Electromagnetic data classification model training method and device, terminal equipment and computer program product

Abstract

The invention is suitable for the technical field of electromagnetic data processing, and provides a training method and device for an electromagnetic data classification model, terminal equipment and a computer program product, and the method comprises the steps: carrying out the collaborative labeling of electromagnetic data in a first sample data set through combining automatic labeling and manual labeling, and obtaining a training sample data set; extracting a target feature set of the training sample data and generating a first feature vector; and training an initial model based on the first feature vector to obtain a target classification model. According to the method provided by the embodiment of the invention, collaborative labeling is performed on the electromagnetic data in the first sample data set by combining automatic labeling and manual labeling to obtain the training sample data set, so that the training sample data with accurate and consistent labeling data can be obtained more quickly; and then extracting a target feature set of the training sample data and generating a first feature vector, and training the initial model based on the first feature vector to obtain a target classification model, thereby improving the training efficiency and classification accuracy of the target classification model.

Description

Technical Field

[0001] The present application belongs to the field of electromagnetic data processing technology, and in particular, relates to a training method, apparatus, terminal equipment and computer program product for an electromagnetic data classification model. Background Art

[0002] As a core technology in the information age, electromagnetic data target recognition is becoming increasingly important. This technology achieves accurate target identification by analyzing the electromagnetic data generated during the operation of electronic equipment. In the modern military field, different weapon systems have unique electromagnetic characteristics. By capturing and analyzing these characteristics, we can not only determine the specific type of electronic equipment used by the opponent, but also further infer their tactical intentions and direction of action. For example, electromagnetic data captured by the reconnaissance system can identify the type of enemy radar and communication system, which is crucial for the acquisition of battlefield intelligence, the construction of the electromagnetic environment, the assessment of electromagnetic threats, and the formulation of command decisions.

[0003] With the development of information technology, the electromagnetic environment on modern battlefields is becoming increasingly complex. On the one hand, the diversification of electronic equipment has led to a sharp increase in the types and quantity of electromagnetic data; on the other hand, the overlapping of operating frequency ranges of radiation sources for different purposes, such as radar and communication equipment, has become increasingly serious, making it increasingly common for multiple signals to appear simultaneously in the same frequency range.

[0004] Therefore, it is necessary to identify an electromagnetic data classification model with higher recognition accuracy to meet the needs of electromagnetic data classification in complex electromagnetic environments. Summary of the invention

[0005] In view of this, embodiments of the present application provide a training method, apparatus, terminal device and computer program product for an electromagnetic data classification model to improve the training efficiency and classification accuracy of the target classification model.

[0006] A first aspect of an embodiment of the present application provides a method for training an electromagnetic data classification model, comprising:

[0007] The electromagnetic data in the first sample data set are collaboratively annotated by combining automatic annotation and manual annotation to obtain a training sample data set; the training sample data set includes training sample data, and the training sample data has annotated data;

[0008] Extracting a target feature set of the training sample data and generating a first feature vector;

[0009] An initial model is trained based on the first feature vector to obtain a target classification model, and the target classification model is used to classify input electromagnetic data.

[0010] In an implementation of the first aspect, the step of collaboratively annotating the electromagnetic data in the first sample data set by combining automatic annotation and manual annotation to obtain a training sample data set includes:

[0011] Annotating the electromagnetic data in the first sample data set by using the target generator and the pre-trained annotation model to obtain electromagnetic data with preliminary annotated data;

[0012] Acquiring electromagnetic data with manually annotated data, wherein the electromagnetic data with manually annotated data is obtained by annotating a portion of the electromagnetic data with preliminary annotated data by an annotator;

[0013] Generating training sample data with final labeled data based on the electromagnetic data with preliminary labeled results and the electromagnetic data with manual labeled results;

[0014] The training sample data with the final labeled data is stored in the training sample data set.

[0015] In an implementation of the first aspect, labeling the electromagnetic data in the first sample data set by using the target generator and the pre-trained labeling model to obtain electromagnetic data with preliminary labeled data includes:

[0016] Annotating the electromagnetic data in the first sample data set by a target generator to obtain electromagnetic data with first annotated data;

[0017] Annotating the electromagnetic data in the first sample data set by using a pre-trained annotation model to obtain electromagnetic data with second annotated data;

[0018] Electromagnetic data with preliminary annotation data are generated according to the confidence level of the first annotation data and the confidence level of the second annotation data.

[0019] In an implementation of the first aspect, before the target generator annotates the electromagnetic data in the first sample data set to obtain the electromagnetic data having the first annotated data, the method further includes:

[0020] Build a generator and introduce an attention mechanism;

[0021] Constructing a discriminator, wherein the discriminator includes a plurality of convolution kernels of different sizes for extracting features of different granularities;

[0022] Iterative adversarial training is performed based on the generator and the discriminator, and a target generator is output after a preset iteration termination condition is met.

[0023] In an implementation of the first aspect, labeling the electromagnetic data in the first sample data set by using a pre-trained labeling model to obtain electromagnetic data with second labeled data includes:

[0024] Inputting electromagnetic data in the first sample data set into a pre-trained annotation model; freezing some layers of the pre-trained annotation model;

[0025] extracting high-level features of the electromagnetic data in the first sample data set based on the partially layer-frozen pre-trained annotation model;

[0026] First annotated data of the electromagnetic data in the first sample data set is generated based on the high-level features to obtain electromagnetic data with second annotated data.

[0027] In an implementation of the first aspect, extracting a target feature set of the training sample data and generating a first feature vector includes:

[0028] Extracting time domain features and frequency domain features of the training sample data;

[0029] Extracting target time-frequency features of the training sample data;

[0030] Extracting statistical features of the time domain features, the frequency domain features, and the target time-frequency features;

[0031] A first feature vector is generated based on the time domain feature, the frequency domain feature, the target feature and the statistical feature.

