Intermediate frequency feature comparative learning method based on multi-scale receptive field and context awareness

Abstract

The invention provides an intermediate frequency characteristic contrast learning method based on a multi-scale receptive field and context awareness. The method comprises the following steps: acquiring a plurality of input electromagnetic signals; estimating the center frequency of the input electromagnetic signal; performing signal preprocessing on the input electromagnetic signal according to the center frequency of the input electromagnetic signal to obtain intermediate frequency I / Q waveform data of the input electromagnetic signal; performing complex decomposition and maximum amplitude normalization on the intermediate frequency I / Q waveform data to form an intermediate frequency signal sample; performing intermediate-frequency high-order feature extraction based on a multi-scale receptive field and context perception on the intermediate-frequency signal sample to obtain a feature vector; performing signal clustering on the feature vectors of the unknown signals according to the feature vectors of the known signals in the plurality of input electromagnetic signals; wherein the signal clustering is carried out according to the cosine similarity of the known signal feature vector and the unknown signal feature vector. Under the condition of different signal-to-noise ratios, the robust performance of extracting the intermediate-frequency high-order features is good, and the generalization ability is excellent.

Description

Technical Field

[0001] The present invention relates to the technical field of signal processing, and in particular, to a method for contrastive learning of intermediate-frequency features based on multi-scale receptive fields and context awareness. Background Art

[0002] Radar signal feature learning can be roughly divided into electromagnetic parameter extraction based on traditional methods and implicit feature extraction based on deep learning. Among them, traditional methods mainly measure radar pulse PDW parameters or extract intra-pulse modulation features based on expert experience, while deep learning methods use artificial intelligence models such as CNN, LSTM, Transformer, etc., without manually extracting features, and learn deeper implicit features that can reflect the inherent characteristics of the radiation source. The prior art is described in detail as follows:

[0003] A. Electromagnetic parameter extraction based on traditional methods.

[0004] In the early days when the electromagnetic environment was relatively simple, traditional methods mainly used Pulse Descriptor Words (PDWs) to achieve the sorting and recognition of radar signals. PDW parameters mainly consist of Direction of Arrival (DOA), Radio Frequency (RF), Pulse Width (PW), Pulse Repetition Interval (PRI), and Pulse Amplitude (PA). Based on the acquisition of PDW parameters, the corresponding radar radiation sources are sorted / recognized by analyzing the characteristic parameters of the pulses. However, with the increase in the number of radars on the battlefield and the emergence of new types of radars, the traditional signal processing relying solely on PDW parameters has difficulty adapting to the changing radar signals. To solve the above problems, scholars at home and abroad have explored more intra-pulse characteristic parameters for signal sorting and recognition, such as extracting the pulse change time, change amplitude, and angle of radar signals, or extracting the wavelet packet features, complexity features, entropy features, intra-pulse similarity coefficients, and intra-pulse instantaneous frequencies of radar signals. However, using the intra-pulse modulation type and waveform parameters of radiation source signals as recognition features has the problem of limited information expression ability, and thus it is difficult to deal with a large number of new signal waveforms generated by advanced electromagnetic devices due to their complex modulation methods and variable emission characteristics or signals that are almost completely similar to other radiation sources. To address such challenges, fingerprint features, which refer to the inherent differences of radar devices brought about by the discreteness of device characteristics and technical indicators caused by process differences and defects, are independent of the modulation type of radar signals, have strong stability, and do not change significantly due to environmental changes, and may be a feasible direction for feature extraction. Extracting intra-pulse fingerprint features from the perspective of time-frequency analysis is a current research hotspot, such as Fourier Transform (FFT), Short-Time Fourier Transform (STFT), Wigner Distribution (WVD), Wavelet Transform (WT), and Hilbert-Huang Transform (HHT), etc. However, the related methods based on explicit mining of fingerprint features still perform unsatisfactorily when dealing with complex and dense electromagnetic environments and advanced radiation source targets.

[0005] B. Implicit Feature Extraction Based on Deep Learning.

[0006] With the widespread application of complex radars, the signal density in the electromagnetic threat environment has reached the million level. At the same time, with the diversification of electromagnetic signals in aspects such as time domain, frequency domain, spatial distribution, and modulation mode, the electromagnetic environment has become increasingly complex. Compared with traditional methods, deep learning has significant advantages in processing complex data. By adopting a deep neural network structure, it can effectively improve the ability to process complex high-dimensional data. Then, deep learning methods can be used to mine and learn deeper hidden features from the original intermediate-frequency signal data of radar pulses or from time-frequency images that can characterize the "fingerprint" features of radars, without manually extracting features from the data, so as to improve the signal analysis and processing ability. In view of this, more and more researchers have begun to use signal feature extraction methods based on deep learning to solve the deficiencies of traditional methods. Although some achievements have been made in radar signal recognition based on deep learning, the recognition accuracy is still relatively low under low signal-to-noise ratio conditions, and the robustness and stability of deep learning models in extracting hidden features remain a technical problem to be overcome.

