Search

CN-122017730-A - Low-sample-dependence high-precision DOA estimation method for complex electromagnetic environment

CN122017730ACN 122017730 ACN122017730 ACN 122017730ACN-122017730-A

Abstract

The invention relates to the technical field of signal processing, in particular to a high-precision DOA estimation method for a complex electromagnetic environment with low sample dependence. The method constructs the pre-training model with the two-way long-short-term memory network BiLSTM and the residual attention network RAN as cores to extract rich characteristics of signal data, and then realizes effective migration of pre-training knowledge in a snapshot number limited scene through model fine tuning so as to efficiently optimize model parameters. Experimental results show that under the condition of limited snapshot numbers, the DOA estimation accuracy is remarkably improved, and meanwhile, excellent robustness is shown. This suggests that the pre-training model can retain efficient information of the signal data, and that real-time prediction of the array in a limited number of scenarios provides a new solution.

Inventors

  • LIU LEI
  • LI MINGYAN

Assignees

  • 新疆大学

Dates

Publication Date
20260512
Application Date
20260129

Claims (7)

  1. 1. The high-precision DOA estimation method for the complex electromagnetic environment with low sample dependence is characterized by comprising the following steps of: Step 1, preprocessing signal data, extracting real parts and imaginary parts of matrix element signals from noise-containing signals received by a uniform linear array ULA, and constructing an input matrix X IQ with the dimension of 2L multiplied by N, bypassing a traditional covariance matrix calculation flow, wherein L is the snapshot number and N is the matrix element number; Step 2, constructing a pre-training model, designing parallel cavity convolution branches to extract multi-scale features, combining a residual error attention module to strengthen key features, and capturing signal bidirectional time sequence correlation through a bidirectional long-short-term memory network BiLSTM; Step 3, model migration and fine tuning, wherein after a large-scale data set pre-trains a model, parameters are divided into a basic feature set and a task self-adaptive set, the former is frozen to keep general knowledge of a source domain, and the latter is fine tuned to adapt to a target domain small snapshot task; And 4, estimating DOA of the classification task, outputting the existence probability of the angle interval through the full-connection layer and the Sigmoid activation function, and determining the angle of the signal source based on the maximum posterior probability criterion.
  2. 2. The method for estimating high-precision DOA in a complex electromagnetic environment with low sample dependence according to claim 1, wherein in step 1, the method comprises the following steps: assuming a uniform linear array ULA consisting of N omni-directional antennas with an array element spacing d=λ/2,K far-field narrowband signals with wavelength λ (K=1, 2,., K) from different directions = [ Theta 1, theta 2, & gttheta K ] is incident on the ULA, the ULA receives a signal at time t: x(t)=A(θ)s(t)+n(t),t=1,...,L (1) Wherein, each parameter is defined as follows: the vector of the received signal of ULA at the moment t, the dimension is N multiplied by 1, N is the number of array elements, and the vector comprises the received signals of N array elements at the moment t; A (θ) is an array manifold matrix, the dimension is N multiplied by K, N is the number of array elements, K is the number of signal sources, and the array manifold matrix is composed of guide vectors corresponding to the signal sources; θ, a set of directions of arrival DOA of K signal sources, θ= [ θ 1, θ 2, ⋯, θK ], wherein θ k represents the direction of arrival of the kth signal source; s (t), the output vector of K far-field narrowband signal sources at the t moment, the dimension is Kx1, and the output vector comprises the complex amplitude of each signal source at the t moment; N (t) is an additive Gaussian white noise vector at the moment t, the dimension is Nx 1, the mean value is 0, the variance is sigma 2 , and complex Gaussian distribution is obeyed; t is a time sampling index, and the value range is 1 to L; the snapshot number is the total sampling frequency of the signal; the source and array manifold matrices are represented as: s(t)=[s 1 (t),s 2 (t),···,s K (t)] T (2) Wherein, each parameter is defined as follows: s (t) is the output vector of K signal sources at the moment t; s k (t) complex amplitude of the kth signal source at time t, k=1, 2..k, K, representing the intensity and phase information of the signal source at time t; a vector transposition operation, which is to convert a row vector into a column vector; A(θ)=[a(θ 1 ),···,a(θ K )]∈C N×K (3) Wherein, each parameter is defined as follows: A (theta) is an array manifold matrix; a (theta K ) is that the guiding vector corresponding to the kth signal source has dimension of Nx 1 and describes the phase relation of signals received by each array element when the signal source is incident on the ULA; C, a complex