects the highest confidence of the model in making identity judgments. This relationship shows the probability corresponding to all identities in the independent variable. For dependent variables The influence of: Due to the The operation emphasizes the nature of "winner takes all", that is, in multi-category judgment, the final judgment result depends on the item with the highest probability, and the change in its value is reflected in the direct comparison of the output of each category, ensuring that the final recognition decision is based on the most powerful evidence, thereby providing the entire identity recognition system with intuitive and explainable output indicators.

[0097] Using the energy spectrum of each complex signal and its corresponding electromagnetic signal identity as labels, the system completes training by dividing the data into training and validation sets and inputting them into the designed CNN network. This model constructs an end-to-end recognition model. A modular structure consisting of convolutional layers, pooling layers, global average pooling, and fully connected layers effectively extracts local and global features from the energy spectrum and provides probabilistic identity predictions through Softmax output. This not only greatly improves the accuracy of automatic feature extraction and classification, but also enables the confidence level of the model output to intuitively reflect the degree of signal matching, providing a solid foundation for subsequent confidence correction using residual signals. This ensures that the overall recognition process achieves a good balance between accuracy and robustness, and can be closely integrated with the previous energy spectrum extraction step.

[0098] Step 3: Divide the energy spectrum of the residual signal into subregions, obtain the energy index of each subregion, summarize the energy indexes of the same subregion of all electromagnetic signals to form an energy index sequence, and form a scatter plot of each subregion based on the confidence level corresponding to each element in each energy index sequence. Based on the relationship between the confidence level and the energy index in the scatter plot, obtain the subregions where the energy index and the confidence level have a definite relationship and their corresponding influencing factors;

[0099] Said step 3 comprises the following contents;

[0100] Step 301: Divide the energy spectrum of the residual signal. The division logic is as follows:

[0101] The energy spectrum is divided into 4x4 sub-regions according to the number of rows and columns, that is, 16 sub-regions of equal size. The number of rows of the energy spectrum is the same as the number of Hamming windows, and the number of columns is the same as the number of sampling points in the Hamming window. The energy index of all sub-regions is obtained respectively according to the following formula:

[0102] ;

[0103] in, For the The energy index of each sub-region, Retrieve variable for subregion number, , , is the total number of Hamming windows, is the number of sampling points in the Hamming window, Indicates the The sum of the normalized energy values ​​of the regions.

[0104] Reflects the average level of signal energy in the sub-area; represents the number of rows of the energy spectrum (i.e. the total number of all Hamming window frames), represents the number of columns of the energy spectrum (that is, the number of sampling points in the Hamming window), and and Respectively represent the number of frames and sampling points contained in each sub-region after the energy spectrum is evenly divided into 4 parts; It is the The result of summing all normalized energy values ​​in a sub-region.

[0105] In a practical sense, It reflects the average intensity of the signal energy in each local area after division and reveals the energy distribution characteristics of the signal in the local time-frequency area, which is of great significance for the analysis of the energy concentration of radar signals in the subdivided area. It means that after data preprocessing in a real environment, the influence of amplitude scale differences between different signals is eliminated, so that the local energy features of different signals can be compared on the same scale, which helps to extract more objective local features.

[0106] The energy value of each sampling point in the sub-region is normalized and accumulated, and then divided by the average value of the total number of sampling points in the sub-region. The change directly affects If the normalized energy value in a sub-region is generally high, then increases, and vice versa. This shows that The signal is obtained by short-time Fourier transform After forming the energy spectrum , then normalized, the energy data of the sampling points within each subregion are scaled uniformly. The average amplifies the local energy distribution, thus accurately reflecting the energy distribution within the subregion. This formula combines the instantaneous energy changes of a large number of sampling points within a local subregion into a stable numerical indicator. This mean value can be used to describe the stability of the local energy distribution when comparing different radar signals and performing subsequent confidence correction analysis, effectively connecting with the confidence assessment in subsequent steps.

[0107] By equally dividing the energy spectrum of the residual signal into 16 sub-regions by row and column, and calculating the energy index of each sub-region separately, a detailed characterization of the local energy distribution of the signal is achieved. On the one hand, the differences between the complex signal and the residual signal in the local region can be discovered, thereby supplementing the information that may be missed in the main feature extraction; on the other hand, by processing each residual signal with its corresponding complex signal on a one-to-one basis, the consistency and refinement of data processing are guaranteed, making the subsequent statistical analysis of the relationship between the energy index and confidence level of each sub-region more reliable, laying the foundation for further correcting the recognition confidence level through local influencing factors, and forming a continuous and effective connection with the signal characteristic analysis link in the entire recognition process.

