A method for detecting and locating incoherent scattering sources based on convolutional neural networks
Through the method based on convolutional neural network, the problem of difficult calculation and strict distribution requirements for signal subspace in incoherent scattering source positioning is solved, and the positioning and source number estimation of multiple scattering sources is realized, which is suitable for large-angle expansion scenarios. Improves positioning accuracy and robustness.
Patent Information
- Application Number
- CN202210478993.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-05-05
- Publication Date
- 2024-12-06
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The prior art has difficulty in accurately calculating signal subspace and noise subspace in the positioning of incoherent scattering sources, requires accurate direction angle power density functions, is only suitable for small angle expansion and point source models, and cannot directly expand to multiple scattering. The scene of the source and requires prior information of the number of sources.
Using a convolutional neural network-based approach, the convolutional neural network is trained to estimate the direction angle and angle extension parameters of the source signal by generating the data set for training, including the sample complex covariance matrix and direction angle power density curve, to avoid sub- Spatial decomposition, using spectral peak search for parameter estimation, without the need for source signal distribution prior information.
The positioning and source number estimation of multiple incoherent scattering sources is realized, the requirements for source signal distribution are relaxed, and it is suitable for large-angle expansion scenarios, without the need for source number prior information, and the positioning accuracy and robustness are improved.
Smart Images

Figure CN115097378B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of array signal processing and deep learning technology, and in particular to an incoherent scattering source positioning method, and specifically to an incoherent scattering source detection and positioning method based on a convolutional neural network. Background Art
[0002] Direction of arrival (DOA) estimation is a basic problem in array signal processing. Currently, there are many high-resolution DOA estimation methods based on the point source model assumption. However, in many practical scenarios, source signals often experience multipath propagation or scattering. For the receiving array, the source signal at this time has spatial distribution characteristics and usually does not meet the assumption of the point source model. The signal needs to be modeled as a scattering source model.
[0003] In an incoherent scattering source, the scattered signals from different directions are unrelated to each other, and the point source model algorithm cannot be directly applied to the incoherent source model. There are also a large number of research results on the positioning methods of scattering sources, but these methods have some limitations. For example, some methods only locate a single source signal and are not suitable for parameter estimation of multiple scattering sources. The signal subspace and noise subspace of the subspace-based scattering source positioning method are difficult to accurately determine, resulting in errors in the estimation results. In addition, it is required that the directional angular power density function of the source signal is accurately known and the distribution type of multiple source signals is the same. When the directional angular power density distribution of the real source signal is inconsistent with the assumed distribution model, the wave direction estimation will be affected by the model mismatch problem. Some methods require a global spectrum peak search, and the algorithm has a large time complexity. Other methods are based on the Taylor expansion approximation and are only applicable to scenarios with small angle expansion.
[0004] At present, some works have applied deep learning technology to solve the DOA estimation problem, but the DOA estimation method based on deep learning is still aimed at the point source model, and there is no deep learning method for the incoherent scattering source model.
[0005] Summarizing the existing methods for locating incoherent scatter sources, there are the following main problems:
[0006] 1. The signal subspace and noise subspace are difficult to calculate accurately. Using a subspace with errors will lead to performance degradation. In addition, this type of method requires the directional angular power density function of the source signal to be accurately known, and the distribution type of multiple source signals must be consistent, requiring a global spectrum peak search; 2. The existing scattering source positioning method based on Taylor expansion approximation is only applicable to scenarios with small angle extensions. 3. The DOA estimation method based on deep learning is only for point source models and cannot be directly extended to incoherent scattering source models. 4. Existing methods require prior information on the number of known sources. Summary of the invention
[0007] The purpose of the present invention is to solve the above-mentioned defects of the prior art and provide a method for detecting and locating incoherent scatter sources based on a convolutional neural network, so as to realize the positioning of multiple incoherent scatter sources and estimate the number of sources at the same time.