domain, wherein matrix elements are represented as complex types; n is the number of array elements of the uniform linear array ULA; K, the number of signal sources; is the ULA relative direction of arrival Is expressed as: ; Wherein, each parameter is defined as follows: a (theta K ) is the steering vector of the kth signal source; d, the array element spacing of ULA, wherein the known condition is d=lambda/2, and lambda is the signal wavelength; Lambda is the wavelength of the incident signal, and is consistent with the definition of a far-field narrowband signal with the wavelength lambda in the known condition; j is an imaginary unit, and meets j 2 = -1; sin (theta K ) is that the sine value of the kth signal source arrival direction theta K is determined by sin (theta K ) because ULA is a linear array; N is the number of array elements; t, vector transposition operation; for the original received signal: ; wherein each row represents a sampled signal on a single element, and the real and imaginary parts of the received signal for each element are then extracted: ; Wherein, each parameter is defined as follows: x IQ , the preprocessed input matrix is used as the input of a subsequent neural network; (. Cndot.) taking the operator of the real part of the complex signal, (X (t)) represents the real part of the received signal vector x (t) at time t; (. Cndot.) taking the operator of the imaginary part of the complex signal, (X (t)) represents the imaginary part of the received signal vector x (t) at time t.
  3. 3. The method for estimating high-precision DOA in a complex electromagnetic environment with low sample dependence according to claim 1, wherein in step 2, the method comprises the following steps: Firstly, extracting local spatial features of signal data by utilizing multi-scale convolution, then, focusing on important features by embedding a residual network enhancement model of the high-efficiency channel attention ECA, reducing model operation amount, inputting the processed features into BiLSTM, further extracting time sequence correlation features of the signals, and finally, carrying out feature mapping through a full-connection layer, and outputting a multi-label classification result by using a Sigmoid activation function.
  4. 4. A low sample-dependent complex electromagnetic environment high-precision DOA estimation method as claimed in claim 3, wherein the method adopts hole convolution to enlarge the receptive field and simultaneously reduce the number of parameters, and uses three kinds of 3 x 1 hole convolution kernels with different expansion rates to extract multi-scale space features in parallel, expressed as: ; Wherein, each parameter is defined as follows: xi, i is the output characteristic diagram of the ith cavity convolution branch, i=1, 2 and 3 correspond to three branches with different expansion rates; DilatedConv (35) carrying out cavity convolution operation, and expanding a receptive field by setting an expansion rate; d, the void ratio of the ith branch is used for expanding the receptive field, and d of the three branches is respectively set to be 1,2 and 3, so that multi-scale feature extraction is ensured; i, branch indexes, wherein the values are 1,2 and 3, and correspond to three parallel cavity convolution branches; Each branch output is sequentially subjected to batch normalization and Mish activation function processing to stabilize training and enhance nonlinear expression capability of the features, and three paths of outputs are respectively X '1, X '2 and X '3, wherein: X′i=Mish(BN(Xi)),i=1,2,3 (8) Wherein, each parameter is defined as follows: x' i is an output characteristic diagram of the ith branch after Batch Normalization (BN) and Mish activation; mish (·) Mish activation function for enhancing the nonlinear expression capacity of a feature; BN (·) batch normalization operation for stabilizing the network training, reducing gradient vanishing risk, normalizing the characteristic dimension of Xi; Xi is the original output of the ith cavity convolution branch; i, branch index; After channel dimension reduction is carried out on each branch feature by adopting 1X 1 convolution, the dimension reduction features of the three branches are cascaded along the channel dimension to form a feature tensor fusing multi-scale information: Xcat=Concat{ X′1, X′2, X′3 } (9) Wherein, each parameter is defined as follows: xcat fusion feature tensors after multi-branch feature cascading; concat {. The feature cascading operation is defined as cascading along the channel dimension, namely, the features of the three branches are spliced in the channel dimension; applying a max-pooling operation on the fused features: Xpool=MaxPool(Xcat) (10) Wherein, each parameter is defined as follows: xpool outputting a characteristic diagram after maximum pooling; MaxPool (DEG), maximum pooling operation; and the compression of the space dimension is carried out, the local significant features are highlighted, and the calculated amount is reduced.
