End-to-end DOA estimation via self-supervised cascaded DNNs with array errors mitigation Authors: Shuang Wu, Ye Yuan, Hongyu Pu, Tai Fu Published: 2026 ABSTRACT In this paper, we present a novel cascaded deep neural network architecture for high-precision direction-of-arrival (DOA) estimation in the presence of multiple array errors. The proposed framework addresses complex nonlinear distortions through an end-to-end feedforward computing strategy, eliminating the need for explicit error modeling. It consists of two key components trained in parallel: (1) four-layer adaptive spatial filters optimized using spatial-domain supervision on covariance matrices, and (2) an eight-layer MUSIC-NET that maps synthetic covariance matrices to MUSIC-like spectrum via self-supervised learning. The output of spatial filters serves as the input for MUSIC-NET, forming a two-stage architecture that enables adaptive array error mitigation during DOA estimation. Simulation results demonstrate that the proposed architecture exhibits superior robustness and generalization capabilities across unseen scenarios, outperforming state-of-the-art methods in low signal-to-noise ratio and limited snapshot conditions. Introduction Direction-of-arrival (DOA) estimation is widely used in radar, communication, sonar, and related fields1. Over the past few decades, numerous methods have been developed to address the DOA estimation challenge with super-resolution, high accuracy, and adaptability in low SNR or small snapshot application scenarios2–8. For instance, the multiple signal classification (MUSIC)2 and estimation of signal parameters via rotational invariance technique (ESPRIT)3 methods, which are based on the subspace orthogonality assumption, have achieved super-resolution direction finding at low SNR. Sparsity-inducing methods9,10, based on the over-complete steering matrix, accurately find direction using a single snapshot by reconstructing the spatial spectrum into a linear weighted representation of an over-complete basis. However, low-cost direction-finding systems often encounter complex array errors such as antenna position perturbations, mutual coupling effects, and gain/phase inconsistencies. These errors make it challenging to establish a mathematical model for the mapping from array output to signal direction, leading to a significant degradation in the performance of model-based methods11. Recent research has introduced deep learning techniques to tackle the DOA estimation challenge12–19. These innovative methods utilize a large training dataset with DOA labels to learn the probability distribution of signal directions for array outputs. These methods do not rely on the ideal array assumption required by classical model-based methods. In particular, the works20–23 have conducted in-depth investigations on the DOA estimation problem in the presence of complex array imperfections. Refs.20,21 propose a supervised pre-training framework for DOA estimation, which addresses array errors using a pre-trained autoencoder and then leverages a substantial amount of labeled data to establish the mapping of array output to signal direction. Refs.22,23 utilize a significant amount of labeled data to model the signal subspace and address array errors, subsequently enabling accurate DOA estimation through spatial spectrum calculation. In practice, obtaining substantial training data with manual DOA labels is challenging24,25. This letter proposes a cascaded framework that combines spatial filters with MUSIC-NET to separate array error mitigation from DOA estimation. The proposed covariance-aware spatial filters reconstructs the covariance matrix to correct array errors, while MUSIC-NET maps this matrix to the MUSIC spectrum. This approach simplifies the training of the spatial filters using limited array-measured data, while MUSIC-NET can be trained with synthesized data based on an idealized perfect array. We introduce a MUSIC fitting loss function that employs self-supervised learning on synthetic covariance matrices to generate MUSIC-like spectrum, facilitating batch-parallel computation of labels on the GPU. Simulation results demonstrate the superiority of our method in array error mitigation and DOA estimation compared to state-of-the-art methods. Problem formulation Assuming that an ideal uniform linear array (ULA) of M antennas spaced at half wavelength d receives K far-field narrowband signals, the incident directions of the signals are \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _k,k=1,...,K$$\end{document}. The array output, sampled uniformly at N instants, is represented by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\boldsymbol{X}} =[{\boldsymbol{x}}({t_1}),{\boldsymbol{x}}({t_2}),...,{\boldsymbol{x}}({t_N})] \in \mathbb {C}{^{M \times N}}$$\end{document}, and the array observation at sampling time \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t_n$$\end{document} is1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} {\boldsymbol{x}}({t_n}) = {\boldsymbol{As}}({t_n}) + {\boldsymbol{n}}({t_n}) = \sum \nolimits _{k = 1}^K {{\boldsymbol{a}}({\theta _k}){s_k}({t_n}) + {\boldsymbol{n}}({t_n})}, \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\boldsymbol{a}}({\theta _k}) = {[1,...,{e^{ - j2\pi (M - 1)d\sin ({\theta _k})/\lambda }}]^\mathrm{{T}}}$$\end{document} represents the steering vector of the kth signal and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\boldsymbol{n}}({t_n}) = {\left[ {{n_1}({t_n}),{n_2}({t_n}), \ldots ,{n_M}({t_n})} \right] ^\mathrm{{T}}}$$\end{document} denotes the sampling vector of additive noise with a zero-mean Gaussian distribution. When considering a finite-length sampling of N instants, the array covariance matrix is denoted by2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} {\boldsymbol{R}} = \frac{1}{N}\sum \nolimits _{i = 1}^N {{\boldsymbol{x}}({t_i}){\boldsymbol{x}}{{({t_i})}^\mathrm{{H}}}} = {{\boldsymbol{U}}_\mathrm{{s}}}{{\boldsymbol{\Sigma }}_\mathrm{{s}}}{\boldsymbol{U}}_\mathrm{{s}}^\mathrm{{H}} + {{\boldsymbol{U}}_\mathrm{{n}}}{{\boldsymbol{\Sigma }}_\mathrm{{n}}}{\boldsymbol{U}}_\mathrm{{n}}^\mathrm{{H}}, \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\boldsymbol{\Sigma }}_\mathrm{{s}}} \in \mathbb {R}{^{K \times K}}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\boldsymbol{\Sigma }}_\mathrm{{n}}} \in \mathbb {R}{^{(M - K) \times (M - K)}}$$\end{document} are the signal eigensubspace corresponding to K large eigenvalues and the noise eigensubspace corresponding to M-K small eigenvalues, respectively. As \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$N \rightarrow \infty$$\end{document}, the signal subspace becomes completely orthogonal to the subspace spanned by the steering vector, i.e3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} {{\boldsymbol{a}}^\mathrm{{H}}}({\theta _k}){{{\boldsymbol{ U}}}_\mathrm{{n}}} \approx 0,k = 1,...,K. \end{aligned}$$\end{document}In many academic studies on the DOA estimation, the MUSIC spectrum function is denoted as4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} {{{\boldsymbol{ p}}}_{\mathrm{{MU}}}}(\theta ) = \frac{1}{{{{\boldsymbol{a}}^\mathrm{{H}}}(\theta ){{{\boldsymbol{ U}}}_\mathrm{{n}}}{\boldsymbol{ U}}_\mathrm{{n}}^\mathrm{{H}}{\boldsymbol{a}}(\theta )}}. \end{aligned}$$\end{document}The array model is typically assumed to be ideal, with no mutual coupling or position deviation of the antennas, and perfectly consistent amplitude and phase gains. However, in practical applications, various array errors result in the array steering vector, denoted as5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \begin{array}{l} {{{\boldsymbol{a}}_e}(\theta ) = {\boldsymbol{G}}(\theta ) \times {\boldsymbol{C}}(\theta ) \times {\boldsymbol{a}}(\theta ,{\boldsymbol{l}})}\\ {{\boldsymbol{G}}(\theta ) = diag({{\boldsymbol{g}}_a}(\theta )\cdot {e^{j{{\boldsymbol{g}}_p}(\theta )}})}\\ {{\boldsymbol{C}}(\theta ) = \mathcal{T}({\boldsymbol{c}}(\theta ))} \end{array}, \end{aligned}$$\end{document}deviating from the ideal \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta \mapsto {\boldsymbol{a}}(\theta )$$\end{document}. Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{\boldsymbol{g}}_a}(\theta )}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\boldsymbol{g}}_p}(\theta )$$\end{document} represent the gain and phase response functions with respect to θ, respectively. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\boldsymbol{c}}(\theta )$$\end{document} denotes the mutual coupling function with respect to θ, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal{T}(\bullet )$$\end{document} is the Toeplitz operator that constructs a Toeplitz matrix using a column vector. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\boldsymbol{l}}$$\end{document} represents the array antenna position deviation vector. Furthermore, the array output sampling and its covariance matrix can be formulated as6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \begin{array}{l} {{\boldsymbol{x}}_e}({t_n}) = \sum \nolimits _{k = 1}^K {{{\boldsymbol{a}}_e}({\theta _k}){s_k}({t_n})} + {\boldsymbol{n}}({t_n})\\ {{\boldsymbol{R}}_e} = \frac{1}{N}\sum \nolimits _{i = 1}^N {{{\boldsymbol{x}}_e}({t_i}){{\boldsymbol{x}}_e}{{({t_i})}^\mathrm{{H}}}} \end{array}. \end{aligned}$$\end{document} Proposed estimator framework As shown in Fig. 1, the proposed DOA estimation framework comprises spatial filters and MUSIC-NET. The solid arrows represent feedforward computation, while the dashed arrows represent the loss function gradient backpropagation. The spatial filters and MUSIC-NET are two separate DNNs, and we will offer a detailed introduction to their mathematical models and training methods.Fig. 1The proposed estimator framework. The proposed estimator framework. Spatial filters The spatial filters consist of 4 dense layers. The output of neurons in each layer undergoes non-linear processing through the Tanh activation function, defined as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathop {\textrm{Tanh}}\nolimits } (x) = ({e^x} - {e^{ - x}})/({e^x} + {e^{ - x}})$$\end{document}. The input to the spatial filters is the column-wise concatenation of the real and imaginary parts of the covariance vector, expressed as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\boldsymbol{y}}_e} = {[\textrm{Re}({\boldsymbol{r}}_e^\mathrm{{T}}), \textrm{Im}({\boldsymbol{r}}_e^\mathrm{{T}})]^\mathrm{{T}}} \in \mathbb {R}{^{H \times 1}}$$\end{document}, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\boldsymbol{r}}_e} = \textrm{Vec}({{\boldsymbol{R}}_e}) \in \mathbb {C}{^{{M^2} \times 1}}$$\end{document}. Here, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{Re}( \bullet )$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{Im}( \bullet )$$\end{document} denote the operations for obtaining the real and imaginary parts, respectively. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{Vec}( \bullet )$$\end{document} represents the matrix vectorization operation, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H = 2M^2$$\end{document} represents the spatial filters’ input dimensions. The output of the spatial filters is the concatenation of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N_{\mathrm{{SF}}}}$$\end{document} spatial subregion filtering results, denoted as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\boldsymbol{\hat{y}}} = {[{\boldsymbol{\hat{y}}}_1^\mathrm{{T}}, \cdots ,{\boldsymbol{\hat{y}}}_{{N_{\mathrm{{SF}}}}}^\mathrm{{T}}]^\mathrm{{T}}} \in \mathbb {R}{^{(H*{N_{\mathrm{{SF}}}}) \times 1}}$$\end{document}, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\boldsymbol{\hat{y}}}_i} \in \mathbb {R}{^{H \times 1}}(i = 1,...,{N_{\mathrm{{SF}}}})$$\end{document} represents the filtering result of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\boldsymbol{y}}_e}$$\end{document} by the ith spatial filter. Consider the potential direction of the signal is \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left[ {{\theta _{\min }},{\theta _{\max }}} \right]$$\end{document}, which is uniformly sampled with an interval of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta \theta$$\end{document}. The set of directions of the incident signal can be represented as7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} {\boldsymbol{\Theta }_{\mathrm{{SF}}}} = \{ {\theta _i}|{\theta _i} = {\theta _{\min }} + (i - 1)\Delta \theta ,i = 1,...,{N_\mathrm{{\theta }}}\}, \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N_\mathrm{{\theta }}} = ({\theta _{\max }} - {\theta _{\min }})/\Delta \theta$$\end{document}. For each \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\theta _i} \in {\Theta _{\mathrm{{SF}}}}$$\end{document}, we collect \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\boldsymbol{y}}^i_e}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\boldsymbol{y}}^i$$\end{document} as the input of the spatial filters and the desired output, respectively, and then build the training set8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} {\boldsymbol{\Omega } _{\mathrm{{SF}}}} = \{ ({\boldsymbol{y}}_e^i,{{\boldsymbol{y}}^i})|i = 1,...,{N_\theta }\}. \end{aligned}$$\end{document}We set \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N_{\mathrm{{SF}}}}$$\end{document} spatial subregions in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left[ {{\theta _{\min }},{\theta _{\max }}} \right]$$\end{document}, where each spatial subregion has a length of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$L = {N_\mathrm{{\theta }}}/{N_{\mathrm{{SF}}}}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$l = \big \lceil {i/L} \big \rceil$$\end{document} (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\big \lceil \bullet \big \rceil$$\end{document} is the upward rounding operation) denotes that \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _i$$\end{document} is in the lth spatial subregion. For \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$j = 1,...,{N_{\mathrm{{SF}}}}$$\end{document}, if j=l,9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \begin{array}{l} {{{\boldsymbol{y}}_j}({\theta _i})} = [\textrm{Re}({\boldsymbol{r}}{({\theta _i})^\mathrm{{T}}}),{[\textrm{Im}{({\boldsymbol{r}}({\theta _i}))^\mathrm{{T}}}]^\mathrm{{T}}}\\ {\boldsymbol{r}}({\theta _i}) = \textrm{Vec}({\boldsymbol{R}}({\theta _i})) \end{array}, \end{aligned}$$\end{document}otherwise \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\boldsymbol{y}}_j}({\theta _i}) = {\boldsymbol{0}}$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\boldsymbol{y}}^i = {[{\boldsymbol{y}}_1^\mathrm{{T}}({\theta _i}), \cdots ,{\boldsymbol{y}}_{{N_{\mathrm{{SF}}}}}^\mathrm{{T}}({\theta _i})]^\mathrm{{T}}}$$\end{document}. In addition, we define the feedforward calculation of the spatial filters as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\boldsymbol{\hat{y}}} = {f_{\mathrm{{SF}}}}({{\boldsymbol{\omega }}_{\mathrm{{SF}}}},{{\boldsymbol{y}}_e})$$\end{document}, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${f_{\mathrm{{SF}}}}( \bullet )$$\end{document} represents the spatial filters model parameterized using the parameter set \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\boldsymbol{\omega }}_{\mathrm{{SF}}}}$$\end{document}. The training of spatial filters can be formulated as a parameter optimization problem under the training set, represented as10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} {\boldsymbol{\omega }}_{\mathrm{{SF}}}^ * = \mathop {\arg \min }\limits _{{{\boldsymbol{\omega }}_{\mathrm{{SF}}}}} ({\mathcal{L}_{\mathrm{{SF}}}}({\boldsymbol{y}},{\boldsymbol{\hat{y}}})). \end{aligned}$$\end{document}Here, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathcal{L}_{\mathrm{{SF}}}}({\boldsymbol{y}},{\boldsymbol{\hat{y}}}) = \frac{1}{{H*{N_{\mathrm{{SF}}}}}}\sum \limits _{i = 1}^{H*{N_{\mathrm{{SF}}}}} { {{{\left( y(i) - \hat{y}(i)\right) }^2}} }$$\end{document} denotes the loss function of the spatial filters. The gradient backpropagation of the loss function is represented as11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} {\boldsymbol{\omega }}_{\mathrm{{SF}}}^{new} = {\boldsymbol{\omega }}_{\mathrm{{SF}}}^{old} - \eta {\nabla _{\boldsymbol{\omega }}}{\mathcal{L}_{\mathrm{{SF}}}}({\boldsymbol{y}},{\boldsymbol{\hat{y}}}), \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\boldsymbol{\omega }}_{\mathrm{{SF}}}^{old}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\boldsymbol{\omega }}_{\mathrm{{SF}}}^{new}$$\end{document} represent the current parameter set and the parameter set updated after one iteration, respectively. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\nabla _{\boldsymbol{\omega }}}$$\end{document} represents the gradient operation concerning the parameter set \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\boldsymbol{\omega }}$$\end{document}, and η represents the learning rate. For the output of spatial filters, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\boldsymbol{\hat{y}}} = {[{\boldsymbol{\hat{y}}}_1^\mathrm{{T}},...,{\boldsymbol{\hat{y}}}_{{N_{\mathrm{{SF}}}}}^\mathrm{{T}}]^\mathrm{{T}}}$$\end{document}, we reshape it into \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{{[}}{{{\boldsymbol{\hat{R}}}}_1},...,{{{\boldsymbol{\hat{R}}}}_{{N_{\mathrm{{SF}}}}}}\mathrm{{]}}^\mathrm{{T}}}$$\end{document}, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{\boldsymbol{\hat{R}}}}_i},i = 1,...,{N_{\mathrm{{SF}}}}$$\end{document} denotes the filtering result of the ith spatial subregion. Finally, we calculate an estimate of the covariance matrix by summing over all filter outputs as12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} {\boldsymbol{\hat{R}}} = \sum \nolimits _{i = 1}^{{N_{\mathrm{{SF}}}}} {{{{\boldsymbol{\hat{R}}}}_i}}. \end{aligned}$$\end{document} MUSIC-NET The proposed MUSIC-NET consists of 4 convolutional layers and 4 dense layers, with each layer’s output being non-linearly processed by the ReLU activation function (ReLU(x) = 0, x łe 0 or x, x > 0). The input of MUSIC-NET comprises two channels, namely \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{Re}}({\boldsymbol{\hat{R}}})$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{Im}}({\boldsymbol{\hat{R}}})$$\end{document}, while the output is denoted as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{\boldsymbol{\hat{p}}}}_{{\text {MU}}}} \in {\mathbb {R}^{{N_o} \times 1}}$$\end{document}. When sampling \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left[ {{\theta _{\min }},{\theta _{\max }}} \right]$$\end{document} uniformly by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta \vartheta$$\end{document}, the output dimension of MUSIC-NET is \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N_o} = ({\theta _{\max }} - {\theta _{\min }})/\Delta \vartheta$$\end{document}. First, we generate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N_{\mathrm{{train}}}}$$\end{document} training data using (2) and create the training set \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\boldsymbol{\Omega }}_{\mathrm{{MU}}}} = \{ {{\boldsymbol{R}}^i}|i = 1,...,{N_{\mathrm{{train}}}}\}$$\end{document}. During the training process, we perform the calculation of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\boldsymbol{p}}_{\mathrm{{MU}}}}$$\end{document} in parallel on the GPU following (4). Additionally, we define the feedforward calculation of the MUSIC-NET as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{\boldsymbol{\hat{p}}}}_{\mathrm{{MU}}}} = {f_{\mathrm{{MU}}}}({{\boldsymbol{\omega }}_{\mathrm{{MU}}}},{\boldsymbol{R}})$$\end{document}, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${f_{\mathrm{{MU}}}}( \bullet )$$\end{document} represents the MUSIC-NET parameterized using parameter set \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\boldsymbol{\omega }}_{\mathrm{{MU}}}}$$\end{document}. The MUSIC-NET training can be described as13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} {\boldsymbol{\omega }}_{\mathrm{{MU}}}^ * = \mathop {\arg \min }\limits _{{{\boldsymbol{\omega }}_{\mathrm{{MU}}}}} ({\mathcal{L}_{\mathrm{{MU}}}}({{\boldsymbol{p}}_{\mathrm{{MU}}}},{{{\boldsymbol{\hat{p}}}}_{\mathrm{{MU}}}})), \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathcal{L}_{\mathrm{{MU}}}}({{\boldsymbol{p}}_{\mathrm{{MU}}}},{{{\boldsymbol{\hat{p}}}}_{\mathrm{{MU}}}}) = \frac{1}{{{N_o}}}\sum \limits _{i = 1}^{{N_o}} { {{{\left( {p_{\mathrm{{MU}}}}(i) - {{\hat{p}}_{\mathrm{{MU}}}}(i)\right) }^2}} }$$\end{document} denotes the loss function of the MUSIC-NET. The gradient backpropagation of the loss function can be formulated as14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} {\boldsymbol{\omega }}_{\mathrm{{MU}}}^{new} = {\boldsymbol{\omega }}_{\mathrm{{MU}}}^{old} - \mu {\nabla _{\boldsymbol{\omega }}}{\mathcal{L}_{\mathrm{{MU}}}}({{\boldsymbol{p}}_{\mathrm{{MU}}}},{{{\boldsymbol{\hat{p}}}}_{\mathrm{{MU}}}}), \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\boldsymbol{\omega }}_{\mathrm{{MU}}}^{old}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\boldsymbol{\omega }}_{\mathrm{{MU}}}^{new}$$\end{document} represent the current parameter set and the parameter set updated after one iteration respectively. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\nabla _{\boldsymbol{\omega }}}$$\end{document} represents the gradient operation concerning the parameter set \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\boldsymbol{\omega }}$$\end{document}, and μ represents the learning rate. Training paradigm The existing methods adopt a unified network framework for array correction and DOA estimation and complete the mapping of the covariance matrix to the spatial spectrum or subspace through supervised learning. However, training such a network usually requires a large amount of training data (e.g., the pairing of the covariance matrix and the corresponding one-hot DOA labels). It seems unrealistic to obtain a large number of covariance matrices in practice and manually label them, especially since each array antenna has a unique array error. We propose a framework characterized by a cascaded structure, wherein the innovation resides in the decoupling of array error correction and DOA estimation. During the training phase, both the spatial filters and MUSIC-NET are trained concurrently using their respective datasets. For the spatial filters component, it is relatively straightforward to obtain a modest amount of array observation data in the presence of array errors. Our experimental results indicate that satisfactory outcomes are trained with as few as 121 covariance matrix samples under specified array error conditions. Concerning MUSIC-NET, the focus is solely on mapping the array covariance matrix to MUSIC spectrum. The training data for this component can be synthesized based on the idealized assumption of a perfect array, utilizing a self-supervised learning paradigm that enables batch-parallel computation of MUSIC spectrum labels on a GPU during training. The details of the self-supervised training paradigm of MUSIC-NET are as follows: Step 1 Perform offline pre-computation of the array manifold matrix and its transpose: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left\{ {\begin{array}{*{20}{l}} {\overline{\boldsymbol{A}} = [{\boldsymbol{a}}({\theta _1}),...,{\boldsymbol{a}}({\theta _{{N_\mathrm{{o}}}}})] \in \mathbb {C}{^{M \times {N_\mathrm{{o}}}}}}\\ {{{\overline{\boldsymbol{A}} }^\mathrm{{H}}} = [{{\boldsymbol{a}}^\mathrm{{H}}}({\theta _1});...;{{\boldsymbol{a}}^\mathrm{{H}}}({\theta _{{N_\mathrm{{o}}}}})] \in \mathbb {C}{^{{N_\mathrm{{o}}} \times M}}} \end{array}} \right. .$$\end{document} Step 2 Taking advantage of the parallel computing capability of GPU, the covariance matrix tensor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathcal {R} } = [{{\boldsymbol{R}}^1},...,{{\boldsymbol{R}}^{{N_{\mathrm{{batch}}}}}}] \in \mathbb {C}{^{{N_{\mathrm{{batch}}}} \times M \times M}}$$\end{document} of batch size \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{N_{\mathrm{{batch}}}}}$$\end{document} is decomposed by eigenvalue to obtain the feature subspace tensor.:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal{R} \rightarrow \mathcal{U} \in \mathbb {C}{^{{N_{\mathrm{{batch}}}} \times M \times M}}$$\end{document}. Step 3 Extract the noise subspace tensor:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathcal{U}_n} \in \mathbb {C}{^{{N_{\mathrm{{batch}}}} \times M \times K}}$$\end{document}. Step 4 Parallel computing tensor:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathcal{G}_n} \in \mathbb {C}{^{{N_{\mathrm{{batch}}}} \times M \times M}}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal{G}_n^i = \mathcal{U}_n^i{(\mathcal{U}_n^i)^\mathrm{{H}}} \in \mathbb {C}{^{M \times M}}$$\end{document}. Step 5 Calculate tensors \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{\overline{\boldsymbol{A}} }^\mathrm{{H}}}}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathcal{G}_n}$$\end{document} by second-dimensional multiplication \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${ \times _2}$$\end{document}:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal{T} = {{\overline{\boldsymbol{A}} }^\mathrm{{H}}}{ \times _2}{\mathcal{G}_n} \in \mathbb {C}{^{{N_\mathrm{{o}}} \times {N_{\mathrm{{batch}}}} \times M}}$$\end{document}. Step 6 Tensor expansion and parallel element dot product:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal{T} \rightarrow \mathcal{T}{_{vec}} \in \mathbb {C}{^{{N_{\mathrm{{batch}}}} \times {N_\mathrm{{o}}}M}}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\boldsymbol{A}} \rightarrow \overline{\boldsymbol{A}} {_{vec}} = [{\boldsymbol{a}}({\theta _1});...;{\boldsymbol{a}}({\theta _{{N_\mathrm{{o}}}}})] \in \mathbb {C}{^{{N_\mathrm{{o}}}M}}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathcal{T}_0} = \mathcal{T}{_{vec}} \odot {\boldsymbol{A}}{_{vec}} \in \mathbb {C}{^{{N_{\mathrm{{batch}}}} \times {N_\mathrm{{o}}}M}}$$\end{document}. Step 7. Tensor reshape and sum the elements in the third dimension:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathcal{T}_0} \rightarrow {\mathcal{T}_1} \in \mathbb {C}{^{{N_{\mathrm{{batch}}}} \times {N_\mathrm{{o}}} \times M}}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal{T}{_2} = {\sum \nolimits _k {({\mathcal{T}_1})} _{i,j,k}} \in \mathbb {C}{^{{N_{\mathrm{{batch}}}} \times {N_\mathrm{{o}}}}}$$\end{document}. Step 8. Obtain the tensor of a batch MUSIC spectrum:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal{P} = \frac{1}{{\mathcal{T}{_2}}} = [{\boldsymbol{p}}_{\mathrm{{MU}}}^1,...,{\boldsymbol{p}}_{\mathrm{{MU}}}^{{N_{\mathrm{{batch}}}}}]$$\end{document}. During the DOA estimation phase, the two components are employed in a cascaded manner. The spatial filters processe the covariance matrix impacted by array errors, and the resulting array error-corrected covariance matrix serves as the input for MUSIC-NET, which subsequently generates MUSIC-like spectrum. Simulations and analyses The simulation settings are outlined in Table 1. The spatial filters comprise \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N_{\mathrm{{SF}}}}=6$$\end{document} filters, with each covering a subregion of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$20^\circ$$\end{document}. The spatial sampling is set with a spacing of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta \theta = {1^ \circ }$$\end{document}, resulting in a spatial grid of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\boldsymbol{\Theta }}_{\mathrm{{SF}}}} = \{ - {60^ \circ } - {59^ \circ },...,{59^ \circ },{60^ \circ }\}$$\end{document}. A single training data with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm {SNR_{train}}$$\end{document} = 10dB is collected for each grid (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${K_{\mathrm{{train}}}} = 1$$\end{document}), resulting in a total of 121 training data for the spatial filters. To accommodate practical cases of array errors, we introduce a parameter \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho \in [0,1]$$\end{document} to control the strength of errors and model the position deviation, gain deviation, phase deviation, and mutual coupling vector of the array antenna in a concise manner as15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} & {\boldsymbol{l}} = \rho \cdot {[\underbrace{ - 0.2,..., - 0.2}_5,\underbrace{0.2,...,0.2}_4]^\mathrm{{T}}} \cdot d, \end{aligned}$$\end{document}16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} & {{\boldsymbol{g}}_a}(\theta ) = \rho \cdot {[0,\underbrace{0.2,...,0.2}_5,\underbrace{ - 0.2,..., - 0.2}_4]^\mathrm{{T}}}, \end{aligned}$$\end{document}17\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} & {{\boldsymbol{g}}_p}(\theta ) = \rho \cdot {[0,\underbrace{ - {{30}^ \circ },..., - {{30}^ \circ }}_5,\underbrace{{{30}^ \circ },...,{{30}^ \circ }}_4]^\mathrm{{T}}}, \end{aligned}$$\end{document}18\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} & {\boldsymbol{c}}(\theta ) = \rho \cdot {[0,{\gamma ^1},...,{\gamma ^{M - 1}}]^\mathrm{{T}}}, \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma = 0.3{e^{j{{60}^ \circ }}}$$\end{document} is the mutual coupling coefficient between adjacent antennas. Then, in each training epoch, we randomly shuffle the training data and train \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N_{epoch}} = 1000$$\end{document} epochs with batch \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N_{batch}} = 64$$\end{document}.Table 1Simulation settings.Array parameters M=10, N=200, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${d}/{\lambda }=0.5$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\theta _{\min }} = - {60^ \circ }$$\end{document},\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\theta _{\max }} = {60^ \circ }$$\end{document}.Spatial filters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N_{\mathrm{{SF}}}}=6$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta \theta = {1^ \circ }$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N_\theta } = 121$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\eta = 10^{ - 3}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^{a}$$\end{document},\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${K_{\mathrm{{train}}}} = 1$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm{{SNR_{train}}} = 10\mathrm{{dB}}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N_{epoch}} = 1000$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N_{batch}} = 64$$\end{document}.MUSIC-NET \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta \vartheta = {0.1^ \circ }$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N_o} = 1201$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N_{\mathrm{{train}}}} = 50000$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${K_{\mathrm{{train}}}} = 2$$\end{document},\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm{{SNR_{train}}}\sim U(0\mathrm{{dB}},10\mathrm{{dB}})$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu = 10^{ - 3}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^{a}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N_{epoch}} = 50$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N_{batch}} = 2048$$\end{document}.The Adam with learning rate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$10^{ - 3}$$\end{document} is adopted for backward gradient propagation27. Table 1Simulation settings.Array parameters M=10, N=200, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${d}/{\lambda }=0.5$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\theta _{\min }} = - {60^ \circ }$$\end{document},\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\theta _{\max }} = {60^ \circ }$$\end{document}.Spatial filters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N_{\mathrm{{SF}}}}=6$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta \theta = {1^ \circ }$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N_\theta } = 121$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\eta = 10^{ - 3}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^{a}$$\end{document},\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${K_{\mathrm{{train}}}} = 1$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm{{SNR_{train}}} = 10\mathrm{{dB}}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N_{epoch}} = 1000$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N_{batch}} = 64$$\end{document}.MUSIC-NET \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta \vartheta = {0.1^ \circ }$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N_o} = 1201$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N_{\mathrm{{train}}}} = 