[0108] Step 302: Obtain the energy spectrum of the residual signal generated by each electromagnetic signal and summarize it to obtain the energy index sequence of each sub-region respectively, based on the following formula:

[0109] ;

[0110] in, Indicates the The energy index sequence of the sub-regions, Retrieve variable for subregion number, , , Indicates the The electromagnetic signal The energy index of each sub-region, Retrieve variable for electromagnetic signal number, , , is the total number of electromagnetic signals.

[0111] By summarizing the energy indices of each subregion within the residual signal generated by each electromagnetic signal, a sequence of energy indices for each subregion is generated. This not only constructs a comprehensive energy index database but also provides sufficient data support for subsequent analysis of the correlation between the energy of each subregion and the recognition confidence. Statistical analysis of the energy index sequence more intuitively reveals the intrinsic connection between local energy characteristics and the output of the recognition model, providing the necessary basis for subsequently deriving influencing factors through scatter plot fitting. This ensures data continuity and rigorous analytical logic between the previous and next steps, further improving the accuracy and interpretability of the overall recognition method.

[0112] Step 303: Plot the energy index sequence of each sub-region into a scatter plot. The scatter plot is a two-dimensional plot based on a rectangular coordinate system. The horizontal axis represents the energy index in the energy index sequence and increases to the right, while the vertical axis represents the confidence level corresponding to the energy index and increases upward. The confidence level point corresponding to each energy index is marked in the scatter plot. The scatter plot is fitted by linear regression to obtain the influencing factor of the energy index causing the confidence level to decrease. The logic is as follows:

[0113] Fit the 16 scatter plots respectively to obtain the corresponding formulas of confidence and energy index. The formulas are as follows:

[0114] ;

[0115] in, is the confidence level, is the energy indicator, is the corresponding formula of confidence and energy index;

[0116] It means "non-confidence" or can be understood as the same as the recognition confidence The complementary value of , that is, when the confidence level is low, The larger the value of , the more uncertainty is reflected in the identification; It represents the local energy index, that is, the normalized energy value in a sub-region. The energy index is described How to influence with certain mathematical relationships , which explains how the degree of reduction in recognition confidence changes with changes in local energy indicators in real environments. This formula establishes a quantitative connection, mathematically expressing the relationship between local energy characteristics and the confidence output by the recognition model. This can help us understand and predict the changing trends in recognition model confidence under different energy conditions, and also provides a theoretical basis for subsequent confidence correction of recognition results based on residual signal energy indicators. Through linear fitting or other regression methods, a specific expression for the relationship between energy and confidence in each subregion can be obtained, thereby achieving precise correction for abnormal or low-confidence situations.

[0117] Obtain the Pearson relationship coefficient of each fitting curve separately and set the coefficient threshold. When the Pearson relationship coefficient of the fitting curve is less than or equal to the coefficient threshold, it means that there is a definite relationship between the energy index and the confidence level of the sub-region. The corresponding formula at this time is the influence factor of the sub-region. When the Pearson relationship coefficients of all fitting curves are greater than the coefficient threshold, it is determined that only the energy index and the confidence level of the sub-region with the smallest Pearson relationship coefficient have a definite relationship. The corresponding formula at this time is the influence factor of the sub-region.

[0118] Summarize the sub-regions where there is a definite relationship between energy index and confidence and their corresponding impact factors.

[0119] By plotting the energy index sequence for each subregion as a scatter plot and using linear regression to fit the relationship between confidence and energy index, the Pearson correlation coefficient for each subregion is calculated. This allows the identification of subregions with a deterministic influence on confidence and their corresponding influencing factors. This not only provides a new quantitative analysis method for residual signals but also effectively supplements and corrects the confidence output of the CNN recognition model, improving the robustness and accuracy of the overall recognition algorithm. Furthermore, the identification of influencing factors provides theoretical and data support for the subsequent correction of recognition results based on the actual received electromagnetic signals, forming a closed-loop correction mechanism with tightly linked upstream and downstream steps, ensuring the rationality and effectiveness of information transmission and feedback between each link.