  5. 5. A low sample-dependent complex electromagnetic environment high-precision DOA estimation method as claimed in claim 3, wherein in the depth network design, a residual attention module based on ResNet structures is introduced, the module consists of four progressive residual attention blocks, each residual attention block comprises convolution, batch normalization BN, an activation function and a channel attention mechanism, gradient transfer and feature multiplexing are enhanced through jump connection, the residual attention block adopts a high-efficiency channel attention ECA mechanism, generates attention weights by global average pooling GAP and one-dimensional convolution Conv1D capture channels, and carries out channel weighting by using Sigmoid functions, and the channel numbers of the four residual attention blocks are 64-128-256-512 in sequence by a progressive channel expansion strategy; After the processing of the residual attention module, the network reduces the dimension through global average pooling AvgPool, and reduces the overfitting risk while preserving global semantics, and then the two-way long-short-term memory network BiLSTM further models the sequence characteristics and captures the two-way context information by utilizing the forward and backward LSTM sub-networks; after the characteristic sequence is encoded, the network maps BiLSTM high-dimensional characteristics output by the full-connection FC layer to a target tag space, and in order to meet the multi-source DOA estimation requirement, the final layer adopts a Sigmoid activation function to restrict the output to an interval [0, 1 ]: (11) Wherein, each parameter is defined as follows: L BCE , binary cross entropy BCE loss function, which is used for measuring the difference between the model predicted value and the real label; The true label of the ith angle interval is valued as 1 or 0; the probability of the signal source exists in the ith angle interval predicted by the model, and the value range is [0,1]; model predicts the logarithm of probability that the ith angle belongs to the positive class; the model predicts the logarithm of the probability that the ith angle belongs to the negative class.
  6. 6. The method for estimating high-precision DOA in a complex electromagnetic environment with low sample dependence according to claim 1, wherein in step 3, the method comprises the following steps: The transfer learning process comprises adopting an 8-element uniform linear array ULA, wherein the element distance is half wavelength, the angle coverage range is-60 degrees to 60 degrees, the number of signal sources is set to be 2, the signal data of a source domain and a target domain are generated by SNRs (different signal to noise ratios), the signal to noise ratio range is-10 dB to 20 dB, the increment is 5dB, and the total number of the signal to noise ratio levels is 7, 200,000 samples are arranged on each signal to noise ratio level, and a source domain data set is randomly divided into training, verifying and testing subsets according to the proportion of 8:1:1; The method comprises the steps of adopting a grouping fine tuning strategy to divide network parameters into a basic feature group and a task self-adaptive group, wherein the layers of the basic feature group learn general signal features in source domain training, so that parameters of the basic feature group are frozen to avoid feature degradation, the layers of the task self-adaptive group are directly related to DOA classification tasks, fine tuning is carried out at a higher learning rate to quickly adapt to a target domain, in addition, a double-learning-rate Adam optimizer is adopted, the learning rate of the basic feature group is set to be 0.0001 to ensure parameter stability, and the learning rate of the task self-adaptive group is set to be 0.001 to accelerate model response.
  7. 7. The method for estimating high-precision DOA in a complex electromagnetic environment with low sample dependence according to claim 1, wherein in step 4, the method comprises the following steps: DOA estimation, discretizing DOA range, equally dividing the angle of the space domain where the incident signal may exist to construct an overcomplete space domain grid set, assuming that the range of the arrival direction of the incident signal is the range of the space domain , Angle interval of- Equally spaced apart to obtain an overcomplete airspace grid set: ; Wherein, each parameter is defined as follows: an overcomplete airspace grid set containing all possible direction of arrival candidate values; θmin, lower limit of DOA estimated angle range; θmax, upper limit of DOA estimated angle range; delta theta is the angle discrete interval; the set can be regarded as a set of all possible direction of arrival angles of the incident signal, which contains the number of elements of For multiple sources, the task is ascribed to predicting the most likely set of angles, ={ -Given by maximizing the posterior probability: ; Wherein, each parameter is defined as follows: DOA estimation value of the kth signal source, namely the optimal candidate angle screened out from the grid set G; searching an operator of an angle theta for maximizing a posterior probability P (theta-X) in a grid set G; The posterior probability, P (θ|x), represents the probability that the angle θ is the true direction of arrival given the input matrix X.

