A fast signal recognition method based on spectrum template
Through the rapid signal recognition method based on spectrum templates, the Gaussian process regression and parallel computing architecture is used to solve the problem of slow signal recognition speed in complex electromagnetic environments, and the ability to quickly identify signals without IQ data is achieved, which satisfies the Real-time monitoring of requirements.
Patent Information
- Application Number
- CN202111324010.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-11-10
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2041-11-10
AI Technical Summary
The existing signal recognition methods are slow to recognize in complex electromagnetic environments and cannot meet the real-time monitoring requirements. The IQ signal-based identification method requires a lot of time and cannot quickly process numerous signals.
Using a rapid signal recognition method based on spectrum templates, a Gaussian process regression model is established by resampling and classifying the spectrum profile information of known signals, and a parallel computing architecture is used for rapid matching and recognition, and the final signal type is performed in combination with Bayesian decisions. Output.
It realizes that the signal can be quickly identified without IQ data, improves the signal recognition efficiency and speed, and meets the real-time monitoring needs in complex electromagnetic environments.
Smart Images

Figure CN114091527B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of spectrum monitoring and signal interference troubleshooting, and in particular relates to a method for rapid signal recognition based on spectrum templates. Background Art
[0002] With the development of electronic technology, various new information equipments are emerging, and the electromagnetic signals radiated by electromagnetic radiation sources are increasing, forming a complex electromagnetic environment with wide frequency band, multiple types and fast changes. Under this circumstance, the demand for signal recognition ability in complex electromagnetic environment is becoming more and more urgent. According to the application requirements, it is necessary to realize real-time and fast signal recognition. However, the existing technical means only have the recognition ability based on IQ signals, and do not have the signal fast recognition ability based on spectrum template matching.
[0003] Moreover, there are two problems with the current signal rapid identification method: first, the principle of the existing signal identification method is based on the identification of IQ signals. In the process of obtaining the IQ signal, a series of processing is required, which takes a certain amount of time, resulting in a slow identification speed; second, there are numerous signals in a complex electromagnetic environment. If each signal is processed and identified, it will seriously affect the work efficiency and cannot meet the requirements of real-time monitoring. Summary of the invention
[0004] In order to solve the above problems, the present invention proposes a signal rapid identification method based on spectrum template, which compares the spectrum contour line of the known signal with the unknown signal. If the similarity is high, the unknown signal is considered to be the known signal, thereby solving the problem of rapid identification of broadband signals and forming the ability to rapidly identify broadband signals.
[0005] The technical solution of the present invention is as follows:
[0006] A signal rapid identification method based on spectrum template comprises the following steps:
[0007] S1. resampling the signal samples of the known spectrum template, and classifying the signals according to known categories to obtain multiple sub-sample sets;
[0008] S2. Establish a signal recognition sub-model for each sub-sample set through Gaussian process regression;
[0009] S3. Input an unknown signal, and input the unknown signal into each signal recognition sub-model simultaneously based on the parallel computing architecture, and each signal recognition sub-model outputs a prediction result that satisfies the Gaussian distribution;
[0010] S4. Through Bayesian decision making, the prediction outputs of all signal recognition sub-models are integrated to output the final predicted signal type.
[0011] Preferably, in step S2, the specific content of Gaussian process regression is:
[0012] S201. A set of training data consisting of n input-output pair samples is denoted as {(x i ,y i )|i=1,…,n}, input vector x i is a d-dimensional column vector, y i is the output and y i =f(x i );
[0013] S202. For a test input vector x * , whose predicted output distribution is an n+1-dimensional joint Gaussian distribution consisting of n training outputs and one test output, expressed as:
[0014]
[0015] The mean function of the joint Gaussian distribution is expressed as:
[0016] m(x * ) = k T (x * )K -1 y (2)
[0017] The variance of the joint Gaussian distribution is expressed as:
[0018] σ 2 (x * )=C(x * ,x * )-k T (x * )K -1 k(x * ) (3)
[0019] Among them, C(x i ,x j ) represents the vector x i and x j The covariance between k(x * ) represents the test input vector x * and the covariance vector between the training input vector, k(x * )=[C(x * ,x1),…C(x * ,x n )] T ; K is the n×n dimensional covariance matrix between the training input vectors, and its matrix elements are represented by K ij =C(x i ,x j );
[0020] S203. Determine the composite covariance function and establish a model. The composite covariance function is expressed as:
[0021] C(x i ,x j )=C SE (x i ,x j )×C PER (x i ,x j ) (4)
[0022]
[0023] C PER (x i ,x j ) = exp[(-2sin 2 [πx i -x j | / v2]) / v3] (6)
[0024] Among them, C SE (x i ,x j ) is the square exponential covariance function, C PER (x i ,x j ) is a periodic covariance function with hyperparameters v1, v2, v3 and ω t The value of (t=1,…d) is obtained by maximizing its log-likelihood estimate.
