CN-118861521-B - Cross-platform complex domain feature enhanced coherent super-resolution DOA estimation method

Abstract

The invention relates to the technical field of DOA (direction of arrival) estimation, and particularly discloses a cross-platform complex domain feature enhanced coherent super-resolution DOA estimation method, which comprises the following steps of firstly, constructing a complex domain CVSIMO learning model; preprocessing the sampled data, CVSIMO learning forward propagation of the model, CVSIMO learning backward propagation of the model and parameter optimization, data reconstruction and signal separation, cross-platform super-resolution DOA estimation, and finally performing a model simulation experiment. The cross-platform complex domain characteristic enhanced coherent super-resolution DOA estimation method can realize characteristic mining of a plurality of point source signals, converts a multi-source estimation problem into a single-point source DOA estimation problem by using a complex domain neural network model, enhances the characteristics of the coherent signals by using the advantages of the complex domain neural network, improves the performance and the precision of a super-resolution DOA estimation algorithm, can also realize data separation of a plurality of coherent signals and cross-platform DOA estimation, and realizes real angle estimation by ingenious characteristic solution.

Inventors

• XIANG HOUHONG
• LI YUXI
• CHEN YUFENG

Assignees

• 合肥工业大学
• 西安电子科技大学杭州研究院

Dates

Publication Date

20260508

Application Date

20240705

Claims (4)

1. 1. A cross-platform complex domain feature enhanced coherent super-resolution DOA estimation method is characterized by comprising the following steps: s1, constructing a complex domain CVSIMO learning model, wherein the complex domain CVSIMO learning model comprises a complex value input layer, four hidden layers and K independent complex value output layers; s2, preprocessing the sampled data; s3, CVSIMO, learning forward propagation of a model, wherein the specific process is as follows: the hidden layer and the output layer are all fully connected layers, assuming that the input of the first layer is The complex weight matrix is The hidden layer is activated by complex number The output after activation is expressed as: (3) Wherein the method comprises the steps of , And Respectively solving real and imaginary components of complex numbers; For an L-layer fully-connected neural network, the output layer consists of K independent fully-connected layers, and the complex weight matrix of the kth output layer is assumed to be The output of the kth output layer is expressed as: (4) Thus, the forward propagation of CVSIMO learning models is expressed as: (5); Wherein, the Represents the t-th sample data; S4, performing back propagation and parameter optimization of a CVSIMO learning model; s5, data reconstruction and signal separation; S6, cross-platform super-resolution DOA estimation, wherein the array data separated by CVSIMO model only comprises one signal source, and the separated signal is subjected to phase enhancement through DNN network by using complex domain characteristic enhancement coherent super-resolution DOA estimation technology, so that noise interference is removed, and DOA estimation is performed; the DOA estimate for different platforms, i.e. when the frequency changes, is corrected as follows: Assume that the training set samples have a wavelength of The angle range of the coherent signal is The test set is cross-platform sampling data, and the wavelength of the test set sample is Unlike the training set samples, the angular range of the coherent signal is And (2) and ; The phase characteristics of the training set and test set samples are expressed as: (19) Test set By phase characteristics of (2) And (3) performing equivalent replacement to obtain: (20) Order the Wherein Phase characteristics Expressed as: (21) the above variants and Consistent, this means that changes in cross-platform wavelength will cause changes in the measured angle of the DOA, and satisfy the following relationship: (22) thus, the DOA under cross-platform conditions is estimated by: (23) analytical formula (23), discussion And Due to the range of values of (2) The method comprises the following steps of: (24) Thus, the first and second substrates are bonded together, The range of the values is as follows: (25) learning data separated by network CVSIMO Is achieved by classical DBF, MUSIC and ML algorithms: the measured angle is recorded as From the correction principle, a wavelength of the light is obtained according to the formula (23) The cross-platform DOA estimation of (2) is expressed as: (26) Wherein, the For the goniometric results of the training set samples, In order to obtain the angle measurement result expected to be output in a cross-platform manner, the forward propagation process of CVSIMO models is utilized to obtain the output of K separated signal data, and the K independent outputs are that , ; S7, performing a model simulation experiment.
2. 2. The cross-platform complex domain feature enhanced coherent super-resolution DOA estimation method as recited in claim 1, wherein in step S2, the specific process of preprocessing the sampled data is as follows: assume that the training set consists of N samples, denoted as: Wherein Representing the data of the t-th sample, Representing the kth marker data, complex-valued gaussian normalization is performed on the real and imaginary parts of the N sample features, as follows: (1) Wherein, the And Respectively solving real and imaginary components of complex numbers; And Respectively representing the statistical average value of the real part and the imaginary part of the training set; And The statistical standard deviation, which represents the real and imaginary parts of the training set respectively, is expressed as: (2)。
3. 3. The cross-platform complex domain feature enhanced coherent super-resolution DOA estimation method as recited in claim 1, wherein in step S4, the loss of CVSIMO learning model is expressed as: (6) Let the parameter update formula of the complex weight matrix W of the first layer p row q column be: (7) Wherein the method comprises the steps of Representing the learning rate, the chain law results in: (8) (9) Wherein, the Representing vectors Is selected from the group consisting of the p-th element of (c), Representing vectors And (3) bringing formula (9) into formula (8): (10) wherein, an error term is set Thus, formula (7) is represented as: (11) When l=l, error term Expressed as: (12) when l=l-1, the mth element of the error term is expressed as: (13) Wherein, the (14) Substituting formula (14) into formula (13): (15) combining the formula (12) and the formula (15), and summarizing the error term obtained by the first layer as follows: (16) And then, adopting an adaptive moment estimation Adam algorithm to optimize weights obtained by CVSIMO learning models until learning errors are converged.
4. 4. The method for cross-platform complex domain feature enhanced coherent super-resolution DOA estimation as claimed in claim 3, wherein in step S5, after optimizing CVSIMO model learning model, the coherent signal of given data is separated by CVSIMO model by assuming arbitrary test data as And is also provided with The test sample is subjected to Gaussian normalization treatment by using the mean value and the standard deviation in the formula (2), and the treated sample is: (17) Data to be normalized Inputting the K signals into a trained SIMO learning model, and obtaining the output of the K separated signal data through the forward propagation process of the model, wherein the K independent outputs are as follows: (18) Wherein, the And Representing separate signal data and noise data.

