Unified direction finding method and system based on generalized deep neural network model

Abstract

The present application relates to a kind of unified direction-finding method and system based on generalized deep neural network model of multiple frequency points, belong to radio direction-finding technical field.The method includes: by antenna array two-dimensional angle (azimuth angle and elevation angle) direction-finding range and observation frequency band produce training sample, based on the input of the training sample of array receiving signal feature information and frequency point information, based on the output of the training sample of signal corresponding two-dimensional angle information;All sample data is classified into several frequency band partitions and non-uniformly classified into several angle partitions;A kind of generalized deep neural network including classification network and regression network is used to train training sample, obtain two-dimensional angle direction-finding model;The direction of arrival of signal is obtained by using the model to carry out two-dimensional angle direction-finding to actual incident signal.The method realizes two-dimensional angle direction-finding, avoids the large amount of correction data of the one frequency point one correction of traditional interferometer direction-finding, realizes high-precision direction-finding in non-ideal scene.

Description

Technical Field

[0001] This invention belongs to the field of radio direction finding technology, and in particular relates to a multi-frequency unified direction finding method and system based on a generalized deep neural network model. Background Technology

[0002] Radio direction finding refers to the process by which a direction-finding station determines the angle between a ray from the station to a radiation source and a specified reference method by measuring and calculating electromagnetic field parameters. Radio direction finding has both military and non-military applications. In military applications, direction finding can obtain directional information about a radiation source, and even its geographical location. Based on this information, for military communication systems, it can preliminarily determine the deployment of enemy radiation sources or the distribution of important communication hubs and nodes; it can perform signal sorting such as frequency hopping and spread spectrum to determine the relationships and communication status of communication networks, and subsequently determine the network attributes and threat levels; it can guide narrow-beam jammers to target the receiving direction or area to carry out jamming, improving the effectiveness and accuracy of jamming, and can even guide fire strikes. In non-military applications, based on direction finding information, it can identify interference sources or illegal radiation sources for radio spectrum monitoring; it can track and locate criminals during communication for arrest and combating criminals; it can use radiation sources at known locations for navigation of moving platforms; it can receive and radiate signals in a specified direction to achieve space division multiple access intelligent communication; and it can be used in scientific research fields such as radio astronomy and Earth remote sensing, as well as in civilian fields such as search and rescue and exploration.

[0003] Most direction-finding methods assume that the direction-finding system has high-precision measurement of signal parameters and an accurate parameter response model. However, complex scenarios in actual engineering applications usually involve various errors. For non-ideal scenarios, typical methods include field tabulation and correction methods for various errors. However, these correction methods still have some shortcomings: First, existing correction methods only target one specific error factor. When multiple error factors exist simultaneously, the construction of the correction model is complex and difficult to effectively correct. Second, existing methods require the correction sample data to correspond one-to-one with the signal frequency. That is, different correction sample data needs to be constructed for different frequencies. Therefore, this type of one-frequency-one-correction model requires a large amount of correction data. Although the engineering implementation is relatively simple, the initial correction process is too cumbersome. Summary of the Invention

[0004] Based on the above analysis, this invention aims to provide a multi-frequency unified direction finding method and system based on a generalized deep neural network model. Leveraging the strong nonlinear modeling capabilities of machine learning, a multi-frequency unified direction finding model is constructed. A model output format suitable for two-dimensional angle (azimuth and elevation) direction finding is proposed, while avoiding the large amount of correction data required by traditional one-frequency-point-one-correction methods, thus achieving high-precision direction finding in non-ideal scenarios.

[0005] On the one hand, the present invention provides a multi-frequency unified direction finding method based on a generalized deep neural network model, specifically including the following steps:

[0006] Construct training samples, wherein the input of the training samples is obtained based on the feature information and frequency information of the array received signals, and the output of the training samples is obtained based on the two-dimensional angle information corresponding to the signals;

[0007] The training samples are partitioned;

[0008] The network parameters of the deep neural network model are trained based on the partitioned training samples to obtain a two-dimensional angle-finding model; wherein, the deep neural network model includes a multi-layer network structure corresponding to the partition;

[0009] Using the aforementioned two-dimensional angle direction finding model, two-dimensional angle direction finding is performed on the incident signal to obtain the azimuth and elevation angles of the incident signal.