[0032] In an implementation of the first aspect, extracting target time-frequency features of the training sample data includes:

[0033] Extracting a first time-frequency feature of the training sample data;

[0034] Extracting a second time-frequency feature of the training sample data;

[0035] The first time-frequency feature and the second time-frequency feature are fused to obtain a target time-frequency feature.

[0036] A second aspect of an embodiment of the present application provides a training device for an electromagnetic data classification model, comprising:

[0037] A collaborative annotation module, used to collaboratively annotate the electromagnetic data in the first sample data set by combining automatic annotation and manual annotation to obtain a training sample data set; the training sample data set includes training sample data, and the training sample data has annotated data;

[0038] A feature extraction module, used to extract a target feature set of the training sample data and generate a first feature vector;

[0039] A model training module is used to train an initial model based on the first feature vector to obtain a target classification model, and the target classification model is used to classify the input electromagnetic data.

[0040] A third aspect of an embodiment of the present application provides a terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method described in the first aspect when executing the computer program.

[0041] A fourth aspect of the embodiments of the present application provides a computer program product, including a computer program, wherein when the computer program is executed, the method described in the first aspect is executed.

[0042] The beneficial effect of the first aspect of the embodiment of the present application is: by combining automatic labeling and manual labeling to collaboratively label the electromagnetic data in the first sample data set, a training sample data set is obtained, and training sample data with accurate and consistent labeling data is obtained more quickly, and then a target feature set of the training sample data is extracted and a first feature vector is generated, and an initial model is trained based on the first feature vector to obtain a target classification model, thereby improving the training efficiency and classification accuracy of the target classification model.

[0043] It can be understood that the beneficial effects of the second to fourth aspects mentioned above can be found in the relevant description of the first aspect mentioned above, and will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0045] Figure 1 It is a schematic diagram of the implementation flow of the training method of the electromagnetic data classification model provided in the embodiment of the present application;

[0046] Figure 2 It is a schematic diagram of the implementation process of the collaborative annotation method provided in the embodiment of the present application;

[0047] Figure 3 It is a schematic diagram of the overall process framework provided by the embodiment of the present application;

[0048] Figure 4 is a schematic diagram of a training device for an electromagnetic data classification model provided in an embodiment of the present application;

[0049] Figure 5is a schematic diagram of a terminal device provided in an embodiment of the present application;

[0050] Figure 6 It is a schematic diagram of a computer program product provided in an embodiment of the present application. DETAILED DESCRIPTION

[0051] In the following description, specific details such as specific system structures, technologies, etc. are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present application. However, it should be clear to those skilled in the art that the present application may also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present application.

[0052] It should be understood that when used in the present specification and the appended claims, the term "comprising" indicates the presence of described features, wholes, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or combinations thereof.

[0053] It should also be understood that the term “and / or” used in the specification and appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0054] As used in the specification and appended claims of this application, the term "if" can be interpreted as "when" or "uponce" or "in response to determining" or "in response to detecting", depending on the context. Similarly, the phrase "if it is determined" or "if [described condition or event] is detected" can be interpreted as meaning "uponce it is determined" or "in response to determining" or "uponce [described condition or event] is detected" or "in response to detecting [described condition or event]", depending on the context.

[0055] In addition, in the description of the present application specification and the appended claims, the terms "first", "second", "third", etc. are only used to distinguish the descriptions and cannot be understood as indicating or implying relative importance.

[0056] References to "one embodiment" or "some embodiments" etc. described in the specification of this application mean that one or more embodiments of the present application include specific features, structures or characteristics described in conjunction with the embodiment. Therefore, the statements "in one embodiment", "in some embodiments", "in some other embodiments", "in some other embodiments", etc. that appear in different places in this specification do not necessarily refer to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized in other ways. The terms "including", "comprising", "having" and their variations all mean "including but not limited to", unless otherwise specifically emphasized in other ways.

[0057] The embodiment of the present application provides a training method for an electromagnetic data classification model, which is used to improve the training efficiency and classification accuracy of the electromagnetic data classification model. The training method for the electromagnetic data classification model provided in the embodiment of the present application, by combining automatic labeling and manual labeling to collaboratively label the electromagnetic data in the first sample data set, obtains a training sample data set, obtains training sample data with accurate and consistent labeling data more quickly, then extracts the target feature set of the training sample data and generates a first feature vector, and trains the initial model based on the first feature vector to obtain a target classification model, thereby improving the training efficiency and classification accuracy of the target classification model (electromagnetic data classification model).

[0058] The training method of the electromagnetic data classification model provided in the embodiment of the present application can be applied to terminal devices such as laptop computers and personal computers (UMPCs). The embodiment of the present application does not impose any restrictions on the specific type of terminal devices.

[0059] like Figure 1 As shown, the embodiment of the present application provides a training method for an electromagnetic data classification model, comprising:

[0060] Step S10, collaboratively annotating the electromagnetic data in the first sample data set by combining automatic annotation and manual annotation to obtain a training sample data set; the training sample data set includes training sample data, and the training sample data has annotated data.

[0061] In the application, a variety of electromagnetic data are collected in real or simulated environments, including radar signals, communication signals, and navigation signals, to ensure the diversity and representativeness of the training sample data used for subsequent model training.

[0062] In the application, the electromagnetic data collected in the real environment or the simulated environment are preprocessed by signal conditioning, analog-to-digital conversion (ADC), noise removal, data cleaning, etc., and the preprocessed electromagnetic data is stored as the first sample data in the first sample data set. The preprocessing operation can improve the data quality, thereby improving the accuracy of the model training and ensuring the generalization ability of the model.