50000$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${K_{\mathrm{{train}}}} = 2$$\end{document},\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm{{SNR_{train}}}\sim U(0\mathrm{{dB}},10\mathrm{{dB}})$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu = 10^{ - 3}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^{a}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N_{epoch}} = 50$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N_{batch}} = 2048$$\end{document}.The Adam with learning rate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$10^{ - 3}$$\end{document} is adopted for backward gradient propagation27. Simulation settings. Array parameters M=10, N=200, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${d}/{\lambda }=0.5$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\theta _{\min }} = - {60^ \circ }$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\theta _{\max }} = {60^ \circ }$$\end{document}. Spatial filters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N_{\mathrm{{SF}}}}=6$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta \theta = {1^ \circ }$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N_\theta } = 121$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\eta = 10^{ - 3}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^{a}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${K_{\mathrm{{train}}}} = 1$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm{{SNR_{train}}} = 10\mathrm{{dB}}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N_{epoch}} = 1000$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N_{batch}} = 64$$\end{document}. MUSIC-NET \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta \vartheta = {0.1^ \circ }$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N_o} = 1201$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N_{\mathrm{{train}}}} = 50000$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${K_{\mathrm{{train}}}} = 2$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm{{SNR_{train}}}\sim U(0\mathrm{{dB}},10\mathrm{{dB}})$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu = 10^{ - 3}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^{a}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N_{epoch}} = 50$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N_{batch}} = 2048$$\end{document}. The Adam with learning rate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$10^{ - 3}$$\end{document} is adopted for backward gradient propagation27. For MUSIC-NET, we use synthetic data for training and do not consider array errors. The spatial sampling is chosen as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta \vartheta = {0.1^ \circ }$$\end{document}, and the corresponding spatial grid is \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\boldsymbol{\Theta }}_{\mathrm{{MU}}}} = \{ - {60^ \circ } - {59.9^ \circ },...,{59.9^ \circ },{60^ \circ }\}$$\end{document}. We generate 50,000 training data sets by randomly selecting the direction of two signals (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${K_{\mathrm{{train}}}} = 2$$\end{document}) from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\boldsymbol{\Theta }}_{\mathrm{{MU}}}}$$\end{document} and setting \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm{{SNR_{train}}} \sim U(0\mathrm{{dB}},10\mathrm{{dB}})$$\end{document}. Then, in each training epoch, we randomly shuffle the training data and train \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N_{epoch}} = 50$$\end{document} epochs with batch \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N_{batch}} = 2048$$\end{document}. After training the spatial filters and MUSIC-NET, we fixed their parameters. Initially, we assess the effectiveness of spatial filters (SF) in the context of array error correction across varying intensities of array errors. This involves retraining the spatial filters for each specific array error intensity configuration. In the single-signal scenario (Fig. 2a, b), the conventional MUSIC with the uncorrected covariance matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\boldsymbol{R}}_e}$$\end{document} shows significant degradation in DOA estimation performance as array error intensity increases. Conversely, the proposed spatial filters efficiently adapt to correct array errors. MUSIC using the post-correction covariance matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\boldsymbol{\hat{R}}}$$\end{document} closely tracks the true DOA, demonstrating its superiority in mitigating array error impacts across varying error intensities. For the two-signal in the same beam scenario (Fig. 2c,d), the proposed array error correction method exhibits robust performance. Despite the signal angle range spanning \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$[-50^{\circ }, -42^{\circ }]$$\end{document} to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$[42^{\circ }, 50^{\circ }]$$\end{document}, the corrected DOA estimates (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _1$$\end{document}(SF+MUSIC) and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _2$$\end{document}(SF+MUSIC)) show smaller biases and better consistency with true DOAs compared to uncorrected counterparts(\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _1$$\end{document}(MUSIC) and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _2$$\end{document}(MUSIC)). These results validate the effectiveness of our spatial filters in handling array errors. As shown in Fig. 3, MUSIC-NET, trained with two signals, is tested under 2/3/4-signal scenarios. For the two-signal within the single-beam case, its amplitude peaks precisely align with true DOAs (marked by triangles), validating accurate estimation. Clustering results reveal that MUSIC-NET estimates (circled markers) closely match actual locations (cross markers). In three-signal and four-signal conditions, the method maintains robust performance. The amplitude spectrum still exhibits distinct peaks at actual directions, even with increased signal counts. Corresponding clustering plots demonstrate that its DOA estimates converge to actual locations across all clusters, which confirms that MUSIC-NET trained with two signals exhibits excellent generalization. The root mean square error (RMSE) of DOA estimation is denoted as19\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textrm{RMSE} = \sqrt{\frac{1}{{{N_{\mathrm{{trial}}}} K}}\sum \limits _{n = 1}^{{N_{\mathrm{{trial}}}}} {\sum \limits _{i = 1}^K {{{({\theta _i} - {{\hat{\theta }}_i})}^2}} } }, \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N_{\mathrm{{trial}}}}$$\end{document} denotes the number of trials. Figure 4a boxplots demonstrate that the DOA estimates of the proposed method are consistently concentrated around the actual values across various ρ levels, thereby highlighting the robustness of the method against variations in the ρ parameter. In Fig. 4b, uncorrected methods (MUSIC2, ESPRIT3, SBL9, SPICE10) exhibit significant accuracy deterioration as ρ increases, whereas the proposed method maintains a lower RMSE and surpasses DNN20 in terms of estimation precision. As shown in Fig. 4c,d, in comparison with corrected methods (SF+MUSIC, SF+ESPRIT, SF+SBL, SF+SPICE) and data-driven methods (DNN20 and SubNET22), the proposed method achieves a rapid reduction in RMSE as SNR or the number of snapshots increases. Notably, the RMSE of the proposed method is closest to the Cramér-Rao lower bound (CRLB)26, even under low SNR or limited snapshot conditions. These results fully validate the superior super-resolution capability, array error mitigation performance, and generalization ability of our method for closely spaced sources. Beyond the estimation accuracy and robustness against array imperfections, computational efficiency and real-time inference performance are critical metrics for practical engineering applications of DOA estimation methods. In this section, we provide a comprehensive analysis of the theoretical computational complexity and the actual average running time of the proposed method and mainstream benchmark methods. The theoretical complexity derivation, and the measured runtime statistics are summarized in Table 2.Table 2Computational complexity and average running time comparison of different methods.MethodComputational complexityAverage computation time (ms) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^{a}$$\end{document}MUSIC\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${O}_{MUSIC} = {O}(M^3) +{O}(N_o M^2)$$\end{document}4.6ESPRIT\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${O}_{ESPRIT} = {O}(M^3)$$\end{document}0.5SBL\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${O}_{SBL} = {O}(T_{iter} \cdot (N_o^2 M))$$\end{document}1190SPICE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${O}_{SPICE} = {O}(T_{iter} \cdot (M^3 + N_o M^2))$$\end{document}142DNN20\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${O}_{DNN} = {O}(M^4)$$\end{document}0.37SubNET22\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${O}_{SubNET} = {O}(M^3) + {O}(N_o M)$$\end{document}0.62Proposed\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${O}_{Proposed} = {O}(M^4) + {O}(N_o M^2)$$\end{document}0.41aThe average computation time is tested on a laptop equipped with an Intel Xeon E3-1535M v6 CPU (8GB RAM) and an NVIDIA Quadro M2200 GPU (4GB VRAM).Fig. 2DOA estimates and biases(\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{SNR}$$\end{document}=10dB,T=200). (a), (b) Single-signal. (c), (d) Two-signal in the same beam.Fig. 3MUSIC-NET output spectrum and K-means clustering results (SNR=0dB,T=200). (d), (e), and (f) correspond respectively to the k-means clustering results with the angle of spectrum line in (a), (b), and (c), based on the known signal number K. (a–d) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _1 = -30.5^{\circ },\theta _2 = -22.5^{\circ }$$\end{document}. (b–e) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _1 = -40.5^{\circ },\theta _2 = -20.5^{\circ },\theta _3 = 30.5^{\circ }$$\end{document}. (c–f) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _1 = -40.5^{\circ },\theta _2 = -20.5^{\circ },\theta _3 = 10.5^{\circ }, \theta _4 = 45.5^{\circ }$$\end{document}.Fig. 4The statistical performance(\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _1 = -5.5^{\circ },\theta _2 = 5.5^{\circ }$$\end{document}). (a) DOA estimates distribution of 1000 trials, SNR=0dB, T=200. (b) RMSE vs ρ, SNR=10dB, T=200. (c) RMSE vs SNR, ρ = 0.2, T=200. (d) RMSE vs Snapshots, ρ = 0.2, SNR=0dB. Table 2Computational complexity and average running time comparison of different methods.MethodComputational complexityAverage computation time (ms) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^{a}$$\end{document}MUSIC\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${O}_{MUSIC} = {O}(M^3) +{O}(N_o M^2)$$\end{document}4.6ESPRIT\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${O}_{ESPRIT} = {O}(M^3)$$\end{document}0.5SBL\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${O}_{SBL} = {O}(T_{iter} \cdot (N_o^2 M))$$\end{document}1190SPICE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${O}_{SPICE} = {O}(T_{iter} \cdot (M^3 + N_o M^2))$$\end{document}142DNN20\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${O}_{DNN} = {O}(M^4)$$\end{document}0.37SubNET22\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${O}_{SubNET} = {O}(M^3) + {O}(N_o M)$$\end{document}0.62Proposed\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${O}_{Proposed} = {O}(M^4) + {O}(N_o M^2)$$\end{document}0.41aThe average computation time is tested on a laptop equipped with an Intel Xeon E3-1535M v6 CPU (8GB RAM) and an NVIDIA Quadro M2200 GPU (4GB VRAM). Computational complexity and average running time comparison of different methods. aThe average computation time is tested on a laptop equipped with an Intel Xeon E3-1535M v6 CPU (8GB RAM) and an NVIDIA Quadro M2200 GPU (4GB VRAM). DOA estimates and biases(\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{SNR}$$\end{document}=10dB,T=200). (a), (b) Single-signal. (c), (d) Two-signal in the same beam. MUSIC-NET output spectrum and K-means clustering results (SNR=0dB,T=200). (d), (e), and (f) correspond respectively to the k-means clustering results with the angle of spectrum line in (a), (b), and (c), based on the known signal number K. (a–d) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _1 = -30.5^{\circ },\theta _2 = -22.5^{\circ }$$\end{document}. (b–e) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _1 = -40.5^{\circ },\theta _2 = -20.5^{\circ },\theta _3 = 30.5^{\circ }$$\end{document}. (c–f) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _1 = -40.5^{\circ },\theta _2 = -20.5^{\circ },\theta _3 = 10.5^{\circ }, \theta _4 = 45.5^{\circ }$$\end{document}. The statistical performance(\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _1 = -5.5^{\circ },\theta _2 = 5.5^{\circ }$$\end{document}). (a) DOA estimates distribution of 1000 trials, SNR=0dB, T=200. (b) RMSE vs ρ, SNR=10dB, T=200. (c) RMSE vs SNR, ρ = 0.2, T=200. (d) RMSE vs Snapshots, ρ = 0.2, SNR=0dB. For classical