[0010] Furthermore, the feature information includes feature vector information composed of a normalized upper half-diagonal array of the received array signals, and the two-dimensional angle information includes vector information composed of the sine and cosine functions of the azimuth and elevation angles.

[0011] Furthermore, the input information includes frequency point information copied N(N-1) times and a normalized upper half-diagonal array of signals received by the array. The resulting characteristic vector is denoted as vector I. Where N represents the number of elements in the receiving array, x (t) This indicates that the array receives signals. x represents (t) The normalized upper half-diagonal matrix of the covariance matrix.

[0012] Furthermore, the relationship between the signal incident angle θ and the frequency f satisfies the function F(θ,f)=χ2; where θ represents the incident angle of the signal, α and β represent the azimuth and elevation angles of the signal respectively, θ=[α,β]; χ2 is the two-dimensional angle information, χ2=[sinα,cosα,sinβ,cosβ].

[0013] Furthermore, the training samples are partitioned based on non-uniform frequency bands and non-uniform angles, the partitioning including:

[0014] All sample data are non-uniformly classified into several frequency band partitions based on their observed frequency bands;

[0015] Within each frequency band partition, the associated data is non-uniformly classified into several angular partitions based on its direction-finding range for the incident angle.

[0016] Furthermore, within each partition, the ratio of the maximum to the minimum value of the error covariance matrix corresponding to the incident angle is no greater than a preset value.

[0017] Furthermore, in the non-uniform angle partitioning, the angle boundaries of adjacent partitions overlap.

[0018] Furthermore, the deep neural network model includes N corresponding to the plurality of frequency band partitions. f Layer classification network, where N f This represents the number of frequency band partitions; each layer of the classification network includes N corresponding to several angle partitions within that frequency band partition. z A regression network, N z This indicates the number of angular partitions in the frequency band partition; wherein, the regression network is used to perform direction finding estimation on the data of the angular partition.

[0019] On the other hand, the present invention also provides a multi-frequency unified direction finding system based on a generalized deep neural network model, comprising:

[0020] The training sample acquisition module is used to acquire training sample data of the deep neural network. It obtains the input of the training sample based on the feature information and frequency information of the array received signal, and obtains the output of the training sample based on the two-dimensional angle information corresponding to the signal.

[0021] The training sample partitioning module is used to partition the training samples;

[0022] A generalized deep neural network model construction and training module is used to construct the deep neural network model and train its network parameters based on the partitioned training samples to obtain a two-dimensional angle-finding model; wherein, the deep neural network model includes a multi-layer network structure corresponding to the partition;

[0023] The two-dimensional angle direction finding module uses the two-dimensional angle direction finding model to perform two-dimensional angle direction finding on the actual incident signal, and obtains the azimuth and elevation angles of the incident signal.

[0024] Furthermore, the training sample partitioning module partitions the training samples based on non-uniform frequency bands and non-uniform angles, the partitioning including:

[0025] All sample data are non-uniformly classified into several frequency band partitions based on their observed frequency bands;

[0026] Within each frequency band partition, the associated data is non-uniformly classified into several angular partitions based on its direction-finding range for the incident angle.

[0027] The present invention can achieve at least one of the following beneficial effects:

[0028] 1. By constructing a generalized deep neural network model, training samples are built and partitioned. The deep neural network model is trained based on the partitioned training samples to obtain a two-dimensional angle (azimuth and elevation) direction finding model, thereby achieving high-precision direction finding in non-ideal scenarios.