[0063] In the application, the electromagnetic data in the first sample data set are collaboratively annotated based on automatic annotation and manual annotation, and the annotated electromagnetic data are stored in the training sample data set as training sample data. The training sample data includes electromagnetic data and annotated data, and the annotated data includes the type of the electromagnetic data. The collaborative annotation takes into account both the efficiency and reliability of the annotation, improves the overall efficiency of the model training, and ensures the accuracy of the model.

[0064] Step S20: extracting a target feature set of the training sample data and generating a first feature vector.

[0065] In the application, the target feature set includes multiple types of features of the training sample data, such as time domain features, frequency domain features, etc. The extracted target features are concatenated into a first feature vector for training the target classification model.

[0066] Step S30: training an initial model based on the first feature vector to obtain a target classification model, wherein the target classification model is used to classify the input electromagnetic data.

[0067] In application, the training method of the electromagnetic data classification model provided in the embodiments of the present application can train a target classification model, which is used to classify the input electromagnetic data. In application, the target classification model can be deployed on mobile terminal devices, computer devices and other devices to realize the recognition of electromagnetic signals in different scenarios.

[0068] In the application, multiple first feature vectors and corresponding labeled data are divided into a training set, a validation set, and a test set. Since the values ​​in the first feature vector may differ greatly in magnitude (for example, the frequency feature and the pulse width have different magnitudes), each feature needs to be standardized to have a mean of 0, a variance of 1, and normalization, so that each feature is scaled to the interval of 0 to 1. If there are missing values ​​in the feature vector, mean filling, median filling, or machine learning algorithms (such as KNN filling) can be used to fill them; abnormal data can be detected and removed by methods such as box plots and Z-scores. In practical applications, some categories of signals may appear less frequently than other categories, which may lead to poor recognition of the model on the minority class. Oversampling can be used to balance the categories by generating new minority class samples, undersampling to reduce the number of majority class samples, or giving higher weights to minority class samples in the loss function.

[0069] In the application, the initial model is a deep learning model such as a convolutional neural network model or a multi-layer perceptron.

[0070] In the application, the initial model is trained based on the training set, and the parameters are adjusted to fit the data; the hyperparameters are tuned and overfitting is detected based on the validation set. During the training process, the performance of the initial model is evaluated on the validation set, and the best hyperparameters (such as learning rate, batch size, etc.) are selected. The gradient descent algorithm is used for the initial model, and the parameters of the initial model are updated according to the gradient value. The learning rate is gradually reduced during the training process to help the model adjust the parameters more finely when it is close to the optimal solution. To prevent overfitting, L2 regularization (weight decay) or L1 regularization can be used to prevent the model from overfitting on the training set or in deep neural network training, some neurons are randomly discarded to reduce overfitting. In the application, the cross-entropy loss function is used to optimize the classification task. During the training process, if the performance of the model on the validation set no longer improves, the training can be stopped in advance to avoid overfitting of the model on the training set.

[0071] In the application, you can choose to evaluate the correct proportion of model predictions through accuracy; evaluate the model's ability to identify positive classes through recall, especially in the case of class imbalance; evaluate the accuracy of the model's predictions of positive samples through precision; and evaluate the performance of the trained target classification model through one or more of the following methods: F1-score: the harmonic mean of precision and recall, which comprehensively evaluates model performance.

[0072] like Figure 2As shown, in one embodiment, the step S10, combining automatic labeling and manual labeling to collaboratively label the electromagnetic data in the first sample data set to obtain a training sample data set, includes:

[0073] Step S11, annotating the electromagnetic data in the first sample data set by using the target generator and the pre-trained annotation model to obtain electromagnetic data with preliminary annotated data.

[0074] In the application, several target generators and pre-trained annotation models are used to improve the annotation efficiency and the accuracy and consistency of preliminary annotation, reduce the workload of manual annotation, and improve the overall annotation efficiency.

[0075] Step S12, acquiring electromagnetic data with manually annotated data, wherein the electromagnetic data with manually annotated data is obtained by annotating a portion of the electromagnetic data with preliminary annotated data by annotators.

[0076] In the application, the preliminary annotation data includes the annotation confidence. For electromagnetic data with low confidence in the preliminary annotation data, the reliability and accuracy of the annotation data can be improved by obtaining manually annotated data.

[0077] Step S13: generating training sample data with final annotation data based on the electromagnetic data with preliminary annotation results and the electromagnetic data with manual annotation results.

[0078] Step S14: storing the training sample data with the final labeled data into a training sample data set.

[0079] In one embodiment, the step S11, annotating the electromagnetic data in the first sample data set by using the target generator and the pre-trained annotation model to obtain electromagnetic data with preliminary annotated data, includes:

[0080] Step S111 : annotating electromagnetic data in a first sample data set by a target generator to obtain electromagnetic data having first annotated data.

[0081] Step S112: annotate the electromagnetic data in the first sample data set using a pre-trained annotation model to obtain electromagnetic data with second annotated data.

[0082] Step S113: generating electromagnetic data with preliminary annotation data according to the confidence of the first annotation data and the confidence of the second annotation data.

[0083] In one embodiment, the step S111, before annotating the electromagnetic data in the first sample data set by the target generator to obtain the electromagnetic data with the first annotated data, further includes:

[0084] Step S1101, construct a generator and introduce an attention mechanism.

[0085] In the application, a deep convolutional generative adversarial network (DCGAN) is used to build a generator, which inputs random noise and outputs preliminary annotation results. By introducing the attention mechanism, the generator can pay more attention to the key features in the data.