[0007] To sum up, scholars and research institutions at home and abroad have carried out extensive research on radar signal feature representation and achieved certain research results. However, most methods use traditional PDW parameters and intra-pulse modulation features of radiation source signals as signal representations, which are difficult to cope with the radar signal analysis and processing problems in complex and changeable electromagnetic scenarios. In view of this, the present invention aims to use deep learning-related technologies to construct a reliable and robust intermediate-frequency high-order feature extraction model for radar pulse signals, with the expectation of making certain breakthroughs in the intelligent analysis of complex radar signals. Summary of the Invention

[0008] The present invention aims to solve at least one of the above technical problems existing in the prior art.

[0009] For this purpose, the present invention provides a method for intermediate-frequency feature contrast learning based on multi-scale receptive fields and context awareness.

[0010] The present invention provides a method for intermediate-frequency feature contrast learning based on multi-scale receptive fields and context awareness, including:

[0011] Obtain a plurality of input electromagnetic signals;

[0012] Estimate the center frequency of the input electromagnetic signals;

[0013] Perform signal preprocessing on the input electromagnetic signals according to the center frequency of the input electromagnetic signals to obtain intermediate-frequency I / Q waveform data of the input electromagnetic signals; the signal preprocessing includes at least digital down-conversion and filtering;

[0014] Perform complex decomposition and maximum amplitude normalization on the intermediate-frequency I / Q waveform data to form intermediate-frequency signal samples;

[0015] Perform intermediate frequency high-order feature extraction based on multi-scale receptive fields and context awareness on the intermediate frequency signal samples to obtain a feature vector;

[0016] Cluster the feature vectors of the unknown signals according to the feature vectors of the known signals in a number of input electromagnetic signals; wherein, signal clustering is performed according to the cosine similarity between the feature vectors of the known signals and the feature vectors of the unknown signals.

[0017] According to the intermediate frequency feature contrast learning method based on multi-scale receptive fields and context awareness of the above technical solutions of the present invention, the following additional technical features may also be provided:

[0018] In the above technical solution, the estimation of the center frequency of the input electromagnetic signal includes:

[0019] Perform fast Fourier transform processing on the input electromagnetic signal, and take the serial number corresponding to the first maximum value, denoted as Idx;

[0020] Estimate the center frequency of the input electromagnetic signal according to Idx:

[0021]

[0022] Wherein, fc represents the estimated result of the center frequency of the input electromagnetic signal; N represents the length of the sampling points of the input electromagnetic signal; Fs represents the sampling rate of the input electromagnetic signal.

[0023] In the above technical solution, the signal preprocessing of the input electromagnetic signal according to the center frequency of the input electromagnetic signal to obtain the intermediate frequency I / Q waveform data of the input electromagnetic signal includes:

[0024] Perform down-conversion processing on the input electromagnetic signal according to the center frequency of the input electromagnetic signal to obtain a complex signal X1(n):

[0025]

[0026] Wherein, X(n) represents the input electromagnetic signal, (n = 1,..., N); j represents the imaginary unit; n represents the serial number of the current input electromagnetic signal;

[0027] Perform filtering processing on X1(n) and perform 4-fold decimation to obtain Wherein, Represents rounding down.

[0028] In the above technical solution, the complex decomposition and maximum amplitude normalization of the intermediate frequency I / Q waveform data to form intermediate frequency signal samples include:

[0029] Decompose the intermediate frequency I / Q waveform data into a real part I(t) and an imaginary part Q(t);

[0030] Perform maximum amplitude normalization on the real part I(t) and the imaginary part Q(t):

[0031]

[0032] where I(t) ′ represents the result after normalization of the real part I(t); Q(t) ′ represents the result after normalization of the imaginary part Q(t);

[0033] According to the normalization results, cascade along the real part and imaginary part dimensions to form an intermediate frequency signal sample X = [I(t) ′ , Q(t) ′ .

[0034] In the above technical solution, performing intermediate frequency high-order feature extraction based on multi-scale receptive fields and context awareness on the intermediate frequency signal sample to obtain a feature vector includes:

[0035] Construct an intermediate frequency high-order feature extraction model. In the intermediate frequency high-order feature extraction model, capture the dependency relationships of the multi-scale receptive fields of the input electromagnetic signal through Inception modules with nested different-scale convolution operations; enable the model to learn the importance of different features through a parallel branch structure and perform feature extraction at different abstraction levels; in the intermediate frequency high-order feature extraction model, also adopt a bidirectional long short-term memory network module Bi-LSTM to capture the context information within different-scale features and learn global sequential information;

[0036] Input the intermediate frequency signal sample into the intermediate frequency high-order feature extraction model for spatial mapping to obtain a feature vector.