[0029] 2. By approximating the training samples based on non-uniform frequency bands and non-uniform angles, and with the multi-layer network structure of the deep neural network model corresponding to the partitions, the large amount of correction data required for traditional one-frequency-point-one-correction is avoided, making the training process more efficient and achieving high-efficiency construction of the direction-finding model.

[0030] Other features and advantages of the invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained from what is particularly pointed out in the description, claims, and drawings. Attached Figure Description

[0031] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0032] Figure 1 This is a flowchart illustrating the method described in an embodiment of the present invention;

[0033] Figure 2 This is a diagram of the DNN model structure (training phase) according to an embodiment of the present invention;

[0034] Figure 3 This is an empirical CDF comparison chart for a non-ideal scenario according to an embodiment of the present invention;

[0035] Figure 4 This is an empirical CDF comparison chart for a 1007.5MHz test set according to an embodiment of the present invention;

[0036] Figure 5 This is a direction finding RMS contour line of 1007.5MHz at various incident angles, according to an embodiment of the present invention.

[0037] Figure 6 This is a structural diagram of the multi-frequency unified direction finding system based on a generalized deep neural network model, as presented in this invention. Detailed Implementation

[0038] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0039] Method Implementation Examples

[0040] Example 1

[0041] A specific embodiment of the present invention discloses a multi-frequency unified direction finding method based on a generalized deep neural network model, which specifically includes the following steps:

[0042] The first step is training sample acquisition. Specifically, the input of the training sample is obtained based on the feature information and frequency information of the array received signal, and the output of the training sample is obtained based on the two-dimensional angle information corresponding to the signal.

[0043] In the direction-finding range of an N-element antenna array, the azimuth angle [α] l α u ] and pitch angle [β l ,β u and observation frequency band [f l f u N is generated within. S There are training samples, where the subscript l represents the lower limit of the range and the subscript u represents the upper limit of the range;

[0044] Assuming the received signal is x(t), its covariance matrix is:

[0045]

[0046] Its normalized upper half-diagonal matrix:

[0047]

[0048] To and The real and imaginary matrix elements are kept consistent, and the frequency point f is copied N(N-1) times. The feature vectors formed together constitute vector I, which serves as the input to the training samples.

[0049]

[0050] The corresponding output is a vector χ2 = [sinα, cosα, sinβ, cosβ] composed of the sine and cosine functions of the azimuth and elevation angles.

[0051] The second step is to partition the training samples. Specifically, within each partition, the ratio of the maximum to the minimum value of the error covariance matrix corresponding to the incident angle of the training samples is no greater than a preset value. All sample data are non-uniformly classified into several frequency band partitions based on their observation frequency bands. Within each frequency band partition, the associated data are non-uniformly classified into several angle partitions based on their direction-finding range for the incident angle, with overlapping angle boundaries between adjacent partitions.

[0052] To improve the model's direction-finding accuracy, it is necessary to consider the error covariance matrix for different incident angles and construct a weighted MSE loss function (WMSE). Assuming that each training sample is generated independently, the WMSE can be expressed as:

[0053]

[0054] Where, N S It is the sample size, F i (θ, f) represents a function of the incident angle θ and the frequency f. It is its estimated value, where the subscript i indicates the i-th sample, Q F (θ, f) represents the covariance.

[0055] Based on the output defined in the first step, we know that F(θ, f) = χ².

[0056]

[0057] A simple derivation yields the following result. The covariance matrix is:

[0058]

[0059] in, and Let represent the mean square errors in estimating α and β, respectively.

[0060] The above equation shows that the covariance matrix is ​​related to the incident angle θ = [α, β], and its angular information is contained in the training labels of the samples, which is referred to here as the angle-labeled loss function. For traditional phase interferometer direction finding, the direction finding error... and All are wavelength-dependent, therefore here It is also related to the frequency f.