[0008] The purpose of the present invention can be achieved by adopting the following technical solutions:
[0009] A method for detecting and locating an incoherent scattering source based on a convolutional neural network, the method comprising the following steps:
[0010] S1. Set the number of array elements M, the array observation space range Θ, the source signal directional angle and angle expansion parameters, the directional angle power density discretization accuracy, and the discretization grid set. The number of grids g;
[0011] S2. Generate a dataset for training convolutional neural networks Dataset Each set of data samples in includes the sample complex covariance matrix R and the corresponding noise-free directional angular power density curve η is the source signal parameter set, η=[η1,η2,...,η K ],η i =(θ i ,σ i ) represents the parameter pair of the i-th source signal, θ i and σ i are the direction angle parameter and angle expansion parameter of the i-th source signal, respectively, K is the number of sources and 1≤i≤K;
[0012] S3. Dataset The sample complex covariance matrix is converted into a dual-channel real matrix and normalized;
[0013] S4. Using the Dataset Iteratively train the convolutional neural network to make the loss function converge to the minimum value and obtain the model parameter set
[0014] S5. Use model parameter set Initialize the convolutional neural network, generate simulated signals or collect real scene signals, and input them into the convolutional neural network after data conversion and normalization to obtain the output is the angular power density estimation curve, is the estimated parameter set of the source signal, represents the estimated parameter pair of the kth source signal, and are the estimated value of the direction angle and the estimated value of the angle extension, respectively. is an estimate of the number of sources;
[0015] S6. Acquisition The number of peaks p, take the number of peaks with p>β As a parameterized estimate of the number of sources, β is the peak judgment threshold;
[0016] S7, according to the parameter estimation method Compute parameterized source signal direction angle estimates and angular spread estimates.
[0017] Furthermore, the calculation process of the source signal direction angle estimation value and the angle extension estimation value in step S7 is as follows:
[0018] The peak value corresponding to the direction angle φ is obtained from the peak value p k ,like At two adjacent peak direction angles φ a With φ b Get within the range The minimum value, a, And ba=1, the minimum value is the split point. Split into share is the directional power density estimation curve of the kth source signal, like Then φ a =-90°,φ b =90°;
[0019] calculate Make the sum of the integrals of each segmentation equal to 1;
[0020] The estimated direction angle is calculated according to the following formula and the angle spread estimate
[0021]
[0022] in, yes The power density value at the direction angle φ. In the parameter estimation process, the source signal distribution prior information is not used. It is only assumed that the directional angle power density function of each source signal is a unimodal function, so the requirements for distribution are relaxed. It is not necessary to accurately know the directional angle distribution density function of the source signal, nor is it required that the distribution type of each source signal is the same.
[0023] Furthermore, in step S7, the spatial spectrum is calculated, and the source signal direction angle and angle extension estimation value are obtained by spectrum peak search, wherein the spatial spectrum function P(η S ) is expressed as:
[0024]
[0025] Among them, ||·|| F is the F-norm, is the distribution used for peak search, η S =(θ,σ),η S Contains parameterized direction angle θ and angle spread σ with adjustable search resolution. θ and σ corresponding to the spectrum peak are estimated results. S ) has nothing to do with the specific array manifold, so P(η S ) has no special requirements on the array shape.
[0026] Furthermore, in step S2, a data set for training a convolutional neural network is generated. The process is as follows:
[0027] Assume that there are K far-field incoherent narrowband scattering source signals received by an arbitrary array of M elements, and the sample complex covariance matrix R is:
[0028]
[0029] Where a(φ) represents the direction vector of the array when the point source direction angle is φ, ρ i (φ,η i ) is the power density value of the ith source signal at the direction angle φ, is the power of the ith source signal, is the noise power, I is the identity matrix of dimension M×M, [·] H represents the conjugate transpose operation, and the dth group of samples is defined as D is the total number of samples, is the angular power density curve of the ith source signal, data set Contains different noise power samples, so that the convolutional neural network can learn common features during the training process, enhance its robustness to noise, and better adapt to real signals.
[0030] Furthermore, in step S3, the sample complex covariance matrix R is decomposed into a dual-channel real matrix R′, where R : ' ,:,1 =Re[R], R : ' ,:,2=Im[R], Re[·] and Im[·] represent the real part and imaginary part operations respectively, and finally R′ is normalized, that is, R′=R′ / max(R′). Since the sample value range is unpredictable under different signal-to-noise ratios and different source signal power conditions, directly using these samples to train the convolutional neural network will cause oscillations in its gradient update phase, affecting the convergence of the loss function, and even causing overfitting problems, which will lead to poor performance. Normalization can avoid these problems.