[0086] Step S1102: construct a discriminator, wherein the discriminator includes a plurality of convolution kernels of different sizes for extracting features of different granularities.

[0087] In the application, the discriminator is used to input the real annotated data and the generated annotated data, and output a probability value representing the authenticity of the input data. In this application, the construction of the discriminator adopts multi-scale feature fusion technology, and 3x3, 5x5 and 7x7 convolution kernels are used in the discriminator to extract fine-grained, medium-grained and coarse-grained features respectively. Then a multi-scale pyramid structure is constructed, and each level corresponds to a feature map of a different scale. After extracting features through convolution layers of different scales, these feature maps are spliced ​​or fused. The feature maps of different scales are spliced ​​in the channel dimension to form a high-dimensional feature map. By learning the importance of features of different scales, the feature maps of different scales are weighted and fused. In the application, a small convolutional network can be used to learn weights; the fused feature map is globally pooled to extract global features; the features after global pooling are spliced ​​to form the final feature vector. The final feature vector is input to the fully connected layer, and a probability value representing the authenticity of the input data is output. Through the above method, the recognition ability of complex signals can be improved.

[0088] Step S1103, performing iterative adversarial training based on the generator and the discriminator, and outputting a target generator after a preset iteration termination condition is met.

[0089] One is to use labeled electromagnetic data as training data. When there is less training data, sample expansion can be performed through data enhancement. First, the original electromagnetic data is enhanced in the time domain, including random cropping, random translation, random scaling, noise addition, and inversion to change the time characteristics of the signal. Then, the signal after time domain enhancement is converted to the frequency domain, and frequency domain enhancement is performed, including frequency domain filtering, frequency domain noise addition, and frequency domain translation to further change the frequency domain characteristics of the signal. Finally, the signal after frequency domain enhancement is inversely transformed back to the time domain to obtain the final enhanced sample. Through the above method, training samples with diverse characteristics can be generated, the number and types of samples can be enriched, and the model's adaptability to electromagnetic data in different scenarios can be improved.

[0090] In the application, the generator and discriminator are optimized based on the loss function. The generator and discriminator are trained alternately. In each adversarial training iteration, the generator is fixed first and the discriminator is trained; then the discriminator is fixed and the generator is trained. In this way, the generator and discriminator can be gradually improved and eventually reach a balance. In the application, the parameters of the discriminator are updated by the gradient descent method to minimize its loss function LD; the parameters of the generator are updated by the gradient ascent method to maximize its loss function LG.

[0091] For example, the discriminator loss function includes:

[0092]

[0093] Where: x is a real data sample, from the real data distribution p data(x) ;

[0094] z is random noise;

[0095] G(z) is the sample generated by the generator;

[0096] D(x) is the output of the discriminator for the real data sample;

[0097] D(G(z)) is the output of the discriminator for the generated data sample;

[0098] λ is the weight hyperparameter of the gradient penalty, usually set to 10;

[0099] is the expected value of the discriminator D for the real data x. data(x) represents the true data distribution; this term encourages the discriminator D to assign higher scores (closer to 1) to the true data x.

[0100] is the expected value of the discriminator D for the generated data G(z). z (z) represents the noise distribution (usually a standard normal distribution), and G(z) is the data generated by the generator. This term encourages the discriminator D to assign a lower score (close to 0) to the generated data G(z).

[0101] is the gradient penalty term. It is the difference sample between the real data x and the generated data G(z), usually obtained by linear interpolation Where α is a scalar sampled from a uniform distribution U(0,1). This term is to force the gradient norm of the discriminator D to be close to 1, so that the Lipschitz constant of D is close to 1. It helps prevent the discriminator from overfitting and makes the training process more stable.

[0102] In the application, the loss function of the generator remains unchanged, and the output of the discriminator on the generated data needs to be maximized:

[0103]

[0104] By adopting Wasserstein distance (WGAN) or spectral normalization, the stability of training and the quality of generation can be improved.

[0105] In the application, the parameters of the generator and the discriminator are optimized through multiple iterations of adversarial training until the labeled data generated by the generator can deceive the discriminator. In the application, the learning rate can be dynamically adjusted through the adaptive learning rate adjustment mechanism to speed up the convergence speed.

[0106] In one embodiment, the step S112, annotating the electromagnetic data in the first sample data set by using a pre-trained annotation model to obtain electromagnetic data with second annotated data, includes:

[0107] Step S1121, inputting the electromagnetic data in the first sample data set into a pre-trained annotation model; freezing some layers of the pre-trained annotation model.

[0108] Step S1122: extracting high-level features of the electromagnetic data in the first sample data set based on the partially layer-frozen pre-trained annotation model.

[0109] Step S1123: generating first labeled data for the electromagnetic data in the first sample data set based on the high-level features, to obtain electromagnetic data with second labeled data.

[0110] In the application, a model pre-trained on a large-scale dataset (such as ImageNet, COCO, etc.) is used as the initial annotation model, such as BERT, and the model parameters are fine-tuned to adapt it to the annotation task of electromagnetic data. Specifically, first, the electromagnetic data is preprocessed into a format suitable for input into the pre-trained model, such as converting time series data into a two-dimensional image or spectrogram. Then, some or all layers of the pre-trained model are frozen to retain its high-level feature extraction capabilities learned on a large-scale dataset, and the output layer and other partial layers of the model are fine-tuned on the electromagnetic dataset. Through the above steps, the intermediate layers of the pre-trained model are used to extract high-level features of the electromagnetic data. These features have higher abstraction and representativeness, which help to improve the performance of subsequent electromagnetic signal classification models.

[0111] In one embodiment, the step S20 of extracting the target feature set of the training sample data and generating a first feature vector includes:

[0112] Step S21, extracting the time domain features and frequency domain features of the training sample data.

