Method for separating and identifying multiple electromagnetic interference sources

Abstract

The invention belongs to the technical field of interference identification, and particularly provides a method for separating and identifying multiple electromagnetic interference sources. According to the blind source separation method based on the convolutional neural network (CNN), separation of mixed signals with unknown source numbers is achieved, and the method does not need assumed conditions and is easy to achieve. The invention provides a deep feature representation learning network based on multi-feature representation learning and similarity measurement to learn deep feature representation of a broadband interference source. According to the method, the characteristic that the interference source is more distinctive can be obtained, so that the identification accuracy is improved.

Description

Technical Field

[0001] The invention belongs to the technical field of interference identification, and specifically proposes a method for separating and identifying multiple electromagnetic interference sources. Background Art

[0002] In recent years, radio technology has been widely used, and various types of radio equipment have been frequently used, making the electromagnetic environment more and more complex. Due to aging, temperature rise, improper use and other reasons, the radiation signal of radio equipment will show broadband characteristics. In a given electromagnetic environment, electromagnetic interference sources working at the same time may affect the normal operation of each other, cause electromagnetic disturbance, electromagnetic interference and even damage to other radio equipment to varying degrees, and may also have a certain degree of impact on the normal life and health of people around. Detecting electromagnetic interference sources from a given electromagnetic environment and identifying their properties are practical problems faced by electromagnetic environment monitoring systems. Since electromagnetic radiation signals do not have a specific modulation mode, the identification of electromagnetic interference sources still faces huge challenges: first, the radiation signal of broadband electromagnetic interference sources is random, and the difference between the spectrums of different interference sources is subtle, making it difficult to extract features; second, in a given electromagnetic environment, two or more electromagnetic interference sources of the same frequency may appear at the same time. At present, there is no quick and effective identification scheme for the identification of multiple broadband electromagnetic interference.

[0003] At present, there are few methods for identifying multiple electromagnetic interference sources in a given environment. The main method is to learn the feature representation of multiple broadband electromagnetic interference sources as a whole and match them with the feature representation of multiple broadband electromagnetic interference sources obtained in advance. This type of method requires first mixing multiple interference sources in pairs or multiples, collecting their mixed signals, and obtaining enough labeled data sets. The Chinese patent with publication number CN113569742 B provides "A method for identifying broadband electromagnetic interference sources based on convolutional neural networks", which mentions the use of convolutional neural networks to identify broadband electromagnetic interference sources around airports, but it cannot identify multiple electromagnetic interference sources at the same time.

[0004] Another idea for identifying multiple electromagnetic interference sources is to first separate multiple broadband electromagnetic interference sources, and then identify the separated single signal type. The performance of this type of method depends on the separation effect of the mixed signal, and it is necessary to consider the problem of low recognition accuracy when only a certain type of feature of a single broadband interference source is used. Therefore, we propose a deep neural network model based on blind source separation and multi-feature representation learning, by exploring the blind source separation method when the number of interference sources is uncertain, and representing the frequency spectrum features and spatial spectrum features of a single broadband interference source. Learning and training. The main contributions of the present invention are two: First, a blind source separation method based on convolutional neural network (CNN) is proposed to achieve mixed signal separation of unknown sources. This method does not require assumptions and is easy to implement; second, a deep feature representation learning network based on multi-feature representation learning and similarity measurement is proposed to learn the deep feature representation of broadband interference sources. This method can obtain more discriminative features of interference sources to improve recognition accuracy. Summary of the invention

[0005] The purpose of the present invention is to provide a method for identifying multiple electromagnetic interference sources based on blind source separation and multi-feature representation based on deep learning, which can realize the separation of mixed signals of unknown sources without any assumptions and can obtain more distinctive features of the interference sources to improve the recognition accuracy.