[0061] Because the gradient calculation of the WMSE loss function is quite complex, efficient training based on the WMSE loss function is difficult. To make the training process more efficient, an approximation of the WMSE needs to be considered. Therefore, a partitioning strategy is adopted, which equally weights the loss functions for different combinations of frequency and angle within a certain range, i.e.:

[0062]

[0063] in,

[0064]

[0065]

[0066]

[0067] in, This indicates that χ2 at the incident angle θ l With frequency f l The covariance matrix, and These represent the lower and upper bounds of the frequency within the i-th partition, respectively. and Let represent the lower and upper bounds of the azimuth angle α within the i-th partition, respectively. and Let f, α, and β represent the lower and upper bounds of the pitch angle β in the i-th partition, respectively, and let k denote the k-th sample. The criteria for determining the upper and lower bounds of f, α, and β are to maximize the covariance matrix max{Q} when the error is maximized. θ (θ l f l )} satisfies the entire partition, where Q θ (θ l f l ) is through The covariance with respect to the incident angle θ is obtained. It is also necessary to ensure that Q is calculated in partitioned calculations. sinα (θ l f l ), Q cosα (θ l f l ), Q sinβ (θ l f l ), Q cosβ (θ l f l The corresponding value is approximately the same. Therefore, the following conditions are set:

[0068] max(Q F (θ l f l )) / min(Q F (θ l f l ))≤ζ

[0069] Here, ζ is a threshold obtained through engineering practice or cross-entropy verification. Simulation analysis shows that the model performs well when 10 ≤ ζ ≤ 20. Since Q F (θ l f l The distribution is usually non-uniform, so non-uniformity is also considered in partitioning.

[0070] In practical direction finding engineering, the frequency can usually be directly measured, but the incident angle θ is unknown beforehand and needs to be estimated. Therefore, a two-step partitioning strategy is proposed: (a) all data are classified into several frequency partitions based on their observation frequency band; (b) within each frequency partition, related data are classified into several angle partitions. Furthermore, to avoid misclassification near the partition boundaries, this invention proposes that adjacent angle boundaries overlap by 5°.

[0071] The third step is the construction and training of a deep neural network model. Specifically, a deep neural network for direction finding is constructed and its network parameters are trained based on training samples.

[0072] Based on the first-step output format and the second-step non-uniform partitioning strategy, a generalized deep neural network (DNN) model structure is proposed for the unified model of two-dimensional direction finding. The training phase of the DNN model structure is as follows: Figure 2 As shown. Where Z f This represents the frequency partition label, with a value range of Z. f ∈[1, N f ], N f Indicates the number of frequency partitions; Z l Indicates the angle partition label, with a value range of Z. l ∈[1, N z ], N z Indicates the number of angular partitions; I(Z) l ) indicates the partition label Z l The model input represents the output of the classification network, which is the predicted partition label; This represents the output of the i-th layer regression network.

[0073] Depend on Figure 2 The model structure shown corresponds to the second step of the non-uniform partitioning strategy, and includes N f (N f +1) layer network, a 1-layer classification network is used to divide the model input into N... z A two-layer regression network estimates the direction of travel for each angular region. During training, samples with known region labels are used to train the regression network. During testing, inputs with region label χ = i are used for direction finding in the i-th regression network. Therefore, the labels I(Z) from the training phase are... l =i) changed to I(χ) c =i) is the testing phase.

[0074] The fourth step is direction finding inference. Specifically, based on the generalized deep neural network direction finding model trained in the third step, and combined with the partitioned training samples obtained in the previous steps, a two-dimensional angle direction finding model is obtained through iterative training of the direction finding angle. This model is then used to perform two-dimensional angle direction finding on the actual incident signal.

[0075] Specifically, regarding the actual incident signal Process the input information of the model:

[0076]

[0077] From the two-dimensional angle direction finding model, the azimuth angle α = arctan 2(sinα, cosα) and the elevation angle β = arctan 2(sinβ, cosβ) of the signal are obtained.

