Multi-radar emitter signal recognition method based on improved VMD-RNN

Abstract

The application discloses a kind of multi-radar radiation source signal identification method based on improved VMD-RNN, belong to radar radiation source identification technical field.The application includes the following steps: composite signal variation modal decomposition;Signal is decomposed into several components by improved VMD method;Time-frequency processing of component signal;Residual neural network is constructed.The center frequency and instantaneous amplitude, instantaneous phase, instantaneous frequency of component are replaced by two-dimensional time-frequency diagram as the input of deep learning, reduce the dimension of data processing, while providing more comprehensive time domain, frequency domain, time-frequency domain feature information, indirectly reduce the number of required samples, improve efficiency;The application decomposes signal into several components with certain bandwidth and center frequency by VMD, respectively carries out deep learning, obtains the deep feature of each component, and the mutual influence between multiple radar radiation source signals is reduced to the minimum.

Description

Technical Field

[0001] This invention belongs to the field of radar radiation source identification technology, specifically relating to a method for identifying multiple radar radiation source signals based on an improved VMD-RNN. Background Technology

[0002] The requirements for radar radiation source identification have evolved from the early identification of a limited number of simple radar signals to the efficient and accurate identification of multiple targets under complex electromagnetic environments, which has greatly increased the difficulty of radar radiation source identification.

[0003] Deep learning models have strong learning capabilities. In recent years, researchers have applied deep learning to radar radiation source identification. By constructing deep neural networks, complex low-level signal features can be transmitted through the neural network to achieve nonlinear function approximation. The computer automatically learns and obtains high-level feature quantities from the data without the need for manual design of feature parameters. It has strong generalization ability and can extract information features that characterize the essence of the data, thereby completing the classification and identification of radar radiation sources.

[0004] Current research focuses on two main directions. One is to extract features from radiation source signals using supervised learning for classification and identification, with CNNs being a typical deep learning model. Leveraging the powerful performance of CNNs in computer vision tasks, the radar radiation source signals are transformed to extract their two-dimensional image features (time-frequency distribution features are commonly used) as samples to train the CNN model, which is then used for classification and identification. The second direction involves extracting features from radiation source signals using unsupervised learning, training the model, then using the backpropagation (BP) algorithm for supervised fine-tuning of the model parameters, and finally using a classifier for identification. Typical deep learning models in this direction are DBN and SAE.

[0005] Disadvantages of existing technology:

[0006] 1. Deep learning requires a large number of samples. 2. Interference between signals from multiple radar sources makes it impossible for a single deep learning method to achieve a high recognition probability. Summary of the Invention

[0007] To address the aforementioned technical problems in existing technologies, this invention proposes a multi-radar radiation source signal identification method based on an improved VMD-RNN (Variational Mode Decomposition-Residual Neural Network). The method is rationally designed, overcomes the shortcomings of existing technologies, and has good performance.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] The multi-radar radiation source signal identification method based on the improved VMD-RNN includes the following steps:

[0010] Step 1: Variational Mode Decomposition of Composite Signal; The signal is decomposed into several components using an improved VMD method;

[0011] Step 2: Time-frequency processing of component signals;

[0012] Step 3: Residual neural network classification and recognition.

[0013] Preferably, step 1 specifically includes the following steps:

[0014] Step 1.1: Convert the original radar signal x k (t) is decomposed into I center frequencies {ω} i The signal components {u} i The sum of the estimated bandwidths of each signal component is minimized by the trend term, as shown in formula (1):

[0015] (1);

[0016] In the formula, {u i}={u1, u2, …, u I} represents I signal components; {ω i}={ω1,ω2, …,ω I} represents the k-th signal

[0017] The center frequency of the component; This represents the partial derivative of a function with respect to time; x k (t) represents the original radar signal; t represents time.