[0113] In applications, time domain feature extraction includes:

[0114] Amplitude changes: The amplitude fluctuations in the time domain are used to analyze the signal strength, signal switching state, etc. Especially for pulse signals, the waveform characteristics in the time domain are particularly critical.

[0115] Pulse characteristics, including pulse width, pulse interval, pulse repetition frequency (PRF), etc., can effectively distinguish different types of radar or communication signals. For example, a pulse signal has a specific width and repetition frequency, while a continuous wave signal has no obvious pulse characteristics.

[0116] In applications, frequency domain feature extraction includes:

[0117] Spectrum analysis analyzes the frequency distribution of a signal by performing Fourier transform on the signal. The peak value, bandwidth, center frequency, etc. in the spectrum can be used to identify radar, communication and navigation signals.

[0118] Bandwidth and modulation features,Extracting the bandwidth of the signal can help distinguish between broadband and narrowband signals. In addition, modulation methods (such as frequency shift keying, phase shift keying, etc.) can also be identified through frequency domain features.

[0119] Step S22: extracting target time-frequency features of the training sample data.

[0120] Step S23, extracting statistical features of the time domain features, the frequency domain features and the target time-frequency features.

[0121] Step S24: generating a first feature vector based on the time domain feature, the frequency domain feature, the target feature and the statistical feature.

[0122] In one embodiment, the step S22, extracting target time-frequency features of the training sample data, includes:

[0123] Step S221: extracting the first time-frequency feature of the training sample data.

[0124] In the application, the first time-frequency feature of the training sample data is extracted based on the short-time Fourier transform. Short-time Fourier transform is an extended Fourier transform method. Its principle is to divide the signal into a series of shorter time periods and perform Fourier transform on the signal in each time period to extract local frequency domain information. Short-time Fourier transform divides a long-time signal into multiple time windows of fixed length, and there will be a certain overlap between the above windows. In each time window, the signal is multiplied by a specific window function to reduce spectrum leakage. Subsequently, the signal in each window is fast Fourier transformed to obtain the spectrum information in the time period. By sliding the window, the change of the signal over time can be captured. The specific formula is as follows:

[0125]

[0126] In the formula, 1 is the coefficient, x(τ) is the input signal, w(t) is the window function, t is time, f is frequency, and X STFT(t,f) is the result of short-time Fourier transform.

[0127] w(t-τ) is the window function w(t) which is centered at time t and shifted at τ. The role of the window function is to localize the signal x(t) so that it is Fourier transformed around time t.

[0128] e -j2πfτ is a complex exponential function that is used to perform a Fourier transform of a signal x(τ) at frequency f by multiplying it by a complex number. j is the imaginary unit, f is the frequency, and τ is the time.

[0129] The window function w(t) in the short-time Fourier transform defines how to segment the signal in the time domain. Because there are inherent limitations on the signal in the time and frequency domain dimensions, the size and position of the window must be adjusted in practical applications. By selecting different window functions (such as rectangular windows, Hamming windows, etc.), the results of the short-time Fourier transform will have different effects. Generally speaking, since a shorter window only contains a part of the signal, it can provide higher frequency domain resolution, but the time domain resolution will be reduced accordingly; while a longer window has multiple cycles of signals and can have higher time domain resolution, but the frequency domain resolution will be reduced accordingly. The output of the short-time Fourier transform is a time-frequency matrix, which is called a "time-frequency diagram" or "spectrum diagram". Each point in the time-frequency diagram represents the local energy of the original signal at a specific time and frequency, and can clearly show the changes in the signal in time and frequency. In the time-frequency diagram, the time axis is usually in the horizontal direction, the frequency axis is in the vertical direction, and the color represents the signal energy. The most critical thing about the short-time Fourier transform is to balance the time resolution and frequency resolution. In practical applications, the corresponding window size is selected according to the characteristics of the electromagnetic signal and the classification task. For scenarios that require detailed analysis of instantaneous signal changes, a smaller time window is generally selected to improve time resolution; for scenarios that require distinguishing signal frequency components, a larger time window is selected to improve frequency resolution.

[0130] Step S222: extracting the second time-frequency features of the training sample data.

[0131] In the application, the second time-frequency feature of the training sample data is extracted based on the Hilbert-Huang transform. The Hilbert-Huang transform uses two methods, the empirical mode decomposition and the Hilbert transform signal analysis method. The Hilbert-Huang transform has better locality and adaptability, and can more accurately reflect the time-frequency characteristics of nonlinear and non-stationary signals. The Hilbert-Huang transform includes the following steps:

[0132] 1) Classical mode decomposition

[0133] Empirical mode decomposition is to complete signal decomposition on the data itself. Its basic principle is to decompose the signal into a set of intrinsic mode functions (IMF) and then extract the time-frequency characteristics of the signal.

[0134] The specific steps are as follows:

[0135] The first step is to extract the upper and lower envelopes of the original signal through multiple extreme points to construct the first intrinsic mode function b1;

[0136] Extract the extreme point sequence p1, p2, …, pM of the signal x(t), where M is the number of extreme points;

[0137] Use the maximum and minimum points to construct the upper and lower envelopes u1(t), l1(t) respectively, and calculate the average values ​​of the upper and lower envelopes:

[0138] m1(t)=(u1(t)+l1(t)) / 2;

[0139] Subtract the signal x(t) from the average value m1(t) to obtain the first eigenmode function:

[0140] b1(t)=(x(t)-m1(t);

[0141] The second step is to repeat the operation on b1 to obtain the second eigenmode function b2;

[0142] b1 is used as the new signal x(t), and step 1 is repeated to obtain the second intrinsic mode function b2;

[0143] If b2 does not satisfy the definition of the eigenmode function, b2 is used as a new signal and iterated again until it is satisfied;

[0144] The third step is to repeat step 2 to obtain n eigenmode functions b1, b2, ...b n , and a residual r(t);

[0145] Put b n-1 As the new signal x(t), repeat step 2 and calculate the nth eigenmode function b n ;

[0146] Define r(t) as the original signal x(t), subtracted from the first n eigenmode functions, that is,

[0147]

[0148] If r(t) does not satisfy the definition of the eigenmode function, r(t) is treated as a new signal and iterated until it satisfies the definition.