[0037] In the above technical solution, the intermediate frequency high-order feature extraction model includes:

[0038] A multi-scale convolution module, at least including convolutional layers with three different sizes of convolutional kernels. Each convolutional layer outputs features of different dimensions and cascades them in the channel dimension to obtain intermediate depth features; among them, the dimension of the depth features is the sum of the dimensions of all convolutional layers; adopt a batch normalization layer to accelerate the training of the multi-scale convolution module and prevent overfitting; adopt a ReLU activation function after the BN layer:

[0039] ReLU(X ′ ) = max(0, X ′ )

[0040] where X ′ represents the intermediate depth features obtained after the intermediate frequency pulse sample X passes through the multi-size convolutional layer and the batch normalization layer;

[0041] The convolutional block is connected to the multi-scale convolutional module to downsample the intermediate depth features to obtain the final depth features;

[0042] The bidirectional long short-term memory network module is connected to the convolutional block to capture the forward information from the beginning to the end of the sequence and the reverse information from the end to the beginning of the sequence according to the depth features:

[0043]

[0044] Among them, represents the hidden state at the current sliding time window w obtained based on the forward information; represents the hidden state at the current sliding time window w obtained based on the reverse information; represents the hidden state at the past sliding time window w - 1, represents the hidden state at the future sliding time window w + 1.

[0045] In the above technical solution, it further includes:

[0046] Create a feature similarity discrimination task to determine whether the input signal data pairs come from the same radiation source target, and then use the contrast loss of "intra-class aggregation and inter-class separation" to constrain the learning of intermediate-frequency high-order features, and use the cross-entropy loss constraint of similarity confidence prediction and the learning of the intermediate-frequency high-order feature extraction model.

[0047] In the above technical solution, in the feature similarity discrimination task, the intermediate-frequency I / Q waveform data pairs are input into two intermediate-frequency high-order feature extraction models, and the two intermediate-frequency high-order feature extraction models have the same structure and share weights.

[0048] In the above technical solution, the use of the contrast loss of "intra-class aggregation and inter-class separation" to constrain the learning of intermediate-frequency high-order features, and the use of the cross-entropy loss constraint of similarity confidence prediction and the learning of the intermediate-frequency high-order feature extraction model includes:

[0049] Randomly construct a number of positive sample pairs (X1 + , X2 + ) and negative sample pairs (X1 - , X2 - ) from the intermediate-frequency I / Q waveform data set, where the two intermediate-frequency I / Q waveform data in the positive sample pair belong to the same radiation source; the two intermediate-frequency I / Q waveform data in the negative sample pair do not belong to the same radiation source;

[0050] Input the sample pairs into the two intermediate-frequency high-order feature extraction models;

[0051] Construct the cross-entropy loss function for predicting the similarity of the intermediate-frequency high-order feature extraction model:

[0052] LCE = -[ylog(f(p + )) + (1 - y)log(1 - f(p - ))]

[0053] p + = (X1 + , X2 + )

[0054] p - = (X1 - , X2 - )

[0055] where L CE (·) represents the cross - entropy loss; y represents the true label of the sample pair. If the data in the sample pair belongs to the same radiation source, then y = 1; otherwise, y = 0; f(·) represents the contrastive learning strategy based on the siamese network;

[0056] Construct the loss function for contrastive learning of intermediate - frequency high - order features in the intermediate - frequency high - order feature extraction model:

[0057] L C = (1 - y)·(s - ) 2 + y·max(0, 1 - s + ) 2

[0058] s + = |H(f(p + ))|

[0059] s - = |H(f(p - ))|

[0060] where L C represents the loss function for contrastive learning of intermediate - frequency high - order features; s + represents the similarity of the positive sample pair; s - represents the similarity of the negative sample pair; H(·) represents the cosine similarity calculation function;

[0061] Construct the overall loss function of the intermediate - frequency high - order feature extraction model:

[0062] L = αL CE + (1 - α)L C

[0063] where L represents the overall loss function of the intermediate - frequency high - order feature extraction model; α represents the weight adjustment coefficient between the two loss functions;

[0064] By optimizing the overall loss function and adjusting the contrastive learning strategy, the distance between positive sample pairs in intermediate-frequency high-order features is reduced, while the distance between negative sample pairs in intermediate-frequency high-order features is increased.

[0065] In the above technical solution, the signal clustering of the feature vector of the unknown signal according to the feature vectors of the known signals in several input electromagnetic signals includes:

[0066] Traverse the intermediate-frequency high-order features based on the chronological order, and calculate the absolute value of the correlation coefficient between the current feature vector and the central feature vector in the classified list as the feature similarity each time;

[0067] If the calculated feature similarity is greater than the upper similarity threshold, add it to the end of the list with the highest similarity;

[0068] If the calculated feature similarity is less than the lower similarity threshold, set it as a new category, create an empty list and add this feature;

[0069] If the calculated feature similarity is between the upper and lower similarity thresholds, it is considered that this feature does not have obvious distinguishability.