[0025] Preferably, the log-likelihood estimate of the hyperparameter is:
[0026]
[0027] Where α=[v1,v2,v3,ω t (t=1,…d)] T represents the vector composed of all hyperparameters, and |K| represents the value of the determinant composed of the elements in the matrix K;
[0028] The partial derivative of the log-likelihood estimate for each hyperparameter is expressed as:
[0029]
[0030] Preferably, a gradient optimization method is used to optimize the value of the hyperparameter to obtain a hyperparameter that can maximize the log-likelihood estimate;
[0031] The gradient optimization method uses the characteristic that the function changes fastest along the gradient direction to quickly find the maximum value of the function; the gradient of the log-likelihood estimation function L(α) is expressed as:
[0032]
[0033] According to the principle of gradient optimization method, the expression of its iterative process is:
[0034] α k+1 =α k +μ k ▽L(α) (10)
[0035] Where k represents the number of iterations, μ k represents the iteration step length;
[0036] When k reaches the set value or α k and α k-1 When the relative error between is less than the set parameter ε, take α k As the final hyperparameter value.
[0037] Preferably, in step S4, for a test input vector x of an unknown signal * Each signal recognition sub-model predicts a set of outputs that conform to the Gaussian distribution, including the mean m j (x * ),j=1,2,…u and variance For the outputs of all sub-models, Bayesian decision is used to merge these outputs into the final output. The specific expression is:
[0038]
[0039]
[0040] E(x * ) and C BCM Respectively represent the mean and variance of the final output, (v1) j , j = 1, 2, ... u represents the function prior distribution of each sub-model; finally, the signal type is obtained based on the mean result.
[0041] Preferably, the spectrum template is effective spectrum data of the signal, including the center frequency, bandwidth, and frequency interval or number of sampling points of the signal samples.
[0042] Beneficial technical effects brought by the present invention:
[0043] There is no need to extract the IQ data of the signal. The recognition model can be established based on the spectral profile information of the known signal and the unknown signal can be quickly matched and identified based on its spectral profile. In order to solve the problem of time-consuming processing of numerous signals, a high-efficiency parallel computing and analysis architecture based on the signal recognition model is established to achieve rapid model generation and rapid identification of unknown signals. BRIEF DESCRIPTION OF THE DRAWINGS
[0044] Figure 1 The flowchart of the method for rapid signal recognition based on spectrum template. DETAILED DESCRIPTION
[0045] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:
[0046] The present invention proposes a method for rapid signal recognition based on spectrum template, which realizes rapid learning of signal spectrum profile information and establishment of a rapid signal recognition model based on spectrum template. Signal recognition can be completed by correlation analysis of signal spectrum profile without obtaining signal IQ data, thereby solving the problem of rapid recognition of broadband signals, forming a rapid recognition capability for broadband signals, and providing advanced technical means for rapid monitoring and recognition of signals in complex electromagnetic environments.
[0047] The present invention firstly establishes a signal recognition model based on spectrum template according to the signal spectrum profile information in the known signal library by using the machine learning method based on Gaussian process and Bayesian decision. Then, according to the constructed signal recognition model, an advanced parallel computing architecture is adopted to complete the rapid matching and recognition of unknown signals by the signal recognition model through parallel computing.
[0048] like Figure 1 As shown, the specific steps are as follows:
[0049] Step 1: Resample the signal samples of the known spectrum template and classify them according to the known categories of the signal to obtain multiple sub-sample sets.
[0050] Step 2: Each sub-sample set can be used to establish a signal recognition sub-model using Gaussian process regression; the input data of the model is the signal spectrum contour line, including the center frequency, bandwidth, sampling point frequency interval or number that need to be determined by the sample, and the corresponding signal type is the output.