[0006] The technical solution of the present invention is:

[0007] A method for separating and identifying multiple electromagnetic interference sources comprises the following steps:

[0008] S1. Build training data:

[0009] The array sampling system is used to collect the source signals S(t) of various known space electromagnetic interference sources to obtain the observed signal X(t); the observed signal X(t) is discretely sampled to obtain the covariance matrix when the number of snapshots is L, and the element in the upper right corner of the covariance matrix is ​​used to obtain the vector r, which is an M(M+1) / 2-dimensional complex matrix. Finally, the real and imaginary parts of r are separated to obtain the vectors r R and r I As sample data;

[0010] S2, constructing a separation and recognition model, including a mapping module, a fusion module, a DOA estimation module, a source signal separation module, a feature extraction module, and a recognition module;

[0011] The mapping module includes two parallel channels, one for r R and r I Processing, r R and r I The spatial probability mapped to the output end is output respectively and point;

[0012] The fusion module and Fusion is performed. The fusion method judges the two outputs. The two values ​​close to 0 take the smaller value, and the two values ​​close to 1 take the larger value, so as to obtain the pseudo-space spectrum, which is defined as

[0013] The DOA estimation module is used to estimate the DOA of the electromagnetic interference source, the source number of the electromagnetic interference source and obtain a mixing matrix, wherein the DOA of the electromagnetic interference source is obtained by calculating the index of the maximum value of the pseudo-space spectrum; the source number of the electromagnetic interference source is the number of set elements, defined as P; the mixing matrix is ​​defined as a p (θ p ) is the direction vector of the pth source signal on the uniform circular array;

[0014] The source signal separation module is used to estimate the separation matrix according to the mixing matrix, and the relationship between the separation matrix B and the mixing matrix is ​​defined as follows: I P×P is the unit matrix. After obtaining the separation matrix B, Restore the source signal;

[0015] The input of the feature extraction module is a positive and negative sample pair, which is obtained by combining the separated signals and calculating the spectrum and spatial spectrum of each separated signal. The output vector of the feature extraction module is defined as and

[0016] The recognition module is based on and Perform target identification;

[0017] S3. Use the training data to train the constructed separation and recognition model to obtain a trained separation and recognition model, wherein the fusion module is optimized using binary cross entropy loss, and the loss function constructed for the source signal separation module is expressed as:

[0018]

[0019] Among them, b p is the p-th row vector of the separation matrix B, a p is the mixing matrix estimate The pth column vector of p Under known circumstances, by continuously optimizing b p Make a p b p is 1, so we get b p;

[0020] The feature extraction module is optimized using binary cross entropy loss;

[0021] The recognition module is optimized using binary cross entropy loss and contrast loss functions;

[0022] S4. Use the trained separation and recognition model to identify the electromagnetic interference source. During the recognition process, use To represent the deep feature representation of the trained electromagnetic interference source, To represent the deep feature representation of the electromagnetic interference source under test, define the threshold η. If the output is 0 or close to 0, it means and The electromagnetic interference sources represented are of the same type. If the output is close to η, then The EMI source represented is an untrained EMI source, i.e., an unknown EMI source type.

[0023] Furthermore, in S1, the observed signal X(t) is defined as:

[0024]

[0025] in, It is an M×P dimensional matrix composed of directional vectors, called the directional vector matrix, that is, the mixing matrix, and N(t) is the noise signal matrix.

[0026] Covariance matrix R XX for:

[0027]

[0028] Then r is:

[0029] r=[r 11 ,…,r 1M ,r 22 ,…,r 2M ,…,r MM ].

[0030] Furthermore, in S2, the two channels of the mapping module are CNN-based subnetworks, which include 5 convolutional layers with RELU, 1 flattening layer, 1 Dropout layer, 1 BN layer and 1 FC layer, and each convolutional layer is followed by a maximum pooling layer.

[0031] Furthermore, in S2, the i-th output after fusion of the fusion module is:

[0032]

[0033] in, and are the i-th output values ​​of the two CNN sub-networks, o i The output label o = [o0, o1, ..., o i ,…,o 360 ], and o i ∈[0,1].