[0120] Step 4: Obtain the electromagnetic signal received by this radar and generate a filtered complex signal and a residual signal. Input the filtered complex signal into the recognition model to obtain the identity. When the recognition model cannot obtain the identity, correct the confidence in the recognition model by the influence factor of the residual signal, set the confidence threshold, and identify the identity again based on the relationship between the corrected confidence and the confidence threshold.

[0121] The step 4 includes the following:

[0122] Step 401: Process the electromagnetic signal actually received by the radar station to obtain a complex signal, and filter the complex signal using the same filtering method as used for recognition model training to obtain the filtered complex signal and residual signal. The filtered complex signal is input into the recognition model to obtain the identity of the electromagnetic signal. If the recognition model cannot identify the identity of the electromagnetic signal, the residual signal's energy spectrum and energy index of each sub-region are obtained from the residual signal according to step 3. The modified confidence is obtained using the influence factor of the sub-region where the energy index has a definite relationship with the confidence. The logic is as follows:

[0123] The maximum confidence is obtained through the fully connected layer of the recognition model. The maximum confidence is the maximum value of all confidences output by the fully connected layer for different identity matches. The corrected confidence is obtained according to the following formula:

[0124] ;

[0125] in, To correct the confidence level, is the maximum confidence level, There is a definite relationship between energy index and confidence level. The energy index of each sub-region, There is a definite relationship between energy index and confidence level. The impact factor of each sub-region, Retrieve variables for sub-regions where there is a definite relationship between energy index and confidence level, , , is the total number of sub-regions for which there is a definite relationship between energy index and confidence.

[0126] This formula aims to convert the traditional recognition result confidence It is fused with the correction information obtained based on the local energy characteristics of the residual signal in order to compensate for the inaccurate recognition problem caused by noise or local anomalies when relying solely on the model output. In the formula, It reflects the recognition confidence of the original signal in the convolutional neural network, and the , then the influence of each sub-region's local energy index on the confidence level is integrated, where each influence factor function This paper quantitatively describes how local energy anomalies in real detection environments reduce the recognition confidence of the model. The value of increases accordingly, thereby increasing the correction value, and the adjusted correction confidence will be higher than the original confidence Higher, reflecting the compensation for detection uncertainty or abnormal noise; if The output is smaller, and This formula establishes a connection between global features and local energy information in the input signal. By uniformly correcting the energy information in each subregion, it enhances the accuracy and robustness of the overall identification process and effectively connects signal preprocessing, deep learning judgment, and subsequent confidence correction.

[0127] For electromagnetic signals received by actual radar stations, the complex signal and residual signal are extracted using the same filtering and preprocessing methods as used during training. A CNN recognition model is then used to obtain a preliminary identity determination. In the event of insufficient recognition, corrections are introduced to the residual signal's energy spectrum and its local influencing factors. This not only ensures consistency in signal processing in practical applications and comparability between new data and model training data, but also utilizes the local energy information of the residual signal as an auxiliary judgment, improving the reliability and robustness of the recognition results. This step serves as a bridge between practical applications and offline models, enabling the method to cope with complex interference and noise environments in real-world scenarios. It also provides sufficient pre-conditioning and assurance for setting confidence thresholds and alarm mechanisms in the next step.

[0128] Step 402: Set a confidence threshold. When the corrected confidence is greater than or equal to the confidence threshold, the identity of the electromagnetic signal with the maximum confidence is determined to match. When the corrected confidence is less than the confidence threshold, the identity of the electromagnetic signal is determined to be unrecognizable and an alarm is issued.

[0129] By setting a threshold for the corrected confidence level of the recognition model, and then deciding whether to accept the recognition result or trigger an alarm mechanism, a clear quality control link is introduced. This threshold setting not only promptly detects situations where confidence is insufficient or the model is unclear, but also enables rapid response and processing in actual operation, avoiding misjudgments and missed judgments, and improving the overall security and operational stability of the system. At the same time, this mechanism effectively connects the front-end signal preprocessing, feature extraction, CNN recognition, and residual signal feedback correction links, establishing a closed-loop feedback mechanism within the overall solution, ensuring that each sub-step works in coordination to achieve highly reliable individual radar emitter identification.

[0130] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters in the formulas are set by technicians in this field according to actual conditions.

[0131] The above embodiments can be implemented in whole or in part by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed by hardware or software depends on the specific application and design constraints of the technical solution.

[0132] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, and may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment as needed.