[0051] The specific steps of sub-model construction are as follows:
[0052] (1) Consider a training data set consisting of n input-output pairs {(x i ,y i )|i=1,…,n}, input vector x i is a d-dimensional column vector, y i is the output and y i=f(x i );
[0053] (2) For a test input vector x * , whose predicted output distribution is an n+1-dimensional joint Gaussian distribution consisting of n training outputs and one test output, which can be expressed as:
[0054]
[0055] Therefore, the mean function of the joint Gaussian distribution can be expressed as:
[0056] m(x * ) = k T (x * )K -1 y (2)
[0057] The variance of the joint Gaussian distribution can be expressed as:
[0058] σ 2 (x * )=C(x * ,x * )-k T (x * )K -1 k(x * ) (3)
[0059] Where C(x i ,x j ) represents the vector x i and x j The covariance between them, k(x * ) represents the test input vector x * and the covariance vector between the training input vector, denoted as k(x * )=[C(x * ,x1),…C(x * ,x n )] T ; K is the n×n dimensional covariance matrix composed of the training input vectors, and its matrix elements can be expressed as K ij =C(x i ,x j ), that is, K = C(x, x). Therefore, once the covariance function is determined, the model can be established.
[0060] (3) The covariance function contains a conjecture about the function to be learned and determines the similarity of the data. There is a restriction on the selection of the covariance function: for any data point in the training data sample, this covariance function must make the covariance matrix a semi-positive definite matrix. This is because the closer the two input data are, the greater their correlation is, and their covariance value should also be larger, and vice versa. The elements on the diagonal of the covariance matrix are the covariances of the input data with itself, which must be greater than or equal to the other elements in the matrix. Therefore, the covariance matrix is a matrix dominated by the main diagonal elements. For a matrix dominated by the main diagonal elements, the determinant values of its reduced-dimensional submatrix must be greater than or equal to 0, so it can be inferred that the covariance matrix is a semi-positive definite matrix.
[0061] In the present invention, a composite covariance function is proposed to be used, which is expressed as:
[0062] C(x i ,x j )=C SE (x i ,x j )×C PER (x i ,x j ) (4)
[0063]
[0064] C PER (x i ,x j ) = exp[(-2sin 2 [π|x i -x j | / v2]) / v3] (6)
[0065] Here C SE (x i ,x j ) is the squared-exponential (SE) covariance function, C PER (x i ,x j ) is a periodic covariance function. The hyperparameters v1, v2, v3 and ω t The value of (t=1,…d) can be obtained by maximizing its log-likelihood estimate. The log-likelihood estimate of the hyperparameter can be expressed as:
[0066]
[0067] where α=[v1,v2,v3,ω t (t=1,…d)] Trepresents the vector of all hyperparameters. |K| represents the value of the determinant of the elements in the matrix K. In order to find the hyperparameters that can achieve the maximum value of the log-likelihood estimate, the gradient optimization method can be used to optimize the value of the hyperparameter. The partial derivative of the log-likelihood estimate for each hyperparameter can be expressed as:
[0068]
[0069] The gradient optimization method is as follows: the gradient of a function at a certain point represents the directional derivative of the function at that point. The function changes fastest along that direction (the direction of the gradient) at that point, and the rate of change is the largest. Therefore, by utilizing the characteristic that the function changes fastest along the gradient direction, the maximum value of the function can be quickly found. According to the definition of the gradient, the gradient of the log-likelihood estimation function L(α) can be expressed as:
[0070]
[0071] According to the principle of gradient optimization method, the expression of its iterative process can be expressed as:
[0072] α k+1 =α k +μ k ▽L(α) (10)
[0073] Where k represents the number of iteration steps, μ k It represents the iteration step length. The initial value α0 can be any value in theory, but different values will result in different number of iteration steps to obtain the optimal value. k The value of should not be too large, otherwise it is easy to miss the optimal value. k and α k-1 The relative error between them is less than the set parameter ε, that is, the change of α obtained in the previous and next steps is very small, so α can be taken k As the final hyperparameter value.
[0074] Each set of sub-training sample sets can be used to establish a signal recognition sub-model S through Gaussian process. j (j=1,2,…u), that is, Gaussian process model, u represents the number of sub-agent models.
[0075] Step 3: For a test input of an unknown signal, each signal recognition sub-model will output a prediction result that satisfies the Gaussian distribution.
[0076] Step 4: The prediction outputs of each sub-model can be combined into the final prediction output through Bayesian decision making. The specific steps are as follows:
[0077] According to the Gaussian process regression introduced in step 2, for a test input vector x * , the signal recognition sub-model will predict a set of outputs that conform to the Gaussian distribution, that is, the mean m j (x * ),j=1,2,…u and variance For the outputs of all signal recognition sub-models, Bayesian decision is used to merge these outputs into the final output. The specific expression is:
[0078]
[0079]
[0080] E(x * ) and C BCM Represented as the mean and variance of the final output, (v1) j , j = 1, 2, ... u represents the function prior distribution of each sub-model. Among them, the signal type of the input unknown signal can be finally obtained through the output mean result.