[0149] The final signal is an eigenmode function of the following form:

[0150]

[0151] Among them, b i (t) represents the i-th eigenmode function, r n (t) represents the remaining term after the nth intrinsic mode function. The obtained n intrinsic mode functions will be used as the input of the Hilbert transform. It does not require any assumptions or prior knowledge about the signal in advance, and can support the decomposition and analysis of any form of signal. In addition, since the intrinsic mode function has good local characteristics, it can better reflect the time-frequency structure of the signal.

[0152] 2) Hilbert spectrum analysis

[0153] The Hilbert-Huang transform can convert a real-valued signal into a complex-valued signal in the frequency domain, and can also extract the amplitude and phase information of the signal. By using the Hilbert spectrum analysis, the time-frequency characteristics of the signal group are obtained by empirical mode decomposition. According to the definition of the Hilbert transform, for a real signal x(t), its Hilbert transform H[x(t)] can be calculated by the following formula.

[0154]

[0155] In the formula, PV is the Cauchy principal value, and t represents the time axis after transformation;

[0156] is the integrand. Specifically, x(τ) is the value of the signal at time τ, and t-τ is the time difference.

[0157] To calculate H[x(t)], we need to first find the analytical signal z(t) of the real signal x(t). The analytical signal z(t) can be obtained from the real and imaginary parts of the Hilbert transform according to the following formula:

[0158] z(t)=x(t)+J·H[x(t)]=A(t)e jφ(t) ;

[0159] In the formula, is the signal amplitude, is the signal phase.

[0160] J is the imaginary unit, J2 = -1; e jφ(t) is a complex exponential function, which represents the instantaneous phase of the signal.

[0161] Therefore, each eigenmode function can be used to calculate its Hilbert transform according to the following process to obtain its Hilbert spectrum analysis.

[0162] Each eigenmode function b i (t) Calculate its Hilbert transform H[b i (t)];

[0163] According to H[b i (t)] calculate its analytical signal z i (t), and the amplitude A is obtained i (t) and phase φ i (t);

[0164] Calculate the Hilbert spectrum S of each eigenmode function i (ω).

[0165]

[0166] In the formula, ω is the frequency, that is, A at each moment i (t) and φ i (t) is derived from the rate of change of is the instantaneous frequency. i (t)) is the Dirac delta function, which represents the frequency ω and the instantaneous frequency ω i (t) The impulse when equal.

[0167] After the above steps, the Hilbert spectrum analysis of each eigenmode function can be analyzed, and the Hilbert spectra of all eigenmode functions are superimposed to obtain the final Hilbert spectrum formula: Hilbert spectrum analysis can analyze the energy distribution of a signal as it changes with time and frequency, thereby realizing time-frequency analysis of the signal. The Hilbert transform is a global transform that does not require a window function and can effectively avoid the boundary effect problem introduced by the window. It has the advantage of maintaining good time-frequency resolution. The result of the Hilbert spectrum analysis is a time-frequency representation that can clearly show the transient components and time-varying characteristics of the signal.

[0168] In the application, the first time-frequency feature and the second time-frequency feature are extracted in different ways, and the specific extraction method is not limited to extracting the first feature based on short-time Fourier transform and extracting the second feature based on Hilbert-Huang transform. According to the frequency change of the electromagnetic signal over time in the actual application, two different methods can be selected to extract the first time-frequency feature and the second time-frequency feature respectively.

[0169] Step S223: fusing the first time-frequency feature and the second time-frequency feature to obtain a target time-frequency feature.

[0170] In the application, the first time-frequency feature and the second time-frequency feature extracted based on two different methods, such as the first time-frequency feature extracted based on the short-time Fourier transform (STFT) and the second time-frequency feature extracted based on the Hilbert-Huang transform (HHT), are fused together to obtain the final target time-frequency feature. This fusion process can be achieved in the following ways:

[0171] Feature splicing: splicing the first time-frequency feature extracted by STFT (such as spectral energy, center frequency, etc.) and the second time-frequency feature extracted by HHT (such as instantaneous frequency, instantaneous amplitude, etc.) according to the feature dimension to form a high-dimensional time-frequency feature vector; or weighted fusion: weighted averaging the features extracted by STFT and HHT, weighting them according to their respective importance, and generating a fused feature representation.

[0172] Time synchronization,To ensure time alignment during feature fusion, the two feature sets can be synchronized on a time scale so that the first time-frequency feature and the second time-frequency feature at each time point correspond to the same signal state.

[0173] Feature selection and dimensionality reduction: If the feature dimension obtained by concatenation or weighted fusion is too high, feature selection or dimensionality reduction techniques (such as principal component analysis PCA, t-SNE, etc.) can be applied to reduce redundant features and improve model training efficiency.

[0174] The fused target time-frequency features contain the comprehensive time-frequency information of the signal, including the changes in time and frequency of the signal, as well as its instantaneous frequency and amplitude characteristics. These fused features will provide rich input for subsequent signal classification or target recognition tasks.

[0175] In the application, the electromagnetic signal is input based on the finally trained target classification model, and the category of the electromagnetic signal is output.