Description

Low-sample-dependence high-precision DOA estimation method for complex electromagnetic environment Technical Field The invention relates to the technical field of signal processing, in particular to a low-sample-dependence high-precision DOA estimation method for a complex electromagnetic environment. Background In the field of array signal processing, direction of arrival (DOA) estimation is a core technology for realizing target positioning, interference suppression and beam forming, and is widely applied to the key fields of radar, communication, sonar and the like, and the core target is to accurately estimate the spatial orientation of a signal source through an array receiving signal. However, in practical engineering, complex electromagnetic environment, hardware resource limitation and dynamic scene requirement often lead to challenges of small snapshot number (i.e. limited observation sample) of signal acquisition, which severely restrict accuracy and stability of DOA estimation. The early beam forming-based method is limited by the Rayleigh diffraction limit, super resolution cannot be realized, while the subspace method breaks through the Rayleigh limit, the limited snapshot number can cause the rank deficiency of a sample covariance matrix in a small snapshot scene, so that the characteristic value distribution of a signal and noise subspace is overlapped, the subspace division method based on characteristic value decomposition is invalid, the model is sensitive to priori errors such as array cross coupling and the like, the robustness is reduced, the sparse representation method relies on grid division to introduce quantization errors, and the iterative optimization is easy to be in local optimum in the small snapshot and has high calculation complexity. The data-driven deep learning method improves the adaptability of complex scenes through end-to-end modeling, but the main stream supervised learning network needs large-scale labeled data training, and the cost for acquiring a large number of labeling samples in a small snapshot scene is extremely high and easily leads to model overfitting: (1) The limited observation sample is difficult to support the traditional statistical model assumption and the deep learning parameter training, so that the risk of 'under fitting' and 'over fitting' coexist, and the parameter space exploration efficiency is low. (2) The traditional optimization method is difficult to cover a complete parameter space under small data, the deep learning optimization is easy to fall into a local extremum due to sample noise, and fusion between physical priori and data driving is absent. (3) The existing method depends on an accurate priori or is completely separated from a physical model, and the robustness can not be improved by effectively utilizing the geometric priori constraint of the array during small snapshot. Disclosure of Invention The invention aims to solve the defects in the prior art, and provides a low-sample-dependence high-precision DOA estimation method for a complex electromagnetic environment, which can effectively solve the problems of insufficient extraction of array signal characteristics, large quantity of model-dependent labeling data and insufficient robustness in a small snapshot scene in the prior art, and realize efficient and accurate signal source positioning in the complex electromagnetic environment. In order to achieve the above purpose, the present invention adopts the following technical scheme: a low-sample-dependence high-precision DOA estimation method for a complex electromagnetic environment comprises the following steps: Step 1, preprocessing signal data, extracting real parts and imaginary parts of matrix element signals from noise-containing signals received by a Uniform Linear Array (ULA), and constructing an input matrix X IQ with the dimension of 2L multiplied by N, bypassing a traditional covariance matrix calculation flow, wherein L is the snapshot number and N is the matrix element number; step 2, constructing a pre-training model, designing parallel cavity convolution branches to extract multi-scale features, combining a residual attention module (including an efficient channel attention ECA) to strengthen key features, and capturing signal bidirectional time sequence correlation through a bidirectional long-short-term memory network BiLSTM; step 3, model migration and fine tuning, wherein after a large-scale data set pre-trains a model, parameters are divided into a basic feature group and a task self-adaptive group, the basic feature group and the task self-adaptive group are frozen, the general knowledge of a source domain is reserved in the former, and the fine tuning is suitable for a target domain small snapshot task in the latter; And 4, estimating DOA of the classification task, outputting the existence probability of the angle interval through the full-connection layer and the Sigmoid activation function, and determ