[0078] This embodiment discloses a multi-frequency unified direction finding method based on a generalized deep neural network model. By constructing a generalized deep neural network model, training samples are built and partitioned. The deep neural network model is then trained based on these partitioned training samples to obtain a two-dimensional angle (azimuth and elevation) direction finding model. Finally, the azimuth and elevation angles of the incident signal are obtained, achieving high-precision direction finding of the incident signal. Specifically, the training samples are approximately partitioned based on non-uniform frequency bands and non-uniform angles, and the multi-layer network structure of the deep neural network model corresponds to these partitions. This avoids the large amount of correction data required for traditional one-frequency-point-one-correction methods, making the training process more efficient and achieving high-efficiency construction of the direction finding model.

[0079] Example 2

[0080] Another specific embodiment of the present invention discloses an application of a multi-frequency unified direction finding method based on a generalized deep neural network model, which specifically includes the following steps:

[0081] The first step is to collect training samples, specifically considering the 1000MHz to 1050MHz frequency band, with azimuth α ∈ [0°, 360°] and elevation β ∈ [30°, 90°]. A four-element array is used to simulate the direction of arrival of the signal. The array element positions are s1 = [0.061, 0.061, 0]. T s2 = [0.061, -0.061, 0] T s3 = [-0.061, 0.061, 0] T s4 = [-0.061, -0.061, 0] T The unit is meters. The element spacing is set to be less than or equal to the wavelength of the signal being measured.

[0082] Considering the common effects of mutual coupling between array elements and amplitude-phase inconsistency between receiving channels in real-world engineering, the signal received by the antenna array can be represented as follows due to these two types of errors:

[0083] x(t)=C P I R As(t)+n(t)

[0084] Wherein, the mutual coupling matrix C P The elements are set as a random complex matrix based on the element spacing as follows:

[0085]

[0086]

[0087] A is the guiding vector, where, c is the signal propagation speed, and λ is the signal wavelength. s i Let r be the position of the i-th array element, and assume the incoming wave direction is θ = [α, β], r = [cosαcosβ, sinαcosβ, sinβ]. T ;

[0088] n(t) is the noise data vector received by the array;

[0089] This indicates a phase inconsistency between the receiving channels, I 4×4 This represents a 4×4 identity matrix. The specific format is as follows:

[0090]

[0091] Among them, g ri d represents the gain relative to channel 1. ri This represents the path difference of the signal relative to array element 1. Here, in this embodiment, it is defined as:

[0092] g r1 =1, g r2 =0.9, g r3 =1.1, g r4 =0.8

[0093] d r1 =0,d r2 =0.003m, d r3 =0.004m, d r4 =-0.005m

[0094] The second step is to partition the training samples. Specifically, in this embodiment, the regions are divided as follows:

[0095]

[0096]

[0097] The third step involves building and training the deep neural network model. Specifically, based on the angle and frequency partitioning strategy given in the second step (according to the above formula), this embodiment considers using two classification networks and six regression networks for direction finding. Each convolutional filter has a kernel function dimension of 5×1×1, and each convolutional layer has 128 filters. The first fully connected layer in the classification network has an output dimension of 16, and the second fully connected layer has an output dimension of 3 based on each given frequency dimension. Furthermore, the first two fully connected layers of the regression network have output dimensions of 16 and 4, respectively. Here, the output dimension of the second fully connected layer is specified as 4, which is a χ²-based dimension.

[0098] Specifically, the training samples are generated according to the following criteria: a) the sample frequencies are traversed from 1000MHz to 1050MHz in 5MHz increments; b) the elevation angle β is traversed from 30° to 88° in 2° increments, with an additional random perturbation ε∈[0°, 2°]; c) the azimuth angle α is traversed from 0° to 358° in 2° increments, with an additional random perturbation ε2∈[0°, 2°]. Each recursion is repeated 10 times, selecting a 60% probability of obtaining training samples in each recursion (i.e., randomly discarding 40% of the samples), and only 5 frequency points are sampled. The total number of training samples is 30*180*10*5*0.6=162000.