[0133] The above is only a specific implementation method of the present application, but the scope of protection of the present application is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed in this application, which should be covered by the scope of protection of the present application.

Claims

1. A radar emitter individual identification method based on deep learning, characterized in that: The specific steps include: Step 1: The historical electromagnetic signals acquired by the radar and their corresponding source identities are aggregated to form a database. The electromagnetic signals are filtered to obtain complex signals and residual signals. The complex signals and residual signals are then divided into the same frames and sampled at the same points to form the energy spectra of the residual signals and the energy spectra of the complex signals. Step 2: Using the identity of the electromagnetic signal corresponding to the energy spectrum of each complex signal as a label, the energy spectrum of the complex signal is input into the CNN network for training, the recognition model is obtained, and the confidence level of the completed identity recognition is obtained based on the output of the fully connected layer; Step 3: Divide the energy spectrum of the residual signal into subregions, obtain the energy index of each subregion, summarize the energy indexes of the same subregion of all electromagnetic signals to form an energy index sequence, and form a scatter plot of each subregion based on the confidence level corresponding to each element in each energy index sequence. Based on the relationship between the confidence level and the energy index in the scatter plot, obtain the subregions where the energy index and the confidence level have a definite relationship and their corresponding influencing factors; Step 4: Obtain the electromagnetic signal received by this radar and generate a filtered complex signal and a residual signal. Input the filtered complex signal into the recognition model to obtain the identity. When the recognition model cannot obtain the identity, correct the confidence in the recognition model by the influence factor of the residual signal, set the confidence threshold, and identify the identity again based on the relationship between the corrected confidence and the confidence threshold.

2. The method for identifying radar emitters based on deep learning according to claim 1, characterized in that: The electromagnetic signals and identities of the radiation sources obtained by the local radar are aggregated into a database. The electromagnetic signals in the database are converted into complex signals through IQ frequency modulation, and then clipped to make each complex signal of the same length. The complex signals are denoised using the wavelet denoising method, and the residual signal is obtained from the filtered complex signals. The formula is as follows: ； in, is the residual signal, is the unfiltered complex signal, is the filtered complex signal, Retrieve variables for time, , the length of time to be retained for clipping; At the same time, sampling points are set at equal time intervals in each complex signal and residual signal, and the sampling points are framed using a Hamming window. The frame shift of the Hamming window has a 50% overlap. Short-time Fourier transform is performed on the complex signal and residual signal of each frame. The sampling point setting and framing of the complex signal and residual signal are exactly the same, based on the following formula: ； in, Indicates in Frame, The time-frequency components at frequencies, is the frame number index variable, , , is the total number of Hamming windows, is the frequency number index variable, , , is the number of sampling points in the Hamming window, Indicates in The first frame in the Hamming window The complex signal of sampling points, Retrieve variables for sampling points, , , is the number of sampling points in the Hamming window, Indicates the first The weight of the sampling points, is the Fourier basis.

3. The method for identifying radar emitters based on deep learning according to claim 2, characterized in that: Obtain the energy spectrum of the residual signal and complex signal of each frame in the Hamming window according to the following formula: ； in, Indicates in Frame, The energy value at a frequency, Indicates in Frame, The time-frequency components at frequencies, is the frame number index variable, , , is the total number of Hamming windows, is the frequency number index variable, , , is the number of sampling points in the Hamming window; The number of rows of each energy spectrum is the number of frames, and the number of columns is the number of sampling points in the Hamming window. The energy spectra of the residual signal and the complex signal are normalized respectively. The energy spectrum of the residual signal corresponds one-to-one to the energy spectrum of the complex signal. The logic of the one-to-one correspondence is: each residual signal is generated by an independent complex signal, and the residual signal uses exactly the same processing method as the complex signal, so the energy spectrum of each residual signal is generated by and only by the energy spectrum of one corresponding complex signal.

4. The method for identifying radar emitters based on deep learning according to claim 3, characterized in that: Label the energy spectrum of each complex signal, where the label indicates the source identity of the electromagnetic signal of the energy spectrum. Divide the energy spectrum of the complex signal into a training set and a validation set in a ratio of 8:

2. Input the training set of the energy spectrum of the complex signal into a CNN network for training to obtain a recognition model. Obtain the output of the fully connected layer in the recognition model and identify the energy spectrum of the complex signal to complete the confidence level of identity recognition. The formula is as follows: ； in, represents the confidence level of the identity recognition completed by the energy spectrum of the complex signal, The energy spectrum of the complex signal output by the fully connected layer in the recognition model is The confidence level of an identity, Retrieve variables for identities, , , is the total number of identities, Select a function for Maximum.