[0176] like Figure 3 The figure shows the overall process framework of the method provided by the present application, including data collection, collaborative labeling, feature extraction, model training and optimization, and signal recognition and analysis. Among them, data collection includes radar signals, navigation signals, broadcast signals, communication signals and other signals; signal preprocessing includes signal conditioning, analog-to-digital conversion, denoising, time-frequency analysis and other steps; collaborative labeling realizes human-computer interaction based on expert labeling and automatic labeling, and the final labeled electromagnetic signal is stored in the database; feature extraction includes time domain features, frequency domain features, time-frequency features and statistical features, and finally obtains the first feature vector; based on the first feature vector, the target classification model is trained and optimized; finally, the target classification model is obtained, which is used to identify electromagnetic signals and perform statistical analysis on the recognition results.

[0177] It should be understood that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0178] The present application also provides a training device for an electromagnetic data classification model, which is used to perform the steps in the training method embodiment of the electromagnetic data classification model. The training device for the electromagnetic data classification model can be a virtual appliance in the first terminal device, which is run by the processor of the first terminal device, or it can be the first terminal device itself.

[0179] like Figure 4 As shown, the training device 4 of the electromagnetic data classification model provided in the embodiment of the present application includes:

[0180] A collaborative annotation module 401 is used to collaboratively annotate the electromagnetic data in the first sample data set by combining automatic annotation and manual annotation to obtain a training sample data set; the training sample data set includes training sample data, and the training sample data has annotated data;

[0181] A feature extraction module 402 is used to extract a target feature set of the training sample data and generate a first feature vector;

[0182] The model training module 403 is used to train an initial model based on the first feature vector to obtain a target classification model, and the target classification model is used to classify the input electromagnetic data.

[0183] In one embodiment, the collaborative annotation module 401 includes:

[0184] A first labeling unit is used to label the electromagnetic data in the first sample data set by using a target generator and a pre-trained labeling model to obtain electromagnetic data with preliminary labeling data;

[0185] A second labeling unit is used to obtain electromagnetic data with manually labeled data, where the electromagnetic data with manually labeled data is obtained by labeling a portion of the electromagnetic data with preliminary labeled data by a labeling staff;

[0186] A labeling fusion unit, used for generating training sample data with final labeling data based on the electromagnetic data with preliminary labeling results and the electromagnetic data with manual labeling results;

[0187] The first storage unit is used to store the training sample data with the final labeled data into the training sample data set.

[0188] In one embodiment, the first marking unit is used to:

[0189] Annotating the electromagnetic data in the first sample data set by a target generator to obtain electromagnetic data with first annotated data;

[0190] Annotating the electromagnetic data in the first sample data set by using a pre-trained annotation model to obtain electromagnetic data with second annotated data;

[0191] Electromagnetic data with preliminary annotation data are generated according to the confidence level of the first annotation data and the confidence level of the second annotation data.

[0192] In one embodiment, the collaborative labeling module 401 further includes a target generator training unit, which is used to:

[0193] Build a generator and introduce an attention mechanism;

[0194] Constructing a discriminator, wherein the discriminator includes a plurality of convolution kernels of different sizes for extracting features of different granularities;

[0195] Iterative adversarial training is performed based on the generator and the discriminator, and a target generator is output after a preset iteration termination condition is met.

[0196] In one embodiment, the first marking unit is used to:

[0197] Inputting electromagnetic data in the first sample data set into a pre-trained annotation model; freezing some layers of the pre-trained annotation model;

[0198] extracting high-level features of the electromagnetic data in the first sample data set based on the partially layer-frozen pre-trained annotation model;

[0199] First annotated data of the electromagnetic data in the first sample data set is generated based on the high-level features to obtain electromagnetic data with second annotated data.

[0200] In one embodiment, the feature extraction module 402 is used to:

[0201] Extracting time domain features and frequency domain features of the training sample data;

[0202] Extracting target time-frequency features of the training sample data;

[0203] Extracting statistical features of the time domain features, the frequency domain features, and the target time-frequency features;

[0204] A first feature vector is generated based on the time domain feature, the frequency domain feature, the target feature and the statistical feature.

[0205] In one embodiment, the feature extraction module 402 is used to:

[0206] The extraction unit extracts the first time-frequency feature of the training sample data;

[0207] Extracting a second time-frequency feature of the training sample data;

[0208] The first time-frequency feature and the second time-frequency feature are fused to obtain a target time-frequency feature.

[0209] In application, each module in the training device of the electromagnetic data classification model can be a software program module, or can be implemented by different logic circuits integrated in a processor, or can be implemented by multiple distributed processors.

[0210] Figure 5 This is a schematic diagram of the structure of a terminal device provided in an embodiment of the present application. Figure 5As shown, the terminal device 5 of this embodiment includes: at least one processor 50 ( Figure 5 Only one is shown in the figure) a processor, a memory 51, and a computer program 52 stored in the memory 51 and executable on the at least one processor 50, and when the processor 50 executes the computer program 52, the steps in any of the above-mentioned method embodiments are implemented.

[0211] The terminal device 5 may be a computing device such as a desktop computer, a notebook, a PDA, or a cloud server. The terminal device may include, but is not limited to, a processor 50 and a memory 51. Those skilled in the art will appreciate that Figure 5 It is only an example of the terminal device 5 and does not constitute a limitation on the terminal device 5. It may include more or fewer components than shown in the figure, or a combination of certain components, or different components. For example, it may also include input and output devices, network access devices, etc.

[0212] The processor 50 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.