[0099] Specifically, test samples (referred to as test group 1) are generated, with a frequency range from 1000MHz to 1050MHz, a step size of 5MHz, β iterating from 30° to 88° in 2° steps, and α iterating from 0° to 359° in 10° steps. The recursive generation of test samples is repeated 10 times. The total number of test samples is 11 * 30 * 36 * 10 = 118800.

[0100] The fourth step is direction-finding inference. Specifically, based on the generalized deep neural network direction-finding model trained in the third step, a two-dimensional angle direction-finding model is obtained. This model is then used to perform two-dimensional angle direction finding on the actual incident signal. Process the input information of the model:

[0101]

[0102] From the two-dimensional angle direction finding model, the azimuth angle α = arctan 2(sinα, cosα) and the elevation angle β = arctan 2(sinβ, cosβ) of the signal are obtained.

[0103] Figure 3 The direction-finding performance of the proposed method in this embodiment on a complex receiving system is presented. In addition, the direction-finding performance of the traditional MUSIC algorithm and the traditional interferometer are also presented. Figure 3 The results show that for complex receiving systems, MUSIC and traditional interferometers have significantly larger direction-finding errors.

[0104] Another set of test samples was generated for 1007.5MHz, which was not covered by the training set. For the test set with unseen frequencies, the performance is as follows: Figure 4 As shown. Figure 4 The results in Figure 3 The results in the previous method did not show a significant decrease, which demonstrates the good generalization ability of the method, while frequency-specific networks cannot achieve this generalization ability.

[0105] Furthermore, based on the results at 1007.5MHz, the contour lines of the root mean square (RMS) of the direction-finding error at each incident angle are as follows: Figure 5 As shown, this demonstrates that generalized deep neural networks (DNNs) still maintain high direction-finding accuracy for complex receiving systems and support the use of ideal system-based CRLBs to guide angle and frequency partitioning.

[0106] System Implementation Examples

[0107] A multi-frequency unified direction finding system based on a generalized deep neural network model includes: a training sample acquisition module, a training sample partitioning module, a generalized deep neural network model construction and training module, and a two-dimensional angle direction finding module.

[0108] The training sample acquisition module is used to acquire training sample data of the deep neural network. It obtains the input of the training sample based on the feature information and frequency information of the array received signal, and obtains the output of the training sample based on the two-dimensional angle information corresponding to the signal. The input and output of the obtained training sample and the specific acquisition process can be referred to the corresponding description in the above method embodiments.

[0109] The training sample partitioning module is used to partition the training samples based on non-uniform frequency bands and non-uniform angles, including: non-uniformly classifying all sample data into several frequency band partitions based on their observation frequency bands; and within each frequency band partition, non-uniformly classifying the associated data into several angle partitions based on their direction-finding range with respect to the incident angle. The methods and processes for non-uniformly classifying the data into several frequency band partitions and non-uniformly classifying the data into several angle partitions can be referred to the corresponding descriptions in the above method embodiments.

[0110] A generalized deep neural network model construction and training module is used to construct the deep neural network model and train its network parameters based on the partitioned training samples to obtain a two-dimensional angle-finding model; wherein, the deep neural network model includes a multi-layer network structure corresponding to the partition. The structure and training method of the deep neural network model can be referred to the corresponding descriptions in the above method embodiments.

[0111] The two-dimensional angle-finding module uses the two-dimensional angle-finding model to perform two-dimensional angle-finding on the actual incident signal, obtaining the azimuth and elevation angles of the incident signal. The method for obtaining the azimuth and elevation angles of the incident signal can be found in the corresponding description in the above method embodiments.

[0112] Since the multi-frequency unified direction finding system based on the generalized deep neural network model and the multi-frequency unified direction finding method based on the generalized deep neural network model are based on the same inventive concept, they can learn from each other and thus achieve the same technical effect.