[0031] Furthermore, the convolutional neural network is composed of 6 neural network layers, wherein the first neural network layer includes a 2D convolution layer with 8 filters and a convolution kernel size of 3×3, an L2 regularization layer with a regularization parameter of 0.02, and an ELU activation layer, which are sequentially connected; the second neural network layer includes a 2D convolution layer with 16 filters and a convolution kernel size of 4×4, an L2 regularization layer with a regularization parameter of 0.02, and an ELU activation layer, which are sequentially connected; the third neural network layer is a vectorization layer; the fourth neural network layer includes a fully connected layer consisting of 400 neurons and an ELU activation layer, which are sequentially connected; the fifth neural network layer includes a fully connected layer consisting of 200 neurons and an ELU activation layer, which are sequentially connected; and the sixth neural network layer includes a fully connected layer consisting of g neurons and a Softmax activation layer, which are sequentially connected. The output of the convolutional neural network is an angular power density estimation curve with a grid number of g, which fits the mapping relationship from the covariance matrix to the angular power density curve, avoiding the decomposition of the subspace.
[0032] Furthermore, the loss function is a symmetric KL divergence, that is:
[0033]
[0034] in, It is a measure and The loss function of the difference between and When they are completely consistent, the value is 0. During the iterative training process, the parameters of the convolutional neural network are continuously optimized to make the loss function converge to the minimum value. After iterative training, the model parameter set is obtained.
[0035] Compared with the prior art, the present invention has the following advantages and effects:
[0036] 1. The traditional subspace-based incoherent scattering source positioning method needs to assume that the directional angular power density function of the source signal is accurately known and the distribution type of each source is the same. The present invention uses a convolutional neural network to output a source signal directional angular power density estimation curve. There is no need to assume that the directional angular power density function of the source signal is known, and it only needs to satisfy that the directional angular power density function of each source signal is a unimodal function.
[0037] 2. The positioning method disclosed in the present invention is based on deep learning technology. It fits the mapping relationship between the covariance matrix and the angular power density curve through a large amount of sample data, avoids the decomposition of the subspace, and can achieve effective parameter estimation without multi-dimensional spectral peak search. At the same time, the spatial spectrum can also be calculated, and the spectral peak search method can be used for parameter estimation.
[0038] 3. The positioning method disclosed in the present invention has no special requirements on the array shape and is applicable to the positioning of multiple source signals. The convolutional neural network outputs the source signal directional angle power density estimation curve. The size of the source signal angle expansion parameter does not affect the output accuracy, so it can be effectively estimated under large angle expansion.
[0039] 4. The method disclosed in the present invention calculates the number of sources and realizes source positioning according to the source signal directional angle power density estimation curve, does not require prior information on the number of sources, and has fewer restrictions. BRIEF DESCRIPTION OF THE DRAWINGS
[0040] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:
[0041] Figure 1 It is a structural diagram of a convolutional neural network used in the incoherent scattering source detection and positioning method based on a convolutional neural network disclosed in the present invention;
[0042] Figure 2 It is a flow chart of the incoherent scattering source detection and positioning method based on convolutional neural network disclosed in the present invention;
[0043] Figure 3 This is a test result diagram of the present invention using simulated data; wherein, Figure 3 (a) is the azimuth root mean square error curve. Figure 3 (b) is the angle expansion root mean square error curve, Figure 3 (c) is the estimation accuracy of the number of source signals;
[0044] Figure 4 This is a graph of the actual data test results collected by the present invention using a 10-element uniform linear microphone array; wherein, Figure 4 (a) is the direction angle estimation result of the present invention, Figure 4(b) is the angle expansion estimation result of the present invention, Figure 4 (c) is the spatial spectrum diagram of the parameter estimation method 2 of the present invention. DETAILED DESCRIPTION
[0045] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0046] Embodiment 1
[0047] This embodiment discloses a method for detecting and locating incoherent scattering sources based on a convolutional neural network. The method uses a large amount of sample data to fit the mapping relationship between the covariance matrix and the directional angular power density of the source signal, obtains the directional angular power density estimation curve, and then performs peak detection. The number of peaks is the number of sources. Then, the curve is segmented according to adjacent peaks to extract the density of each scattering source, and finally the directional angle and angle expansion parameters are solved. At the same time, the directional angular power density curve can also be used to construct a spatial spectrum, and the parameters are solved by a spectrum peak search method.