5. The method for identifying individual radar emitters based on deep learning according to claim 4, characterized in that: The energy spectrum of the residual signal is divided, and the division logic is as follows: The energy spectrum is divided into 4x4 sub-regions according to the number of rows and columns, that is, 16 sub-regions of equal size. The number of rows of the energy spectrum is the same as the number of Hamming windows, and the number of columns is the same as the number of sampling points in the Hamming window. The energy index of all sub-regions is obtained respectively according to the following formula: ； in, For the The energy index of each sub-region, Retrieve variable for subregion number, , , is the total number of Hamming windows, is the number of sampling points in the Hamming window, Indicates the The sum of the normalized energy values ​​of the regions.

6. The method for identifying individual radar emitters based on deep learning according to claim 5, characterized in that: The energy spectrum of the residual signal generated by each electromagnetic signal is obtained and summarized to obtain the energy index sequence of each sub-region respectively. The formula is as follows: ； in, Indicates the The energy index sequence of the sub-regions, Retrieve variable for subregion number, , , Indicates the The electromagnetic signal The energy index of each sub-region, Retrieve variable for electromagnetic signal number, , , is the total number of electromagnetic signals.

7. The method for identifying individual radar emitters based on deep learning according to claim 6, characterized in that: The energy index sequence of each sub-region is plotted as a scatter plot. The scatter plot is a two-dimensional plot based on a rectangular coordinate system. The horizontal axis is the energy index in the energy index sequence and increases to the right, and the vertical axis is the confidence level corresponding to the energy index and increases upward. The confidence level point corresponding to each energy index is marked in the scatter plot. The scatter plot is fitted by linear regression to obtain the influencing factor of the energy index leading to a decrease in confidence. The logic is as follows: Fit the 16 scatter plots respectively to obtain the corresponding formulas of confidence and energy index. The formulas are as follows: ； in, is the confidence level, is the energy indicator, is the corresponding formula of confidence and energy index; Obtain the Pearson relationship coefficient of each fitting curve separately and set the coefficient threshold. When the Pearson relationship coefficient of the fitting curve is less than or equal to the coefficient threshold, it means that there is a definite relationship between the energy index and the confidence level of the sub-region. The corresponding formula at this time is the influence factor of the sub-region. When the Pearson relationship coefficients of all fitting curves are greater than the coefficient threshold, it is determined that only the energy index and the confidence level of the sub-region with the smallest Pearson relationship coefficient have a definite relationship. The corresponding formula at this time is the influence factor of the sub-region. Summarize the sub-regions where there is a definite relationship between energy index and confidence and their corresponding impact factors.

8. The method for identifying individual radar emitters based on deep learning according to claim 7, characterized in that: The electromagnetic signal actually received by the radar station is processed to obtain a complex signal, and the complex signal is filtered using the same filtering method as the recognition model training to obtain the filtered complex signal and the residual signal. The filtered complex signal is input into the recognition model to obtain the identity of the electromagnetic signal. When the recognition model cannot identify the identity of the electromagnetic signal, the energy spectrum of the residual signal and the energy index of each sub-region are obtained through the residual signal according to step 3. The modified confidence is obtained through the influence factor of the sub-region where the energy index has a definite relationship with the confidence. The logic is as follows: The maximum confidence is obtained through the fully connected layer of the recognition model. The maximum confidence is the maximum value of all confidences output by the fully connected layer for different identity matches. The corrected confidence is obtained according to the following formula: ； in, To correct the confidence level, is the maximum confidence level, There is a definite relationship between energy index and confidence level. The energy index of each sub-region, There is a definite relationship between energy index and confidence level. The impact factor of each sub-region, Retrieve variables for sub-regions where there is a definite relationship between energy index and confidence level, , , is the total number of sub-regions for which there is a definite relationship between energy index and confidence.

9. The method for identifying individual radar emitters based on deep learning according to claim 8, characterized in that: Set a confidence threshold. When the corrected confidence is greater than or equal to the confidence threshold, the identity of the electromagnetic signal with the maximum confidence is determined to match the identity. When the corrected confidence is less than the confidence threshold, the identity of the electromagnetic signal is determined to be unrecognizable and an alarm is issued.

Images