[0213] In some embodiments, the memory 51 may be an internal storage unit of the terminal device 5, such as a hard disk or memory of the terminal device 5. In other embodiments, the memory 51 may also be an external storage device of the terminal device 5, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. equipped on the terminal device 5. Further, the memory 51 may also include both an internal storage unit and an external storage device of the terminal device 5. The memory 51 is used to store an operating system, an application program, a boot loader (BootLoader), data, and other programs, such as the program code of the computer program. The memory 51 may also be used to temporarily store data that has been output or is to be output.

[0214] It should be noted that the information interaction, execution process, etc. between the above-mentioned devices / units are based on the same concept as the method embodiment of the present application. Their specific functions and technical effects can be found in the method embodiment part and will not be repeated here.

[0215] The technicians in the relevant field can clearly understand that for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In practical applications, the above-mentioned function allocation can be completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated in a processing unit, or each unit can exist physically separately, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, which will not be repeated here.

[0216] An embodiment of the present application further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps in the above-mentioned method embodiments can be implemented.

[0217] like Figure 6 As shown, an embodiment of the present application provides a computer program product 6, including a computer program 60. When the computer program 60 is executed, the steps in the above-mentioned various method embodiments are executed.

[0218] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present application implements all or part of the processes in the above-mentioned embodiment method, which can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned various method embodiments when executed by the processor. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may at least include: any entity or device that can carry the computer program code to the device / terminal device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electric carrier signal, a telecommunication signal, and a software distribution medium. For example, a USB flash drive, a mobile hard disk, a magnetic disk or an optical disk. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electric carrier signals and telecommunication signals.

[0219] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0220] Those of ordinary skill 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 in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0221] In the embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the modules or units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

Claims

1. A training method for an electromagnetic data classification model, characterized in that: include: The electromagnetic data in the first sample data set are collaboratively annotated by combining automatic annotation and manual annotation to obtain a training sample data set; the training sample data set includes training sample data, and the training sample data has annotated data; Extracting a target feature set of the training sample data and generating a first feature vector; An initial model is trained based on the first feature vector to obtain a target classification model, and the target classification model is used to classify input electromagnetic data.

2. The method for training an electromagnetic data classification model according to claim 1, characterized in that: The method of collaboratively annotating the electromagnetic data in the first sample data set by combining automatic annotation and manual annotation to obtain a training sample data set includes: Annotating the electromagnetic data in the first sample data set by using the target generator and the pre-trained annotation model to obtain electromagnetic data with preliminary annotated data; Acquiring electromagnetic data with manually annotated data, wherein the electromagnetic data with manually annotated data is obtained by annotating a portion of the electromagnetic data with preliminary annotated data by an annotator; Generating training sample data with final labeled data based on the electromagnetic data with preliminary labeled results and the electromagnetic data with manual labeled results; The training sample data with the final labeled data is stored in the training sample data set.

3. The method for training an electromagnetic data classification model according to claim 2, characterized in that: The step of labeling the electromagnetic data in the first sample data set by using the target generator and the pre-trained labeling model to obtain electromagnetic data with preliminary labeled data includes: Annotating the electromagnetic data in the first sample data set by a target generator to obtain electromagnetic data with first annotated data; Annotating the electromagnetic data in the first sample data set by using a pre-trained annotation model to obtain electromagnetic data with second annotated data; Electromagnetic data with preliminary annotation data are generated according to the confidence level of the first annotation data and the confidence level of the second annotation data.

4. The method for training an electromagnetic data classification model according to claim 3, characterized in that: Before the target generator is used to label the electromagnetic data in the first sample data set to obtain the electromagnetic data with the first labeled data, the method further includes: Build a generator and introduce an attention mechanism; Constructing a discriminator, wherein the discriminator includes a plurality of convolution kernels of different sizes for extracting features of different granularities; Iterative adversarial training is performed based on the generator and the discriminator, and a target generator is output after a preset iteration termination condition is met.

5. The method for training an electromagnetic data classification model according to claim 3, characterized in that: The method of labeling the electromagnetic data in the first sample data set by using the pre-trained labeling model to obtain electromagnetic data with second labeled data includes: Inputting electromagnetic data in the first sample data set into a pre-trained annotation model; freezing some layers of the pre-trained annotation model; extracting high-level features of the electromagnetic data in the first sample data set based on the partially layer-frozen pre-trained annotation model; First annotated data of the electromagnetic data in the first sample data set is generated based on the high-level features to obtain electromagnetic data with second annotated data.

6. The method for training an electromagnetic data classification model according to any one of claims 1 to 5, characterized in that: The step of extracting a target feature set of the training sample data and generating a first feature vector includes: Extracting time domain features and frequency domain features of the training sample data; Extracting target time-frequency features of the training sample data; Extracting statistical features of the time domain features, the frequency domain features, and the target time-frequency features; A first feature vector is generated based on the time domain feature, the frequency domain feature, the target feature and the statistical feature.

7. The method for training an electromagnetic data classification model according to claim 6, characterized in that: The step of extracting target time-frequency features of the training sample data includes: Extracting a first time-frequency feature of the training sample data; Extracting a second time-frequency feature of the training sample data; The first time-frequency feature and the second time-frequency feature are fused to obtain a target time-frequency feature.

8. A training device for an electromagnetic data classification model, characterized in that: include: A collaborative annotation module, used to collaboratively annotate the electromagnetic data in the first sample data set by combining automatic annotation and manual annotation to obtain a training sample data set; the training sample data set includes training sample data, and the training sample data has annotated data; A feature extraction module, used to extract a target feature set of the training sample data and generate a first feature vector; A model training module is used to train an initial model based on the first feature vector to obtain a target classification model, and the target classification model is used to classify the input electromagnetic data.

9. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer program product, characterized in that The invention comprises a computer program, which, when executed, enables the method according to any one of claims 1 to 7 to be performed.

Images