[0048] Figure 1 is a structural diagram of a convolutional neural network in the incoherent scattering source detection and positioning method based on a convolutional neural network disclosed in this embodiment. Figure 2 The flowchart of the incoherent scattering source detection and positioning method based on convolutional neural network disclosed in this embodiment is shown. The present invention includes the following steps:
[0049] S1, set the number of array elements M = 10, the array observation space range -90°≤Θ≤90°, the angular power density discretization accuracy is 1°, and the discretization grid set is Grid number g = 181;
[0050] S2: Generate a data set for training a convolutional neural network based on the parameters in step S1 Dataset Each set of data samples in includes the sample complex covariance matrix R and the corresponding noise-free directional angular power density curve η is the source signal parameter set, η=[η1,η2,...,η K ],η i =(θ i ,σ i ) represents the parameter pair of the i-th source signal, θ i and σ iare their direction angle parameters and angle expansion parameters respectively, K is the number of sources and 1≤i≤K;
[0051] S3. Dataset The sample complex covariance matrix is converted into a dual-channel real matrix and normalized;
[0052] S4. Using the Dataset The convolutional neural network is iteratively trained to make the loss function converge to the minimum value. The loss function is the symmetric KL divergence, that is:
[0053]
[0054] in, It is a measure and The loss function of the difference between After iterative training, the model parameter set is obtained
[0055] S5. Use Initialize the convolutional neural network, set two source signals that satisfy the Gaussian distribution, with parameters η1 = (-30°, 3°) and η2 = (30°, 3°), and a signal-to-noise ratio of 20 dB. Generate an analog signal, which is input into the convolutional neural network after data conversion and normalization to obtain the angular power density estimation curve. is the estimated parameter set of the source signal, represents the estimated parameter pair of the kth source signal, and are the estimated value of the direction angle and the estimated value of the angle extension, respectively. is an estimate of the number of sources;
[0056] S6. Acquisition The number of peaks p, take the number of peaks with p>β As the parameterized source quantity estimate, β is the peak judgment threshold. In this example, β is set to 0.02 to prevent Medium to small noise impact;
[0057] S7, according to the parameter estimation method Calculate the parameterized source signal direction angle estimate and angle spread estimate. The specific calculation process is as follows:
[0058] Parameter estimation method 1:
[0059] The peak value corresponding to the direction angle φ is obtained from the above peak value p k ,like At two adjacent peak direction angles φ aWith φ b Get within the range The minimum value, a, And ba=1, the minimum value is the split point. Split into share is the directional power density estimation curve of the kth source signal, like Then φ a =-90°,φ b =90°;
[0060] calculate Make the sum of the integrals of each segmentation equal to 1;
[0061] The estimated direction angle is calculated according to the following formula and the angle spread estimate
[0062]
[0063] in, yes The power density value at the direction angle φ.
[0064] Parameter estimation method 2:
[0065] use Calculate the spatial spectrum and obtain the source signal direction angle and angle extension estimation value by searching the spectrum peak, where the spectrum function P(η S ) is expressed as follows:
[0066]
[0067] Among them, ||·|| F is the F-norm, is the distribution used for peak search, η S =(θ,σ),η S The parameterized central angle θ and angle extension σ with adjustable search resolution are included, and the θ and σ corresponding to the spectrum peak are the estimated results. is a Gaussian distribution.
[0068] The test results of this embodiment are as follows Figure 3 As shown, Figure 3 (a) is the azimuth root mean square error curve; Figure 3 (b) is the angle expansion root mean square error curve, Figure 3 (c) is the source quantity estimation accuracy. It can be seen that the performance of the method disclosed in the present invention is significantly better than the other two subspace algorithms.
[0069] Embodiment 2
[0070] Combine the following Figure 2 , describing the specific implementation steps of the detection and positioning of the incoherent scattering source detection and positioning method based on convolutional neural network in the real signal scenario of the present invention.