[0081] In summary, the present invention aims to solve the problem that electromagnetic signals in complex electromagnetic environments are diverse and cannot be recognized by the naked eye. According to the signal spectrum template, an effective machine learning method is used to quickly establish a signal rapid recognition model based on the spectrum template; in response to the need for rapid recognition of electromagnetic signals, an advanced parallel computing architecture is adopted, and a signal rapid recognition model based on the spectrum template is used to quickly match and identify unknown signals.
[0082] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the protection scope of the present invention.
Claims
1. A fast signal recognition method based on spectrum template, characterized in that: The steps include: S1. resampling the signal samples of the known spectrum template, and classifying the signals according to known categories to obtain multiple sub-sample sets; S2. Establish a signal recognition sub-model for each sub-sample set through Gaussian process regression; S3. Input an unknown signal, and input the unknown signal into each signal recognition sub-model simultaneously based on the parallel computing architecture, and each signal recognition sub-model outputs a prediction result that satisfies the Gaussian distribution; S4. The prediction outputs of all signal recognition sub-models are integrated through Bayesian decision making, and the final predicted signal type is output; In step S2, the specific content of Gaussian process regression is: S201. A set of training data consisting of n input-output pair samples is denoted as {(x i ,y i )|i=1,…,n}, input vector x i is a d-dimensional column vector, y i is the output and y i =f(x i ); S202. For a test input vector x * , whose predicted output distribution is an n+1-dimensional joint Gaussian distribution consisting of n training outputs and one test output, expressed as: The mean function of the joint Gaussian distribution is expressed as: m(x * )=k T (x * )K -1 y (2) The variance of the joint Gaussian distribution is expressed as: σ 2 (x * )=C(x * ,x * )-k T (x * )K -1 k(x * ) (3) Among them, C(x i ,x j ) represents the vector x i and x j The covariance between k(x * ) represents the test input vector x * and the covariance vector between the training input vector, k(x * )=[C(x * ,x1),…C(x * ,x n )] T ; K is the n×n dimensional covariance matrix between the training input vectors, and its matrix elements are represented by K ij =C(x i ,x j ); S203. Determine the composite covariance function and establish a model. The composite covariance function is expressed as: C(x i ,x j )=C SE (x i ,x j )×C PER (x i ,x j ) (4) Among them, C SE (x i ,x j ) is the square exponential covariance function, C PER (x i ,x j ) is a periodic covariance function with hyperparameters v1, v2, v3 and ω t The value of (t=1,…d) is obtained by maximizing its log-likelihood estimate; The log-likelihood estimate of the hyperparameters is: Where α=[v1,v2,v3,ω t (t=1,…d)] T represents the vector composed of all hyperparameters, and |K| represents the value of the determinant composed of the elements in the matrix K; The partial derivative of the log-likelihood estimate for each hyperparameter is expressed as: The gradient optimization method is used to optimize the value of the hyperparameter to obtain the hyperparameter that can maximize the log-likelihood estimate; The gradient optimization method uses the characteristic that the function changes fastest along the gradient direction to quickly find the maximum value of the function; the gradient of the log-likelihood estimation function L(α) is expressed as: According to the principle of gradient optimization method, the expression of its iterative process is: Where k represents the number of iterations, μ k represents the iteration step length; When k reaches the set value or α k and α k-1 When the relative error between is less than the set parameter ε, take α k As the final hyperparameter value.
2. The method for rapid signal recognition based on spectrum template according to claim 1, characterized in that: In step S4, for a test input vector x of an unknown signal * Each signal recognition sub-model predicts a set of outputs that conform to the Gaussian distribution, including the mean m j (x * ),j=1,2,…u and variance j=1,2,…u; For the outputs of all sub-models, Bayesian decision is used to merge these outputs into the final output. The specific expression is: E(x * ) and C BCM Respectively represent the mean and variance of the final output, (v1) j , j = 1, 2, ... u represents the function prior distribution of each sub-model; finally, the signal type is obtained based on the mean result.
3. The method for rapid signal recognition based on spectrum template according to claim 1, characterized in that: The spectrum template is valid spectrum data of the signal, including the center frequency, bandwidth, and frequency interval or number of sampling points of the signal samples.