Description

Cross-platform complex domain feature enhanced coherent super-resolution DOA estimation method Technical Field The invention relates to the technical field of direction of arrival estimation, in particular to a cross-platform complex domain feature enhanced coherent super-resolution DOA estimation method. Background The direction of arrival (Directionofarrival, DOA) refers to the direction of arrival of the spatial signals (the angle of arrival of each signal at a reference element of the array, simply referred to as the direction of arrival). The method is an important concept in the theory of spatial spectrum estimation, and the spatial spectrum estimation is one of two research directions of array signal processing. Typical DOA estimation methods include monopulse goniometry (amplitude-versus-monopulse, phase-versus-monopulse) and array super-resolution goniometry. The single-pulse angle measurement technology has the advantages of small calculated amount, strong real-time performance, low precision, inapplicability to multi-source DOA estimation and more suitability to the problem of angle estimation of a coherent source. Currently, the problem of DOA estimation of coherent signals is mainly to adopt an array signal processing method. For DOA estimation of coherent sources, classical super-resolution algorithms include multiple signal classification algorithms (MultipleSignalClassification, MUSIC) and maximum likelihood algorithms (MaximumLikelihood, ML). The ML algorithm knows the statistical distribution characteristics of noise, and the algorithm performance is better. For single source signals, an ML algorithm can be directly adopted for angle estimation, but if the problem of multi-source estimation is involved, the ML algorithm involves the problem of projection matrix calculation of multiple dimensions, the calculated amount is large, the real-time performance is poor, an optimization algorithm of multi-dimensional alternate iteration can be adopted, the calculation complexity is reduced, but the alternate iteration process cannot always be optimized to a globally optimal solution, the performance is slightly lost, if a MUSIC algorithm is adopted, solution coherence processing is needed first to recover the rank of a covariance matrix, and a classical rank recovery means is realized by adopting a subarray smoothing method, namely a space smoothing MUSIC algorithm (SpatialSmoothingMUSIC, SSMUSIC). Obviously, the smoothing treatment causes array aperture loss and angle measurement performance reduction, and in addition, the MUSIC algorithm involves eigenvalue decomposition operation, so that the calculation amount is large. For the super-resolution algorithms, under the condition of an ideal far-field plane wave model, if the number of snapshots and the Signal-to-noise ratio (SNR) are high, good estimation performance can be obtained, but when the actual received Signal model does not meet the requirement of the far-field plane wave model, the matching degree is high, and if the number of snapshots is less, the performance of the super-resolution algorithm is greatly reduced. Artificial intelligence has become one of the hot topics of today's society, mainly due to the rapid development of deep learning, while neural networks (NeuralNetworks) are important models of deep learning, mainly applied to the field of computer vision, and have achieved many excellent results in this field. The existing super-resolution algorithm based on deep learning has high performance, but only utilizes the spatial domain characteristics of the target, namely the spatial sparsity, and learns the mapping relation between the array received data and the target elevation angle through a deep neural network, so that the data characteristics of the array received data per se are lack to be analyzed, and the requirement on the array received data is high in snapshot number. Most neural networks, however, are based on real representations, and rarely on complex representations. Recent studies have shown that networks based on complex representation have better expressive power than networks based on real representation. In 2010, haenschR and HellwichO apply the complex neural network to the full-polarization synthetic aperture radar image for classification, and experimental results show that the complex performance is superior to that of the complex full-connection network. In 2016, DANIHELKA et al proposed a complex long and short-term memory network that achieved better performance than a real long and short-term memory network. In 2017 Popa et al proposed a multi-layer complex neural network and achieved better performance than real neural networks on MNIST and CIFAR-10 datasets. In 2018, trabelsi et al proposed a Complex residual network (Complex-valuedResidualNetworks, CResNet) with a core module of cResNet blocks. In 2019, cheng Huitao applied a complex neural network to reconstruction of nuclear magnetic resonance im