[0071] In this embodiment, considering the real data collected by a 10-element uniform linear microphone array, the incoherent scattering source detection and positioning method based on a convolutional neural network disclosed in the present invention specifically includes the following steps:
[0072] S1, set the number of array elements M = 10, the array formation is a uniform linear array, the array observation space range is -90°≤Θ≤90°, the directional angle power density discretization accuracy is 1°, and the discretization grid set is The number of grids g = 181;
[0073] S2: Generate a data set for training a convolutional neural network based on the parameters in step S1 Dataset Each set of data samples in includes the sample complex covariance matrix R and the corresponding noise-free directional angular power density curve η is the source signal parameter set, η=[η1,η2,...,η K ],η i =(θ i ,σ i ) represents the parameter pair of the i-th source signal, θ i and σ i are their direction angle parameters and angle expansion parameters respectively, K is the number of sources and 1≤i≤K;
[0074] S3. Dataset The sample complex covariance matrix is converted into a dual-channel real matrix and normalized;
[0075] S4. Using the Dataset The convolutional neural network is iteratively trained to make the loss function converge to the minimum value. The loss function is the symmetric KL divergence, that is:
[0076]
[0077] in, It is a measure and The loss function of the difference between After iterative training, the model parameter set is obtained
[0078] S5. Use Initialize the convolutional neural network and collect a single tone signal of a real scene. The sound source frequency is 13.076KHz, the straight-line distance between the sound source and the array is 2.0 meters, and the audio recording time is 55 seconds. The collected data is divided into 97 parts at a time interval of 500 milliseconds. Each piece of data is recorded as a sampling point. The sampling point data is input into the convolutional neural network after data conversion and normalization to obtain the directional angle power density estimation curve of the sampling point data is the estimated parameter set of the source signal, represents the estimated parameter pair of the kth source signal, and are the estimated value of the direction angle and the estimated value of the angle extension, respectively. is an estimate of the number of sources;
[0079] S6. Acquisition The number of peaks p, take the number of peaks with p>β As the parameterized source quantity estimate, β is the peak judgment threshold. In this example, β is set to 0.1 to prevent Medium to small noise impact;
[0080] S7, according to the parameter estimation method Calculate the parameterized source signal direction angle estimate and angle spread estimate. The specific calculation process is as follows:
[0081] Parameter estimation method 1:
[0082] The peak value corresponding to the direction angle φ is obtained from the above peak value p k ,like At two adjacent peak direction angles φ a With φ b Get within the range Minimum value, And ba=1, the minimum value is the split point. Split into share is the directional power density estimation curve of the kth source signal, like Then φ a =-90°,φ b =90°;
[0083] calculate Make the sum of the integrals of each segmentation equal to 1;
[0084] The estimated direction angle is calculated according to the following formula and the angle spread estimate
[0085]
[0086] in, yes The power density value at the direction angle φ.
[0087] Parameter estimation method 2:
[0088] use Calculate the spatial spectrum and obtain the source signal direction angle and angle extension estimation value by searching the spectrum peak, where the spectrum function P(η S ) is expressed as follows:
[0089]
[0090] Among them, ||·|| F is the F-norm, is the distribution used for peak search, η S =(θ,σ),η S The parameterized central angle θ and angle extension σ with adjustable search resolution are included, and the θ and σ corresponding to the spectrum peak are the estimated results. is a Gaussian distribution.
[0091] Figure 4 This is a graph of the actual data test results collected using a 10-element uniform linear microphone array in this embodiment, where Figure 4 (a) is the direction angle estimation result of each sampling point data, Figure 4 (b) is the angle expansion estimation result of each sampling point data, Figure 4 (c) is the spatial spectrum diagram of parameter estimation method 2. It can be seen that the variance of the estimation result of the present invention is small and the result value is relatively stable; the peak of the spatial spectrum calculated by parameter estimation method 2 is sharp; only using simulated data simulation training can adapt to the real collected signal.
[0092] In summary, the embodiment of the present invention proposes a method for detecting and locating incoherent scattering sources based on a convolutional neural network. Different from the traditional subspace-based algorithm, the main idea of the present invention is to train the convolutional neural network with a large amount of sample data, so that it fits the mapping relationship between the covariance matrix and the directional angular power density of the source signal, and then perform peak detection after obtaining the directional angular power density curve. The number of peaks is the number of sources, and then the density of each scattering source is extracted by curve segmentation according to the adjacent peaks of the directional angular power density curve, and finally the directional angle and angle expansion parameters are solved; at the same time, the directional angular power density curve can also be used to construct a spatial spectrum, which is solved by the spectrum peak search method. Therefore, the present invention completes the estimation of the number of sources while estimating the directional angle and angle expansion parameters of the source signal, avoids the problem of subspace effective dimension selection of the traditional subspace-based algorithm, relaxes the requirements for the distribution of the source signal, does not require the distribution prior information of the source signal, can be effectively estimated under large angle expansion, and only uses simulated data to train the convolutional neural network to adapt it to the real signal, with excellent overall detection, positioning performance and generalization ability.