[0070] In summary, due to the adoption of the above technical features, the beneficial effects of the present invention are:

[0071] The present invention constructs an intermediate-frequency high-order feature extraction model based on deep learning, and the network includes a multi-scale convolutional receptive field and a context-aware attention structure design; and proposes a contrastive learning strategy based on a Siamese network, where the feature extractors between pulse sample pairs share weight parameters, and the network model is trained with a contrastive loss of "intra-class aggregation and inter-class divergence" combined with a cross-entropy loss for feature similarity prediction; the robust performance of the present invention in extracting intermediate-frequency high-order features under different signal-to-noise ratio conditions is good, and it has excellent generalization ability. The high-order features of the radar monopulse intermediate-frequency signal are signal feature representations extracted from I / Q waveform data that can capture the differences between radiation source individuals, have the natural advantage of signal clustering and sorting at the individual level, and have the ability to generalize to the clustering and sorting of unknown targets.

[0072] The additional aspects and advantages of the present invention will become apparent in the following description section, or be learned through the practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] The above and / or additional aspects and advantages of the present invention will become apparent and be readily understood from the description of the embodiments in conjunction with the following drawings, where:

[0074] Figure 1 is a schematic diagram of the network architecture of the intermediate-frequency feature contrastive learning method based on multi-scale receptive fields and context awareness according to an embodiment of the present invention;

[0075] Figure 2 It is a schematic diagram of the statistical quantity of data of 10 signal sources in a specific embodiment of the present invention;

[0076] Figure 3 It is a schematic diagram of the visual comparison of I / Q signal samples of 10 signal sources in a specific embodiment of the present invention. Specific embodiments

[0077] In order to more clearly understand the above objects, features and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, without conflict, the embodiments of the present application and the features in the embodiments can be combined with each other.

[0078] In the following description, many specific details are set forth in order to fully understand the present invention. However, the present invention can also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited by the specific embodiments disclosed below.

[0079] The following refers to Figures 1 to 3 to describe a mid-frequency feature contrast learning method based on multi-scale receptive fields and context awareness provided according to some embodiments of the present invention.

[0080] Some embodiments of the present application provide a mid-frequency feature contrast learning method based on multi-scale receptive fields and context awareness.

[0081] The first embodiment of the present invention proposes a mid-frequency feature contrast learning method based on multi-scale receptive fields and context awareness, including the following steps S1-S6.

[0082] S1. Obtain a plurality of input electromagnetic signals. Specifically, the feature extraction method of the present disclosure can be adapted to real-time input electromagnetic signals. In the present disclosure, the real-time input radar monopulse raw intermediate frequency I / Q signal is taken as an example of the input electromagnetic signal for illustration. In a specific embodiment, assume that the real-time input electromagnetic signal is X(n), (n = 1,..., N), where N is the length of the sampling points and the sampling rate Fs = 600 MHz.

[0083] S2. Estimate the center frequency of the input electromagnetic signal.

[0084] In some embodiments, step S2 includes:

[0085] Perform fast Fourier transform processing on the input electromagnetic signal, and take the serial number corresponding to the first maximum value, denoted as Idx;

[0086] Estimate the center frequency of the input electromagnetic signal according to Idx:

[0087]

[0088] Among them, fc represents the estimated result of the center frequency of the input electromagnetic signal; N represents the length of the sampling points of the input electromagnetic signal; Fs represents the sampling rate of the input electromagnetic signal.

[0089] S3. Perform signal preprocessing on the input electromagnetic signal according to the center frequency of the input electromagnetic signal to obtain the intermediate frequency I / Q waveform data of the input electromagnetic signal; the signal preprocessing includes at least digital down-conversion and filtering.

[0090] In some embodiments, step S3 includes:

[0091] Perform down-conversion processing on the input electromagnetic signal according to the center frequency fc of the input electromagnetic signal to obtain a complex signal X1(n):

[0092]

[0093] Among them, X(n) represents the input electromagnetic signal, (n = 1,..., N); j represents the imaginary unit; n represents the serial number of the current input electromagnetic signal;

[0094] Perform filtering processing on X1(n) and perform 4-fold decimation to obtain Among them, represents rounding down.

[0095] S4. Perform complex decomposition and maximum amplitude normalization on the intermediate frequency I / Q waveform data to form an intermediate frequency signal sample.

[0096] In some embodiments, step S4 includes:

[0097] Decompose the intermediate frequency I / Q waveform data X2(m) into a real part I(t) and an imaginary part Q(t);

[0098] Perform maximum amplitude normalization on the real part I(t) and the imaginary part Q(t):

[0099]

[0100] Among them, I(t) ′ represents the result after normalization of the real part I(t); Q(t) ′ represents the result after normalization of the imaginary part Q(t);

[0101] According to the normalization result, cascade along the real part and imaginary part dimensions to form an intermediate frequency signal sample X = [I(t) ′ , Q(t) ′ , and the intermediate frequency signal sample is a two-dimensional vector.