[0018] Constructing the augmented Lagrange function by introducing the Lagrange multiplication operator and the quadratic penalty parameter: (2);

[0019] In the formula, λ is the Lagrange multiplier operator, and α is the quadratic penalty parameter;

[0020] In the alternating direction multiplier algorithm, the solution to equation (1) is found as the saddle point of the augmented Lagrange; u i n+1 ω i n+1 and λ n+1 The iterative formulas are shown in (3)-(6):

[0021] (3);

[0022] (4);

[0023] (5);

[0024] (6);

[0025] In the formula, τ is the noise tolerance parameter; u i n+1 ω represents the i-th signal component in iteration (n+1); i n+1 λ represents the center frequency of the i-th signal component in iteration (n+1); n+1 Let represent the Lagrange multiplication operator iterated (n+1) times; stop iterating after the discrimination accuracy is satisfied, and obtain the corresponding i signal components, as shown in formula (7):

[0026] (7);

[0027] In the formula, ε is the discrimination precision, u i n To iterate n times to decompose the i-th single-component amplitude-frequency modulated signal;

[0028] According to formula (8), calculate the I component signals and the original signal x. k The correlation coefficient δ between (t) and ki ;

[0029] (8);

[0030] Based on the principle of removing unrelated modes, the noise component and trend term are obtained, denoted as R(t);

[0031] Step 1.2: Determine max{δ} k1 ,…,δ kI}<min{δ (k-1)1 ,…,δ (k-1)I Does this hold true?

[0032] If the result of the judgment is max{δ k1 ,…,δ kI}<min{δ (k-1)1 ,…,δ (k-1)I If the condition is met, the process ends.

[0033] Or the result of the judgment is max{δ k1 ,…,δ kI}<min{δ (k-1)1 ,…,δ (k-1)I If the condition is not met, proceed to step 1.3;

[0034] Step 1.3: Select the component u with the highest correlation coefficient. ki As the k-th component, the original signal is subtracted from this component and R(t) to obtain the new original signal, and the variational mode decomposition process described above is repeated.

[0035] Preferably, step 2 specifically includes the following steps:

[0036] Step 2.1: Determine each component signal u k The local maxima and local minima are found, where k=1,…,N. Cubic spline interpolation is used to form the upper and lower envelopes of the data, and the average of the upper and lower envelopes is defined as u. k The envelope, u kFM =u k -u kAM Repeat the envelope solution until... Each component signal is decomposed into a frequency modulation part u. kFM and amplitude modulation part u kAM ;

[0037] Step 2.2: For u respectively kFM and u kAM The Hilbert transform is expressed as follows:

[0038] (9);

[0039] The instantaneous phase θ(t) is defined as follows:

[0040] (10);

[0041] instantaneous frequency The definition is as follows:

[0042] (11);

[0043] u is obtained through formulas (9) and (10) kFM and u kAM The instantaneous phase and instantaneous frequency of the signal.

[0044] Preferably, step 3 specifically includes the following steps:

[0045] Step 3.1: Construct a classification model based on a ResNet34 residual neural network, with the input being a set of intrinsic mode functions u obtained through variational mode decomposition. i Frequency modulation u obtained by envelope demodulation iFM and amplitude modulation u iAM The component and its corresponding instantaneous frequency, instantaneous phase, and center frequency ω i Normalized multiple 7×7 matrices of 7 time series combined together;

[0046] Step 3.2: For each component u k Training and classification recognition are performed.

[0047] Step 3.3: Present the recognition results.

[0048] The beneficial technical effects of this invention are as follows:

[0049] 1. This invention replaces the two-dimensional time-frequency diagram with the center frequency, instantaneous amplitude, instantaneous phase, and instantaneous frequency of the component as input for deep learning, reducing the data processing dimension and providing more comprehensive time-domain, frequency-domain, and time-frequency-domain feature information of the signal, indirectly reducing the number of samples required and improving efficiency.

[0050] 2. This invention divides the signal into several components with a certain bandwidth and center frequency through VMD, performs deep learning on each component to obtain the deep features of each component, and minimizes the mutual influence between multiple radar radiation source signals. Attached Figure Description

[0051] Figure 1 This is a flowchart of the method of the present invention;

[0052] Figure 2 This is a diagram showing the overall structure of the RNN network model. Detailed Implementation

[0053] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0054] This invention presents a multi-radar radiation source signal identification algorithm based on an improved VMD-RNN. By modifying the VMD method, the signal is decomposed into several components with specific center frequencies and bandwidths, achieving the separation of multiple radar radiation source signals. Each component is then input into a deep learning neural network, and radar radiation source signal identification is achieved through the mining of deep-level features. The algorithm flow described in this invention is as follows: Figure 1 As shown.