[0093] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.
Claims
1. A method for detecting and locating incoherent scatter sources based on convolutional neural networks, characterized in that: The detection and positioning method comprises the following steps: S1. Set the number of array elements , Array observation space range , source signal direction angle and angle expansion parameters, direction angle power density discretization accuracy, discretization grid set , Grid Number ; S2. Generate a dataset for training convolutional neural networks , dataset Each set of data samples includes the sample complex covariance matrix And the corresponding noise-free angular power density curve , is the source signal parameter set, , Indicates The parameter pairs of source signals, and Respectively Direction angle parameters and angle expansion parameters of source signals, is the number of sources and ; S3. Dataset The sample complex covariance matrix is converted into a dual-channel real matrix and normalized; S4. Using the Dataset Iteratively train the convolutional neural network to make the loss function converge to the minimum value and obtain the model parameter set ; S5. Use model parameter set Initialize the convolutional neural network, generate simulated signals or collect real scene signals, and input them into the convolutional neural network after data conversion and normalization to obtain the output , is the angular power density estimation curve, is the estimated parameter set of the source signal, , Indicates The estimated parameter pairs of source signals are and are the estimated value of the direction angle and the estimated value of the angle extension, respectively. , is an estimate of the number of sources; S6. Acquisition Several peaks ,Pick The number of peaks is used as an estimate of the number of parameterized sources , is the peak judgment threshold; S7, according to the parameter estimation method Calculate the parameterized source signal direction angle estimate and angle spread estimate, where the calculation process of the source signal direction angle estimate and angle spread estimate is as follows: By peak Get the peak corresponding to the angle ,like , at two adjacent peak direction angles and Get within the range Minimum value, and , with the minimum value as the split point exist The range is divided into share , For the The directional power density estimation curve of the source signal is: ;like ,but ; calculate , so that the sum of the integrals of each segmented part is 1; The estimated direction angle is calculated according to the following formula and the angle spread estimate : in, yes At the direction angle The power density value.
2. The method for detecting and locating incoherent scatter sources based on convolutional neural networks according to claim 1, characterized in that: In step S7, the spatial spectrum is calculated, and the source signal direction angle and angle extension estimation value are obtained by spectrum peak search, wherein the spatial spectrum function The expression is: in, is the F-norm, is the distribution used for peak search, , Contains parameterized direction angles with adjustable search resolution and angle expansion , the peak corresponds to and For the estimated results.
3. The incoherent scatter source detection and positioning method based on convolutional neural network according to claim 1, characterized in that: The step S2 generates a data set for training a convolutional neural network. The process is as follows: Assume that A far-field incoherent narrowband scattering source signal is Arbitrary array of array elements received, sample complex covariance matrix for: in, Indicates that the array is at a point source direction angle of The direction vector at It is The source signal has a direction angle of The power density value, For the The power of the source signal, is the noise power, The dimension is The identity matrix of represents the conjugate transpose operation, The group sample is defined as , , is the total number of samples, , For the Directional power density curve of source signals, data set .
4. The incoherent scatter source detection and positioning method based on convolutional neural network according to claim 1, characterized in that: In step S3, the sample complex covariance matrix Decompose into a two-channel real matrix ,in , , and Respectively represent the real part and imaginary part operations, and finally Normalize the data, that is .
5. The incoherent scatter source detection and positioning method based on convolutional neural network according to claim 1, characterized in that: The convolutional neural network consists of 6 neural network layers, wherein the first neural network layer includes 8 filters connected sequentially and the convolution kernel size is 2D convolutional layer, L2 regularization layer with regularization parameter of 0.02 and ELU activation layer; the second neural network layer includes 16 filters connected sequentially and the convolution kernel size is The third neural network layer is a vectorization layer; the fourth neural network layer includes a fully connected layer consisting of 400 neurons connected in sequence and an ELU activation layer; the fifth neural network layer includes a fully connected layer consisting of 200 neurons connected in sequence and an ELU activation layer; the sixth neural network layer includes a fully connected layer consisting of The fully connected layer consists of neurons and the Softmax activation layer.
6. The incoherent scatter source detection and positioning method based on convolutional neural network according to claim 1, characterized in that: The loss function is a symmetric KL divergence, namely: in, It is a measure and The loss function of the difference between , , after iterative training, the model parameter set is obtained .