[0102] Obtaining input samples adapted to the deep learning model through step S4 can effectively improve the model convergence speed and training stability.

[0103] S5. Perform intermediate-frequency high-order feature extraction based on multi-scale receptive fields and context awareness on the intermediate-frequency signal samples to obtain a feature vector FeaVec. In a specific embodiment, the feature vector FeaVec can be expressed as [a1, a2, …, a K , where K = 64 is the length of the feature vector.

[0104] In some embodiments, step S5 includes:

[0105] Construct an intermediate-frequency high-order feature extraction model. In the intermediate-frequency high-order feature extraction model, capture the dependency relationships of multi-scale receptive fields of the input electromagnetic signals through Inception modules that nest different-scale convolution operations; enable the model to learn the importance of different features through this parallel branch structure and perform feature extraction at different abstraction levels; at the same time, a bidirectional long short-term memory network module Bi-LSTM is also adopted in the intermediate-frequency high-order feature extraction model to capture the context information within different-scale features and learn more global sequential information;

[0106] Input the intermediate-frequency signal samples into the intermediate-frequency high-order feature extraction model for spatial mapping to obtain a feature vector.

[0107] As Figure 1 shown, in some embodiments, the intermediate-frequency high-order feature extraction model includes: a multi-scale convolution module, a convolution block, and a bidirectional long short-term memory network module.

[0108] The multi-scale convolution module (Multi-scale Convolution, MSC) includes at least three convolutional layers (Convolutioanal Layer, Conv) with different-sized convolutional kernels, such as 3×3, 5×5, 7×7; each convolutional layer outputs features with dimensions of 174, 173, and 172 respectively, and performs concatenation in the channel dimension to obtain intermediate-depth features; among them, the dimension of the depth features is the sum of the dimensions of all convolutional layers. In this embodiment, the dimension of the depth features is 519; then, a batch normalization layer (Batch Normalization, BN) is adopted to accelerate the training of the multi-scale convolution module and prevent overfitting; after the BN layer, a ReLU activation function is used to prevent gradient disappearance or gradient explosion, and the calculation formula is as follows:

[0109] ReLU(X ′ ) = max(0, X ′ )

[0110] where X ′It represents the intermediate depth features obtained after the intermediate frequency pulse sample X passes through the multi-size convolutional layer and the batch normalization layer;

[0111] The convolutional block (Convolutioanal Block), which is connected to the multi-scale convolutional module, downsamples the intermediate depth features to obtain the final depth features; in a specific embodiment, the depth features of different scales extracted above are input into two convolutional blocks for downsampling. The two convolutional blocks are a convolutional block with 32-dimensional channels and a 7×7 convolutional kernel and a convolutional block with 64-dimensional channels and a 3×3 convolutional kernel, and the final model outputs depth features of 128 dimensions.

[0112] The bidirectional long short-term memory network module, which is connected to the convolutional block, captures the forward information from the beginning to the end of the sequence and the reverse information from the end to the beginning of the sequence according to the depth features. Specifically, the single-pulse signal sample is essentially a sequence data. By using Bi-LSTM to capture the forward information from the beginning to the end of the sequence and the reverse information from the end to the beginning of the sequence simultaneously, more robust and effective intermediate frequency high-order features can be extracted. In some embodiments, the calculation formula of the bidirectional long short-term memory network module is:

[0113]

[0114] where, represents the hidden state at the current sliding time window w obtained based on the forward information; represents the hidden state at the current sliding time window w obtained based on the reverse information; represents the hidden state at the past sliding time window w - 1, represents the hidden state at the future sliding time window w + 1.

[0115] In some embodiments, to improve the effective separability of the extracted intermediate frequency high-order features, the present disclosure creates a feature similarity discrimination task based on the intermediate frequency high-order feature extraction model, determines whether the input signal data pair comes from the same radiation source target (outputs 1 if it belongs to the same radiation source, otherwise outputs 0), and then uses the contrastive loss of "intra-class aggregation and inter-class separation" to constrain the learning of the intermediate frequency high-order features, and uses the cross-entropy loss constraint of the similarity confidence prediction and the learning of the intermediate frequency high-order feature extraction model.

[0116] In a specific embodiment, in the feature similarity discrimination task, the intermediate frequency I / Q waveform data pair is input into two intermediate frequency high-order feature extraction models, and the two intermediate frequency high-order feature extraction models have the same structure and share weights.