[0034] Furthermore, in S3, the specific identification process of the identification module is as follows:

[0035] First, the frequency spectrum and spatial spectrum of each sample are calculated, which are expressed as and Sample pair x i and x j The frequency spectrum and spatial spectrum of and

[0036] Secondly, the sample pair (x i ,x j ) is fused with the spatial spectrum to obtain the deep features. express;

[0037] Then, As input, As output. The corresponding output label is o = [o i1 ,o i2 ,…,o iK ,y,o j1 ,o j2 ,…,o jK ], each element in the vector o belongs to {0,1}. K represents the number of types of electromagnetic interference sources. W1 and W2 represent weights, b1 and b2 represent biases. In order to reduce the training parameters of the network, let W1=W2, b1=b2. Let but

[0038]

[0039] Similarly, we can calculate In is the similarity between two multi-feature representations, expressed as Euclidean distance

[0040]

[0041] Finally, the loss function is used for classification. The loss function is

[0042]

[0043] Using contrast loss function to achieve and To judge whether they belong to the same type, the contrast loss function is expressed as

[0044]

[0045] Where Q q is the sample logarithm, is the similarity of the qth sample pair, and η is the threshold.

[0046] Beneficial effects of the present invention:

[0047] (1) A blind source separation method based on convolutional neural network (CNN) is proposed to separate mixed signals with unknown number of sources. This method does not require any assumptions and is easy to implement.

[0048] (2) A deep feature representation learning network based on multi-feature representation learning and similarity measurement is proposed to learn the deep feature representation of broadband interference sources. This method can obtain more discriminative features of interference sources to improve recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is the overall flow chart of the present invention.

[0050] Figure 2 Schematic diagram of the multi-feature representation process of the separated signal.

[0051] Figure 3 Schematic diagram of signal identification after separation. DETAILED DESCRIPTION

[0052] The present invention is described in detail below with reference to the accompanying drawings.

[0053] like Figure 1 As shown, the present invention mainly comprises the following steps:

[0054] The array sampling system is used to collect the source signal S(t) of the space electromagnetic interference source to obtain the observation signal X(t).

[0055] X(t)=AS(t)+N(t)

[0056] When the observed signal X(t) is discretely sampled and the number of snapshots is L, the expression of its covariance matrix is:

[0057]

[0058] Because the covariance matrix R XX is a symmetric matrix, whose upper right corner elements can be combined into a vector r = [r 11 ,…,r 1M ,r 22 ,…,r 2M,…,r MM ] is a complex matrix of J = M(M+1)k2 dimensions. Currently, there is no mature deep neural network that can be used to process complex numbers. The present invention separates the real part and the imaginary part of r into two vectors, respectively using r R and r I To express.

[0059] The real vector r of the covariance matrix is ​​transformed through two CNN-based sub-networks R and the imaginary vector r I The CNN subnetwork consists of 5 convolutional layers with RELU, 1 flattening layer, 1 dropout layer, 1 batch normalization layer, and 1 FC layer. Each convolutional layer is followed by a maximum pooling layer.

[0060] The outputs of the two sub-networks are fused. First, the two values ​​close to 0 in the outputs of the corresponding positions of the two sub-networks are taken as the smaller value, and the two values ​​close to 1 are taken as the larger value to obtain the "needle-shaped" pseudo-space spectrum. In other words, the i-th output after fusion is

[0061]

[0062] In the formula and are the i-th output values ​​of the two sub-networks respectively. Therefore, the fused output vector can be expressed as

[0063]

[0064] Then, in order to make the fused output vector "needle-shaped", the estimation of the pseudo-spatial spectrum is formulated as a multi-label multi-classification problem, and the binary cross entropy loss is used for optimization, that is,

[0065]

[0066] Where Q is the number of samples. is the i-th output label of the q-th sample, and is the i-th network output value of the q-th sample and That is, the estimated value of the pseudo-space spectrum of the qth sample in direction i. Obviously, if the direction of a source signal is i, its corresponding output label o i is 1, and the Adam optimization algorithm makes L PSSENet When the minimum value is reached, the estimated value of the pseudo-spatial spectrum output after fusion is is 1, and the estimated value of the pseudo-space spectrum of the electromagnetic interference source in other directions is 0. The pseudo-space spectrum estimation method proposed in the present invention does not need to assume the relationship between the number of sources and the number of array elements, and is not constrained by positive, overdetermined and underdetermined conditions.