subspace-based methods, the computational complexity of the standard MUSIC is dominated by two core operations: the eigenvalue decomposition (EVD) of the M× M array covariance matrix, which brings a complexity of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${O}(M^3)$$\end{document}, and the full-angle traversal for spatial spectrum calculation over \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$N_o$$\end{document} angular grids, which contributes a complexity term of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${O}(N_o M^2)$$\end{document}. The ESPRIT method avoids the full-angle spectrum traversal process, resulting in a reduced overall complexity of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${O}_{ESPRIT} = {O}(M^3)$$\end{document}, where the EVD operation still accounts for the dominant computational overhead. For Sparsity-inducing methods, the SBL has a theoretical complexity of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${O}_{SBL} = {O}(T_{iter} \cdot (N_o^2 M))$$\end{document}, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T_{iter}$$\end{document} denotes the number of iterations required for algorithm convergence, and matrix operations dominate the per-iteration complexity for posterior probability update and hyperparameter optimization. The SPICE yields a complexity of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${O}_{SPICE} = {O}(T_{iter} \cdot (M^3 + N_o M^2))$$\end{document}, with the per-iteration overhead mainly coming from covariance matrix inversion and spatial spectrum estimation. It is worth noting that the actual computational cost of these iterative methods is heavily dependent on the convergence speed, which deteriorates significantly in harsh scenarios such as low SNR, limited snapshots, and array imperfections. For existing deep learning-based DOA estimation methods, the DNN20 framework presents an overall complexity of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${O}_{DNN} = {O}(M^4)$$\end{document}, which is dominated by the fully connected layer operations with the input dimension proportional to the square of the number of array elements M. The SubNET22 achieves a theoretical complexity of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${O}_{SubNET} = {O}(M^3) + {O}(N_o M)$$\end{document}, where the EVD operation inherited from classical subspace methods remains the core computational bottleneck. The network module contributes the linear complexity term with respect to the angular sampling number \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$N_o$$\end{document}. For our proposed cascaded framework, the overall computational complexity is \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${O}_{Proposed} = {O}(M^4) + {O}(N_o M^2)$$\end{document}. The \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${O}(M^4)$$\end{document} term comes from the fully connected layers of the 4-layer adaptive spatial filters, whose input and output dimensions are proportional to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$M^2$$\end{document}, while the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${O}(N_o M^2)$$\end{document} term is attributed to the convolutional and fully connected operations in the 8-layer MUSIC-NET for high-precision spatial spectrum mapping. Despite the relatively higher theoretical complexity of the proposed method compared with classical MUSIC and ESPRIT, there exists a fundamental difference in hardware acceleration potential and actual inference efficiency between the model-driven and data-driven frameworks, which is well validated by the measured average computation time. The inference process of the proposed cascaded DNNs consists entirely of parallelizable tensor operations, including matrix multiplication and element-wise nonlinear activation. These operations can be efficiently accelerated via pipeline optimization and massively parallel computing on mainstream hardware platforms such as GPUs and digital signal processors (DSPs), thus achieving extremely low end-to-end inference latency. In contrast, the computational overhead of classical subspace methods is dominated by the EVD operation, which is an inherently serial process with very limited parallelization capability. Especially in large-scale array and wide angular scanning scenarios, the actual runtime of these classical methods increases drastically with the cubic growth of EVD complexity, leading to inferior real-time performance compared with the proposed method. Compared with the iterative Sparsity-inducing methods (SBL and SPICE), the proposed method exhibits overwhelming superiority in both computational efficiency and inference stability. As shown in Table 2, the average running time of SBL and SPICE is 3 to 4 orders of magnitude higher than that of the proposed method. This is because the actual computational cost of iterative methods is determined by the number of convergence iterations \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T_{iter}$$\end{document}, which often increases to tens or even hundreds in challenging non-ideal scenarios. Moreover, these methods face non-negligible risks of iteration divergence and unstable estimation under low SNR and array imperfection conditions. In sharp contrast, the proposed method adopts an end-to-end feedforward inference architecture without any iterative convergence process, yielding fixed and controllable computational overhead for each single-sample inference, which is highly favorable for engineering applications with stringent real-time requirements. When compared with other deep learning-based benchmarks, the proposed method also shows comprehensive advantages. Compared with the DNN20, the proposed method achieves competitive inference latency. For SubNET22, although it has a lower theoretical complexity, it retains the serial EVD bottleneck of classical subspace methods, which cannot fully exploit the parallel acceleration capability of modern hardware. Conclusion In this paper, we exploit deep learning to address realistic array imperfections and DOA estimation by decoupling error mitigation and DOA estimation. The proposed method attains robust performance across diverse error intensities, SNR levels, and snapshot conditions. Simulation results reveal that the spatial filters effectively rectify array errors, while MUSIC-NET achieves accurate DOA estimation. The framework surpasses state-of-the-art methods in terms of RMSE and generalization capabilities, especially in low-SNR and limited-snapshot circumstances.