[0117] In some embodiments, the learning of intermediate-frequency high-order features is constrained by a contrastive loss that "aggregates within classes and separates between classes", and the learning of the intermediate-frequency high-order feature extraction model is constrained by a cross-entropy loss for similarity confidence prediction, including:

[0118] Randomly construct a number of positive sample pairs (X1 + , X2 + ) and negative sample pairs (X1 - , X2 - ) from the intermediate-frequency I / Q waveform dataset, where the two intermediate-frequency I / Q waveform data in a positive sample pair belong to the same radiation source; the two intermediate-frequency I / Q waveform data in a negative sample pair do not belong to the same radiation source;

[0119] Input the sample pairs into two intermediate-frequency high-order feature extraction models;

[0120] Construct a cross-entropy loss function for similarity prediction of the intermediate-frequency high-order feature extraction model:

[0121] L CE = -[ylog(f(p + )) + (1 - y)log(1 - f(p - ))]

[0122] p + = (X1 + , X2 + )

[0123] p - = (X1 - , X2 - )

[0124] where L CE (·) represents the cross-entropy loss; y represents the true label of the sample pair, y = 1 if the data in the sample pair belong to the same radiation source, otherwise y = 0; f(·) represents a contrastive learning strategy based on a siamese network;

[0125] Construct a loss function for contrastive learning of intermediate-frequency high-order features in the intermediate-frequency high-order feature extraction model:

[0126] L C = (1 - y)·(s - ) 2 + y·max(0, 1 - s + ) 2

[0127] s + = |H(f(p + ))|

[0128] s - = |H(f(p - ))|

[0129] Among them, L C represents the loss function of intermediate-frequency high-order feature contrast learning; s + represents the similarity of positive sample pairs; represents; s - represents the similarity of negative sample pairs; H(·) represents the cosine similarity calculation function;

[0130] Construct the overall loss function of the intermediate-frequency high-order feature extraction model:

[0131] L = αL CE +(1 - α)L C

[0132] Among them, L represents the overall loss function of the intermediate-frequency high-order feature extraction model; α represents the weight adjustment coefficient between the two loss functions;

[0133] By optimizing the overall loss function and adjusting the contrast learning strategy, the distance of the intermediate-frequency high-order features of the positive sample pairs is reduced, and at the same time, the distance of the intermediate-frequency high-order features of the negative sample pairs is increased.

[0134] S6. According to the feature vectors of the known signals in a plurality of input electromagnetic signals, perform signal clustering on the feature vectors of the unknown signals; among them, signal clustering is performed according to the cosine similarity between the known signal feature vectors and the unknown signal feature vectors.

[0135] In some embodiments, the real-time input electromagnetic signals are processed according to steps S1 - S6. For the p-th input electromagnetic signal, the intermediate-frequency high-order feature obtained after processing is FeaVec p = [a 1,p , a 2,p , …, a K,p , (p = 1, 2, …, P), where P is the number of known signal samples. The unknown signal samples to be clustered are also processed according to steps 1 - 5 to obtain the intermediate-frequency high-order feature as FeaVec ID = [a 1,ID , a 2,ID , …, a K,ID . It should be noted that the known signal refers to the radiation source of the current electromagnetic signal that can be directly marked and classified, and vice versa is an unknown signal.

[0136] The performing signal clustering on the feature vectors of the unknown signals according to the feature vectors of the known signals in a plurality of input electromagnetic signals includes:

[0137] Traverse the intermediate-frequency high-order features based on the chronological order, and calculate the absolute value of the correlation coefficient between the current feature vector and the central feature vector in the classified list as the feature similarity each time;

[0138] If the calculated feature similarity is greater than the upper threshold of similarity, add it to the end of the list of the category with the highest similarity;

[0139] If the calculated feature similarity is less than the lower threshold of similarity, set it as a new category, create an empty list and add the feature;

[0140] If the calculated feature similarity is between the upper and lower thresholds of similarity, it is considered that the feature does not have obvious distinguishability. The part of the signal can be marked or discarded.

[0141] In a specific embodiment, to verify the effectiveness of the multi-scale receptive field and context-aware intermediate frequency feature contrast learning method proposed in the present disclosure, a large number of simulated signal data are generated by a radar signal source to form a training / verification data set. Specifically, 10 different signal sources are used to collect signal samples in a large frequency range, with different pulse widths and various modulation types, and a total of 544,811 signal samples are collected. The data statistical results are as Figure 2 shown. It can be seen that the data set constructed in this embodiment is a relatively balanced data set, and each signal source contains at least 48,727 and at most 64,018 I / Q signal samples; at the same time, it can be seen from the visualization of the I / Q waveform (as Figure 3 shown) that the waveforms of each signal source are highly similar and difficult to distinguish. Positive and negative sample pairs required for training and verification are constructed based on the constructed large-scale signal source data.

[0142] Construction of positive sample pairs: Randomly select any signal source category, and randomly select two samples from the samples under this category as a positive sample pair. This random selection method ensures that the sampling of each category is balanced.

[0143] Construction of negative sample pairs: Randomly select two different radiation source categories, and randomly select one sample from the samples under each of the two categories as a negative sample pair. This method also ensures that the sampling of categories is balanced.