[0067] Calculate the number of sources, angle of arrival (DoA) and mixing matrix of the electromagnetic interference source. In theory, the DoA of the electromagnetic interference source is the index corresponding to the value of 1 in the pseudo-space spectrum vector. However, the loss function of the network is sometimes difficult to converge completely to 0, and may be a very small value close to 0. In other words, It is also difficult to reach 1, and is a value close to 1. Therefore, the DoA of the electromagnetic interference source can be estimated by calculating the index of the maximum value of the pseudo-space spectrum, which can be mathematically expressed as

[0068]

[0069] The number of electromagnetic interference sources is the number of elements in the set. The frequency of the electromagnetic interference source can be obtained by Fourier transforming the observed signal.

[0070] Calculate the mixing matrix according to the existing formula a p (θ p ) is the direction vector of the p-th source signal on the UCA, expressed as

[0071] a p (θ p )=[exp(j2πf p τ1(θ p )),exp(j2πf p τ2(θ p )),…,exp(j2πf p τ m (θ p ))] T

[0072] Where f p is the frequency of the pth source signal. If the source signal is a broadband signal, f p It represents the frequency of the interference frequency. Assuming that a uniform circular array (UCA) antenna is used to receive signals from multiple radiation sources, τ m (θ p ) is the time delay between the mth array element and the reference point when receiving the pth source signal. When the center of the circle is taken as the reference point and the radius of the circular array is R, τ m (θ p ) can be expressed as

[0073]

[0074] Where C is a constant representing the speed of light.

[0075] Estimate the separation matrix and restore the source signal. Accurately restore the source signal needs to meet Assumption b pis the p-th row vector of the separation matrix B, a p is the mixing matrix estimate The present invention constructs a new loss function as the criterion for optimizing separation matrix estimation (SME), namely

[0076]

[0077] Where P is the number of electromagnetic interference sources in the environment, L SME It is called the loss function of the separation matrix estimation. p Under known circumstances, by continuously optimizing b p Make a p b p is 1, so we get b p The optimization process is expressed as

[0078]

[0079] Where h is the number of iterations, κ b It is called the learning rate. YesL SME About b p The partial derivative of can be expressed as

[0080]

[0081] Where a p T is a p The transpose of . Each row vector in the separation matrix B can be obtained through multiple iterations.

[0082] According to the formula Recover the source signal, that is, the source signal separated from the observed signal.

[0083] like Figure 2 As shown, the separated signals are combined in pairs to form positive sample pairs and negative sample pairs, and the frequency spectrum and spatial spectrum of each separated signal are calculated respectively.

[0084] The spectrum and spatial spectrum of the separated signal are input into the CNN network to obtain the multi-feature representation of each separated signal. The output vector is and

[0085] The output labels of the same electromagnetic interference source from different CNN networks should be the same, so the output vector is selected One of the values ​​with the same index in is output. That is, the output vector The kth value in is

[0086]

[0087] When the label y k When it is 0, select The minimum value in When k When it is 1, select The maximum value in The selected values ​​constitute the output vector The learning of multi-feature representation is regarded as a multi-label multi-classification problem, represented by binary cross entropy loss, that is,

[0088]

[0089] In the formula, Q represents the number of samples, and K is also the number of types. is the output label of the qth sample. is the network output value of the qth sample. Through the Adam optimization algorithm, L MFRL When the minimum value 0 is reached, the estimated value of the output corresponding to label 1 is 1, and the estimated value corresponding to label 0 is 0. After multiple forward and backward propagations until L MFRL The training is completed when convergence occurs. By optimizing the loss function, the selected minimum value is close to 0 and the maximum value is close to 1, thereby obtaining a deeper multi-feature representation of the electromagnetic interference source. The multi-feature representation proposed in the present invention does not generate new training parameters and does not increase the training time of the network.