[0055] Step 1: Variational Mode Decomposition of Composite Signal; The signal is decomposed into several components using an improved VMD method; Specifically, it includes the following steps:

[0056] Step 1.1: Convert the original radar signal x k (t) is decomposed into I center frequencies {ω} i The signal components {u} i The sum of the estimated bandwidths of each signal component is minimized by the trend term, as shown in formula (1):

[0057] (1);

[0058] In the formula, {u i}={u1, u2, …, u I} represents I signal components; {ω i}={ω1,ω2, …,ω I} represents the k-th signal

[0059] The center frequency of the component; This represents the partial derivative of a function with respect to time; x k (t) represents the original radar signal; t represents time.

[0060] Constructing the augmented Lagrange function by introducing the Lagrange multiplication operator and the quadratic penalty parameter:

[0061] (2);

[0062] In the formula, λ is the Lagrange multiplier operator, and α is the quadratic penalty parameter;

[0063] In the alternating direction multiplier algorithm, the solution to equation (1) is found as the saddle point of the augmented Lagrange; u i n+1 ω i n+1 and λ n+1 The iterative formulas are shown in (3)-(6):

[0064] (3);

[0065] (4);

[0066] (5);

[0067] (6);

[0068] In the formula, τ is the noise tolerance parameter; u i n+1 ω represents the i-th signal component in iteration (n+1); i n+1 λ represents the center frequency of the i-th signal component in iteration (n+1); n+1 Let represent the Lagrange multiplication operator iterated (n+1) times; stop iterating after the discrimination accuracy is satisfied, and obtain the corresponding i signal components, as shown in formula (7):

[0069] (7);

[0070] In the formula, ε is the discrimination precision, u i nTo iterate n times to decompose the i-th single-component amplitude-frequency modulated signal;

[0071] According to formula (8), calculate the I component signals and the original signal x. k The correlation coefficient δ between (t) and ki ;

[0072] (8);

[0073] Based on the principle of removing unrelated modes, the noise component and trend term are obtained, denoted as R(t);

[0074] Step 1.2: Determine max{δ} k1 ,…,δ kI}<min{δ (k-1)1 ,…,δ (k-1)I Does this hold true?

[0075] If the result of the judgment is max{δ k1 ,…,δ kI}<min{δ (k-1)1 ,…,δ (k-1)I If the condition is met, the process ends.

[0076] Or the result of the judgment is max{δ k1 ,…,δ kI}<min{δ (k-1)1 ,…,δ (k-1)I If the condition is not met, proceed to step 1.3;

[0077] Step 1.3: Select the component u with the highest correlation coefficient. ki As the k-th component, the original signal is subtracted from this component and R(t) to obtain the new original signal, and the variational mode decomposition process described above is repeated.

[0078] Step 2: Time-frequency processing of component signals; specifically including the following steps:

[0079] Step 2.1: Determine each component signal u k The local maxima and local minima are found, where k=1,…,N. Cubic spline interpolation is used to form the upper and lower envelopes of the data, and the average of the upper and lower envelopes is defined as u. k The envelope, u kFM =u k -u kAM Repeat the envelope solution until... Each component signal is decomposed into a frequency modulation part u. kFM and amplitude modulation part u kAM ;

[0080] Step 2.2: For u respectively kFMand u kAM The Hilbert transform is expressed as follows:

[0081] (9);

[0082] The instantaneous phase θ(t) is defined as follows:

[0083] (10);

[0084] instantaneous frequency The definition is as follows:

[0085] (11);

[0086] u is obtained through formulas (9) and (10) kFM and u kAM The instantaneous phase and instantaneous frequency of the signal.

[0087] Step 3: Residual neural network classification and recognition;

[0088] Construct a classification model based on a ResNet34 residual neural network, with the input being a set of intrinsic mode functions u obtained through variational mode decomposition. i Frequency modulation u obtained by envelope demodulation iFM and amplitude modulation u iAM The component and its corresponding instantaneous frequency, instantaneous phase, and center frequency ω i The seven time series are combined into a normalized 7×7 matrix; then, for each component u... k The process involves training and classification; finally, the recognition results are given, which include the number of signal components and their corresponding center frequency and modulation type.

[0089] Key points of this invention:

[0090] The signal is decomposed into several components with a certain center frequency and bandwidth, and deep learning is performed on each component to reduce mutual interference between multiple radar radiation source signals and improve the recognition probability.

[0091] Each VMD component is transformed into a frequency-modulated and amplitude-modulated signal through its envelope. Its center frequency, instantaneous amplitude, instantaneous phase, and instantaneous frequency are used as input matrices to replace the time-frequency plot of the VMD component. This reduces the computational cost of deep learning while preserving the important time-frequency characteristics of the signal.