[0144] The intermediate-frequency high-order feature extraction model proposed by the present disclosure does not rely on traditional electromagnetic parameters, directly extracts high-order features that can reflect the essential differences of radiation sources from the original intermediate-frequency signal data, and improves the effective separability of intermediate-frequency high-order features through the task learning of feature similarity discrimination. In this embodiment, any 6 signal source samples are selected from the analog signal source dataset for training and testing, and the samples are discriminated according to the method described in steps S1-S6. The discrimination accuracy of the intermediate-frequency high-order features can be as high as 90.17%, and it shows strong robustness under different signal-to-noise ratio conditions (i.e., 17 dB to 51 dB). Further, the trained model is directly generalized to the remaining 4 signal source samples that have not participated in the training at all, and the discrimination accuracy of the intermediate-frequency high-order features can be as high as 82.1%. Sufficient experiments prove that the intermediate-frequency feature extraction method designed by the present invention can well extract effective and separable essential features of radiation sources, and has strong robustness and generalization ability. The experimental results are shown in the following table:

[0145]

[0146]

[0147] In this specification, the schematic descriptions of the above terms do not necessarily refer to the same embodiments or examples. Moreover, the specific features, structures, materials, or characteristics described may be combined in a suitable manner in any one or more embodiments or examples.

[0148] Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A mid-frequency feature contrast learning method based on multi-scale receptive field and context perception, characterized in that: include: obtaining a number of input electromagnetic signals; estimating a center frequency of the input electromagnetic signal; Preprocessing the input electromagnetic signal according to the center frequency of the input electromagnetic signal to obtain intermediate frequency I / Q waveform data of the input electromagnetic signal; the signal preprocessing includes at least digital down-conversion and filtering; Performing complex number decomposition and maximum amplitude normalization on the intermediate frequency I / Q waveform data to form intermediate frequency signal samples; Extracting intermediate frequency high-order features based on multi-scale receptive field and context perception from the intermediate frequency signal samples to obtain feature vectors; According to the characteristic vectors of known signals in a number of input electromagnetic signals, the characteristic vectors of unknown signals are clustered; wherein, the signal clustering is performed according to the cosine similarity between the characteristic vectors of the known signals and the characteristic vectors of the unknown signals.

2. The method for comparing and learning intermediate frequency features based on multi-scale receptive field and context perception according to claim 1, characterized in that: The estimating the center frequency of the input electromagnetic signal comprises: Perform fast Fourier transform on the input electromagnetic signal and take the serial number corresponding to the first maximum value, which is recorded as Idx; Estimate the center frequency of the input electromagnetic signal based on Idx: Wherein, fc represents the estimation result of the center frequency of the input electromagnetic signal; N represents the length of the sampling points of the input electromagnetic signal; and Fs represents the sampling rate of the input electromagnetic signal.

3. The method for comparing and learning intermediate frequency features based on multi-scale receptive field and context perception according to claim 2 is characterized in that: The method of performing signal preprocessing on the input electromagnetic signal according to the center frequency of the input electromagnetic signal to obtain intermediate frequency I / Q waveform data of the input electromagnetic signal includes: The input electromagnetic signal is down-converted according to its center frequency to obtain a complex signal X1(n): Wherein, X(n) represents the input electromagnetic signal, (n=1,…,N); j represents an imaginary unit; n represents the serial number of the current input electromagnetic signal; X1(n) is filtered and decimated by 4 times to obtain X2(m). in, Indicates rounding down.

4. The method for comparing and learning intermediate frequency features based on multi-scale receptive field and context perception according to claim 1, characterized in that: The step of performing complex number decomposition and maximum amplitude normalization on the intermediate frequency I / Q waveform data to form an intermediate frequency signal sample includes: Decompose the intermediate frequency I / Q waveform data into the real part I(t) and the imaginary part Q(t); The real part I(t) and the imaginary part Q(t) are normalized to their maximum magnitude: Among them, I(t) ′ represents the normalized result of the real part I(t); Q(t) ′ Represents the normalized result of the imaginary part Q(t); According to the normalization result, the intermediate frequency signal sample X = [I(t) ′ ,Q(t) ′ ].

5. The method for comparing and learning intermediate frequency features based on multi-scale receptive field and context perception according to claim 1, characterized in that: The step of extracting intermediate frequency high-order features based on multi-scale receptive field and context perception from the intermediate frequency signal sample to obtain a feature vector includes: A medium-frequency high-order feature extraction model is constructed, in which the dependency of the multi-scale receptive field of the input electromagnetic signal is captured by nesting the Inception module of convolution operations of different scales in the medium-frequency high-order feature extraction model; the model learns the importance of different features through a parallel branching structure, and extracts features at different abstraction levels; the medium-frequency high-order feature extraction model also adopts a bidirectional long short-term memory network module Bi-LSTM to capture contextual information in features of different scales and learn global serialization information; The intermediate frequency signal samples are input into the intermediate frequency high-order feature extraction model for spatial mapping to obtain a feature vector.