[0090] Finally, a multi-classification layer based on similarity measurement is used to classify the sample pairs (x i ,x j ) identification. Figure 3 As shown, the input of the multi-classification layer is the multi-feature representation of the sample pair The output is the probability distribution of each type represented by multiple features, and y i With y j The vector composed of the similarity of can be expressed as vector K represents the number of types of electromagnetic interference sources. W1 and W2 are weights, b1 and b2 are biases, and the upper and lower parts of the network have the same functions. In order to reduce the training parameters of the network, let W1 = W2, b1 = b2. Vector o = [o i1 ,o i2 ,…,o iK ,y,o j1 ,o j2 ,…,o jK ] is the output label of the network, and each element in the vector o belongs to {0,1}.

[0091] make but After being mapped by the Softmax function, it is expressed as:

[0092]

[0093] In the formula, k = 1, 2, ..., K. Similarly, we can calculate In is the similarity between two multi-feature representations, expressed as Euclidean distance

[0094]

[0095] In order to identify and The types of electromagnetic interference sources represented by them are respectively, and the loss function for achieving classification is obtained according to the binary cross entropy loss function:

[0096]

[0097] Use cross entropy to evaluate the difference between the actual output and the expected output label. If the difference is large, continue to adjust the network parameters until the difference between the two is 0 or close to 0. It should be noted that and The corresponding output labels are the same, i.e. o ik =o jk In order to judge and Whether they belong to the same type, the contrast loss function of the multi-classification layer to realize the identification of unknown interference sources can be expressed as

[0098]

[0099] Where Q q is the sample logarithm, is the similarity of the qth sample pair, η is the threshold. The total loss function L Unknown It can be defined as L Unknown =L Class +L Class-SM . Use L Class-SM Realize the measurement of the similarity of two electromagnetic interference sources, using L Class To identify the specific type of electromagnetic interference source. Unknown Regarding the gradient of each parameter, and continuously updating each parameter through the Adam optimization algorithm, L Unknown Minimize, so that the proposed classification layer has the ability to classify the trained electromagnetic interference source types and the ability to identify unknown electromagnetic interference sources. To represent the deep feature representation of the trained electromagnetic interference source, To represent the deep feature representation of the electromagnetic interference source being tested. If the output of the network 0 or close to 0, indicating and The electromagnetic interference sources represented belong to the same type, and the specific type is determined by the corresponding output vector The value in is used to determine. The value is large and close to η, indicating The represented electromagnetic interference source is an untrained electromagnetic interference source, that is, an unknown electromagnetic interference source type. Therefore, after training, the classification layer of the present invention can identify known electromagnetic interference sources and unknown electromagnetic interference sources.