[0092] Protection points of this invention:

[0093] Instead of a time-frequency graph, an input matrix consisting of center frequency, instantaneous amplitude, instantaneous phase, and instantaneous frequency is constructed to reduce the computational cost of deep learning.

[0094] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A method for identifying multiple radar radiation source signals based on an improved VMD-RNN, characterized in that: Includes the following steps: Step 1: Variational Mode Decomposition of Composite Signal; The signal is decomposed into several components using an improved variational mode decomposition method; Step 2: Time-frequency processing of component signals; Step 3: Residual neural network classification and recognition; Step 1 specifically includes the following steps: Step 1.1: Convert the original radar signal x k (t) is decomposed into I center frequencies {ω} i The signal components {u} i The sum of the estimated bandwidths of each signal component is minimized by the trend term, as shown in formula (1): (1)； In the formula, {u i }={u1, u2, …, u I } represents I signal components; {ω i }={ω1,ω2, …,ω I } represents the k-th signal The center frequency of the component; This represents the partial derivative of a function with respect to time; x k (t) represents the original radar signal; t represents time. Constructing the augmented Lagrange function by introducing the Lagrange multiplication operator and the quadratic penalty parameter: （2）； In the formula, λ is the Lagrange multiplier operator, and α is the quadratic penalty parameter; In the alternating direction multiplier algorithm, the solution to equation (1) is found as the saddle point of the augmented Lagrange; u i n+1 ω i n+1 and λ n+1 The iterative formulas are shown in (3)-(6): （3）； （4）； （5）； （6）； In the formula, τ is the noise tolerance parameter; u i n+1 Represents the i-th signal component in iteration (n+1); ω i n+1 λ represents the center frequency of the i-th signal component in iteration (n+1); n+1 Let represent the Lagrange multiplication operator iterated (n+1) times; stop iterating after the discrimination accuracy is satisfied, and obtain the corresponding i signal components, as shown in formula (7): （7）； In the formula, ε is the discrimination precision, u i n To iterate n times to decompose the i-th single-component amplitude-frequency modulated signal; According to formula (8), calculate the I component signals and the original signal x. k The correlation coefficient δ between (t) and ki ; （8）； Based on the principle of removing unrelated modes, the noise component and trend term are obtained, denoted as R(t); Step 1.2: Determine max{δ} k1 ,…,δ kI }<min{δ (k-1)1 ,…,δ (k-1)I Does this hold true? If the result of the judgment is max{δ k1 ,…,δ kI }<min{δ (k-1)1 ,…,δ (k-1)I If the condition is met, the process ends. Or the result of the judgment is max{δ k1 ,…,δ kI }<min{δ (k-1)1 ,…,δ (k-1)I If the condition is not met, proceed to step 1.3; Step 1.3: Select the component u with the highest correlation coefficient. ki As the kth component, the original signal is subtracted from this component and R(t) to obtain the new original signal, and the above variational mode decomposition process is repeated; Step 3 specifically includes the following steps: Step 3.1: Construct a classification model based on a ResNet34 residual neural network, with the input being a set of intrinsic mode functions u obtained through variational mode decomposition. i Frequency modulation u obtained by envelope demodulation iFM and amplitude modulation u iAM The component and its corresponding instantaneous frequency, instantaneous phase, and center frequency ω i Normalized multiple 7×7 matrices of 7 time series combined together; Step 3.2: For each component u k Training and classification recognition are performed. Step 3.3: Present the recognition results.

2. The method for identifying multiple radar radiation source signals based on improved VMD-RNN according to claim 1, characterized in that: Step 2 specifically includes the following steps: Step 2.1: Determine each component signal u k The local maxima and local minima are found, where k=1,…,N. Cubic spline interpolation is used to form the upper and lower envelopes of the data, and the average of the upper and lower envelopes is defined as u. k The envelope, u kFM =u k -u kAM Repeat the envelope solution until... Each component signal is decomposed into a frequency modulation part u. kFM and amplitude modulation part u kAM ; Step 2.2: For u respectively kFM and u kAM The Hilbert transform is expressed as follows: (9)； The instantaneous phase θ(t) is defined as follows: (10)； instantaneous frequency The definition is as follows: (11)； u is obtained through formulas (9) and (10) kFM and u kAM The instantaneous phase and instantaneous frequency of the signal.

Images