6. The method for comparing and learning intermediate frequency features based on multi-scale receptive fields and context perception according to claim 5, characterized in that: The intermediate frequency high-order feature extraction model includes: The multi-scale convolution module includes at least three convolution layers with different sizes of convolution kernels. Each convolution layer outputs features of different dimensions and cascades them in the channel dimension to obtain intermediate deep features. The dimension of the deep feature is the sum of the dimensions of all convolution layers. The batch normalization layer is used to accelerate the training of the multi-scale convolution module and prevent overfitting. The ReLU activation function is used after the BN layer: ReLU(X ′ )=max(0,X ′ ) Among them, X ′ It represents the intermediate depth features obtained after the medium frequency pulse sample X passes through the multi-size convolution layer and batch normalization layer; The convolution block is connected to the multi-scale convolution module to downsample the intermediate depth features to obtain the final depth features; The bidirectional LSTM module, connected to the convolutional block, captures the forward information from the beginning to the end of the sequence and the reverse information from the end to the beginning of the sequence based on the deep features: in, Represents the hidden state in the current sliding time window w obtained based on the forward information; Represents the hidden state in the current sliding time window w obtained based on reverse information; represents the hidden state in the past sliding time window w-1, Represents the hidden state of the future sliding time window w+1.

7. The method for comparing and learning intermediate frequency features based on multi-scale receptive fields and context perception according to claim 5, characterized in that: Also includes: Create a feature similarity discrimination task to determine whether the input signal data pair comes from the same radiation source target, and then use the contrast loss of "intra-class aggregation and inter-class separation" to constrain the learning of intermediate frequency high-order features, and use the cross-entropy loss of similarity confidence prediction to constrain and predict the learning of the intermediate frequency high-order feature extraction model.

8. The method for comparing and learning intermediate frequency features based on multi-scale receptive fields and context perception according to claim 7, characterized in that: In the feature similarity discrimination task, the intermediate frequency I / Q waveform data pair is input into two intermediate frequency high-order feature extraction models, and the two intermediate frequency high-order feature extraction models have the same structure and share weights.

9. The method for comparing and learning intermediate frequency features based on multi-scale receptive fields and context perception according to claim 8, characterized in that: The method of using the contrast loss of "intra-class aggregation and inter-class separation" to constrain the learning of intermediate-frequency high-order features and using the cross entropy loss of similarity confidence prediction to constrain and predict the learning of the intermediate-frequency high-order feature extraction model includes: According to the intermediate frequency I / Q waveform data set, several positive sample pairs (X1 + ,X2 + ) and negative sample pairs (X1 - ,X2 - ), wherein the two intermediate frequency I / Q waveform data in the positive sample pair belong to the same radiation source; and the two intermediate frequency I / Q waveform data in the negative sample pair do not belong to the same radiation source; The sample pairs are input into two intermediate frequency high-order feature extraction models; Construct the cross entropy loss function for similarity prediction of the intermediate frequency high-order feature extraction model: L CE =-[ylog(f(p + ))+(1-y)log(1-f(p - ))] p + =(X1 + ,X2 + ) p - =(X1 - ,X2 - ) Among them, L CE (·) represents the cross entropy loss; y represents the true label of the sample pair. If the data in the sample pair belongs to the same radiation source, y = 1, otherwise y = 0; f(·) represents the contrastive learning strategy based on the twin network; Construct the loss function of the intermediate frequency high-order feature contrast learning in the intermediate frequency high-order feature extraction model: L C =(1-y)·(s - ) 2 +y max(0.1-s + ) 2 s + =|H(f(p + ))| s - =|H(f(p - ))| Among them, L C represents the loss function for contrastive learning of intermediate frequency high-order features; s + Represents the similarity of positive sample pairs; represents; s - represents the similarity of negative sample pairs; H(·) represents the cosine similarity calculation function; Construct the overall loss function of the intermediate frequency high-order feature extraction model: L=αL CE +(1-α)L C Among them, L represents the overall loss function of the intermediate frequency high-order feature extraction model; α represents the weight adjustment coefficient between the two loss functions; By optimizing the overall loss function and adjusting the contrastive learning strategy, the distance between positive samples and medium-frequency high-order features is shortened, while the distance between negative samples and medium-frequency high-order features is increased.

10. The method for comparing and learning intermediate frequency features based on multi-scale receptive fields and context perception according to claim 1, characterized in that: The signal clustering of the feature vectors of the unknown signal according to the feature vectors of the known signals in the plurality of input electromagnetic signals comprises: Based on the chronological order, the intermediate frequency high-order features are traversed, and the absolute value of the correlation coefficient between the current feature vector and the central feature vector in the classified list is calculated each time as the feature similarity; If the calculated feature similarity is greater than the upper similarity threshold, it is added to the end of the category list with the highest similarity; If the calculated feature similarity is less than the similarity lower threshold, set it as a new category, create an empty list and add the feature; If the calculated feature similarity is between the upper and lower similarity thresholds, it is considered that the feature is not clearly distinguishable.

Images