Claims

1. A method for separating and identifying multiple electromagnetic interference sources, characterized in that: The following steps are involved: S1. Build training data: The array sampling system is used to collect source signals S(t) of various known space electromagnetic interference sources to obtain observation signals X(t); Discretely sample the observed signal X(t) to obtain the covariance matrix when the number of snapshots is L. Use the element in the upper right corner of the covariance matrix to obtain the vector r, which is an M(M+1) / 2-dimensional complex matrix. Finally, separate the real and imaginary parts of r to obtain the vectors r R and r I As sample data; S2, constructing a separation and recognition model, including a mapping module, a fusion module, a DOA estimation module, a source signal separation module, a feature extraction module, and a recognition module; The mapping module includes two parallel channels, one for r R and r I Processing, r R and r I The spatial probability mapped to the output end is output respectively and point; The fusion module and Fusion is performed. The fusion method judges the two outputs. The two values ​​close to 0 take the smaller value, and the two values ​​close to 1 take the larger value, so as to obtain the pseudo-space spectrum, which is defined as The DOA estimation module is used to estimate the DOA of the electromagnetic interference source, the source number of the electromagnetic interference source and obtain a mixing matrix, wherein the DOA of the electromagnetic interference source is obtained by calculating the index of the maximum value of the pseudo-space spectrum; the source number of the electromagnetic interference source is the number of set elements, defined as P; the mixing matrix is ​​defined as a p (θ p ) is the direction vector of the pth source signal on the uniform circular array; The source signal separation module is used to estimate the separation matrix according to the mixing matrix, and the relationship between the separation matrix B and the mixing matrix is ​​defined as follows: I P×P is the unit matrix. After obtaining the separation matrix B, Restore the source signal; The input of the feature extraction module is a positive and negative sample pair, which is obtained by combining the separated signals and calculating the spectrum and spatial spectrum of each separated signal. The output vector of the feature extraction module is defined as and The recognition module is based on and Perform target identification; S3. Use the training data to train the constructed separation and recognition model to obtain a trained separation and recognition model, wherein the fusion module is optimized using binary cross entropy loss, and the loss function constructed for the source signal separation module is expressed as: Among them, b p is the p-th row vector of the separation matrix B, a p is the mixing matrix estimate The pth column vector of p Under known circumstances, by continuously optimizing b p Make a p b p is 1, so we get b p ; The feature extraction module is optimized using binary cross entropy loss; The recognition module is optimized using binary cross entropy loss and contrast loss functions; S4. Use the trained separation and recognition model to identify the electromagnetic interference source. During the recognition process, use To represent the deep feature representation of the trained electromagnetic interference source, To represent the deep feature representation of the electromagnetic interference source under test, define the threshold η. If the output is 0 or close to 0, it means and The electromagnetic interference sources represented are of the same type. If the output is close to η, then The EMI source represented is an untrained EMI source, i.e., an unknown EMI source type.

2. The method for separating and identifying multiple electromagnetic interference sources according to claim 1, characterized in that: In S1, the observed signal X(t) is defined as: in, It is an M×P dimensional matrix composed of directional vectors, called the directional vector matrix, that is, the mixing matrix, and N(t) is the noise signal matrix; Covariance matrix R XX for: Then r is: r＝[r 11 ,…,r 1M ,r 22 ,…,r 2M ,…,r MM ]。 3. The method for separating and identifying multiple electromagnetic interference sources according to claim 2, characterized in that: In S2, the two channels of the mapping module are CNN-based subnetworks. The CNN subnetwork includes 5 convolutional layers with RELU, 1 flattening layer, 1 Dropout layer, 1 BN layer and 1 FC layer. There is a maximum pooling layer after each convolutional layer.

4. The method for separating and identifying multiple electromagnetic interference sources according to claim 3, characterized in that: In S2, the i-th output after fusion of the fusion module is: in, and are the i-th output values ​​of the two CNN sub-networks, o i The output label o = [o0, o1, ..., o i ,…,o 360 ], and o i ∈[0,1].

5. The method for separating and identifying multiple electromagnetic interference sources according to claim 3, characterized in that: In S3, the specific recognition process of the recognition module is as follows: First, the frequency spectrum and spatial spectrum of each sample are calculated, which are expressed as and Sample pair x i and x j The frequency spectrum and spatial spectrum of and Secondly, the sample pair (x i ,x j ) is fused with the spatial spectrum to obtain the deep features. express; Then, As input, As output, the corresponding output label is o = [o i1 ,o i2 ,…,o iK ,y,o j1 ,o j2 ,…,o jK ], each element in the vector o belongs to {0,1}, K represents the number of types of electromagnetic interference sources; W1 and W2 represent weights, b1 and b2 represent biases, in order to reduce the training parameters of the network, let W1=W2, b1=b2; let but Similarly, we can calculate In is the similarity between two multi-feature representations, expressed as Euclidean distance Finally, the loss function is used for classification. The loss function is Using contrast loss function to achieve and Whether they belong to the same type, the contrast loss function is expressed as Where Q q is the sample logarithm, is the similarity of the qth sample pair, and η is the threshold.

Images