[0113] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-frequency unified direction finding method based on a generalized deep neural network model, characterized in that, Includes the following steps: Construct training samples, wherein the input of the training samples is obtained based on the feature information and frequency information of the array received signals, and the output of the training samples is obtained based on the two-dimensional angle information corresponding to the signals; The input includes copying. Frequency information of the array and signals received by the array Normalized upper half-diagonal matrix The characteristic vector formed is represented as a vector. ； in, , Indicates the number of elements in the receiving array. , for the array to receive signals The covariance matrix; The training samples are partitioned based on non-uniform frequency bands and non-uniform angles; The network parameters of the deep neural network model are trained based on the partitioned training samples to obtain a two-dimensional angle-finding model; wherein, the deep neural network model includes a multi-layer network structure corresponding to the partition; Using the aforementioned two-dimensional angle direction finding model, two-dimensional angle direction finding is performed on the incident signal to obtain the azimuth and elevation angles of the incident signal.

2. The multi-frequency unified direction finding method according to claim 1, characterized in that, The feature information includes feature vector information composed of the normalized upper half-diagonal array of the received signals from the array, and the two-dimensional angle information includes vector information composed of the sine and cosine functions of the azimuth and elevation angles.

3. The multi-frequency unified direction finding method according to claim 2, characterized in that, The incident angle of the signal received by the array With frequency The relationship satisfies the function ;in, This indicates the angle of incidence of the signal received by the array. , These represent the azimuth and elevation angles of the signal received by the array, respectively. ; The two-dimensional angle information, .

4. The multi-frequency unified direction finding method according to claim 3, characterized in that, The training samples are partitioned based on non-uniform frequency bands and non-uniform angles, including: All sample data are non-uniformly classified into several frequency band partitions based on their observed frequency bands; Within each frequency band partition, the associated data is non-uniformly classified into several angular partitions based on its direction-finding range for the incident angle.

5. The multi-frequency unified direction finding method according to claim 4, characterized in that, Within each partition, the ratio of the maximum to the minimum value of the error covariance matrix corresponding to the incident angle is no greater than a preset value.

6. The multi-frequency unified direction finding method according to claim 4 or 5, characterized in that, The non-uniform angle partitions have overlapping angle boundaries between adjacent partitions.

7. The multi-frequency unified direction finding method according to claim 4 or 5, characterized in that, The deep neural network model includes components corresponding to the aforementioned frequency band partitions. Layer classification network, in which This indicates the number of frequency band partitions; each layer of the classification network includes several angle partitions corresponding to that frequency band partition. A regression network, This indicates the number of angular partitions in the frequency band partition; wherein, the regression network is used to perform direction finding estimation on the data of the angular partition.

8. A multi-frequency unified direction finding system based on a generalized deep neural network model, characterized in that, include: The training sample acquisition module is used to acquire training sample data of the deep neural network. It obtains the input of the training sample based on the feature information and frequency information of the array received signal, and obtains the output of the training sample based on the two-dimensional angle information corresponding to the signal. The input includes copying. Frequency information of the array and signals received by the array Normalized upper half-diagonal matrix The characteristic vector formed is represented as a vector. ； in, , Indicates the number of elements in the receiving array. , for the array to receive signals The covariance matrix; A training sample partitioning module is used to partition the training samples based on non-uniform frequency bands and non-uniform angles. A generalized deep neural network model construction and training module is used to construct the deep neural network model and train its network parameters based on the partitioned training samples to obtain a two-dimensional angle-finding model; wherein, the deep neural network model includes a multi-layer network structure corresponding to the partition; The two-dimensional angle direction finding module uses the two-dimensional angle direction finding model to perform two-dimensional angle direction finding on the actual incident signal, and obtains the azimuth and elevation angles of the incident signal.

9. The direction-finding system according to claim 8, characterized in that, The training sample partitioning module partitions the training samples based on non-uniform frequency bands and non-uniform angles, including: All sample data are non-uniformly classified into several frequency band partitions based on their observed frequency bands; Within each frequency band partition, the associated data is non-uniformly classified into several angular partitions based on its direction-finding range for the incident angle.

Images