CN113466782A - Deep Learning (DL) -based cross-coupling correction D O A estimation method - Google Patents

Deep Learning (DL) -based cross-coupling correction D O A estimation method Download PDF

Info

Publication number
CN113466782A
CN113466782A CN202110636128.7A CN202110636128A CN113466782A CN 113466782 A CN113466782 A CN 113466782A CN 202110636128 A CN202110636128 A CN 202110636128A CN 113466782 A CN113466782 A CN 113466782A
Authority
CN
China
Prior art keywords
mutual coupling
array
vector
signal
training
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202110636128.7A
Other languages
Chinese (zh)
Other versions
CN113466782B (en
Inventor
童美松
张晨婧
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Tongji University
Original Assignee
Tongji University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Tongji University filed Critical Tongji University
Priority to CN202110636128.7A priority Critical patent/CN113466782B/en
Publication of CN113466782A publication Critical patent/CN113466782A/en
Application granted granted Critical
Publication of CN113466782B publication Critical patent/CN113466782B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
    • G01S3/00—Direction-finders for determining the direction from which infrasonic, sonic, ultrasonic, or electromagnetic waves, or particle emission, not having a directional significance, are being received
    • G01S3/02—Direction-finders for determining the direction from which infrasonic, sonic, ultrasonic, or electromagnetic waves, or particle emission, not having a directional significance, are being received using radio waves
    • G01S3/14—Systems for determining direction or deviation from predetermined direction
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Abstract

The invention relates to the technical field of signal processing, in particular to a mutual coupling correction D O A estimation method based on Deep Learning (DL). A DNN structure is designed in the method and consists of a rough division and mutual coupling correction module and a subdivision mapping module, wherein the rough division and mutual coupling correction module and the subdivision mapping module are respectively realized by a self-encoder and a ResNet which are fully connected, and the realization process is as follows: firstly, processing signals to obtain vectors containing characteristic information as input, then preliminarily decomposing the vectors into a plurality of angle domains through a rough separation and mutual coupling correction module, carrying out mutual coupling correction, and finally merging prediction results of all the subdivision mapping networks through subdivision mapping networks corresponding to each angle domain to determine accurate arrival directions of all signal sources. The structure can estimate DOA of any information source, including information sources with different information source numbers and different signal-to-noise ratios, and can reduce the influence of array element coupling on DOA precision. Compared with the MUSIC algorithm with the precision of 0.1 degree, the method has the advantage that the operation speed is improved by 50 times.

Description

Deep Learning (DL) -based cross-coupling correction D O A estimation method
Technical Field
The present invention relates to the technical field of signal processing.
Background
DOA estimation research is an important branch of array signal processing, and has numerous applications in various fields of communication, radar, geophysical, acoustic, biomedical engineering, etc., and is particularly important in wireless communication because it can help optimize the radiation pattern of an antenna array by using an adaptive beamforming algorithm, thereby minimizing interference. There are currently many algorithms to solve the DOA estimation problem. They are mainly based on eigenvalue decomposition of the spatial covariance matrix of the signals received by the antenna elements.
The subspace-based approach is called Multiple Signal Classification (MUSIC), known for its super-resolution function. MUSIC, however, performs an intensive spectral search process to provide an accurate DOA estimate. Therefore, it is difficult to satisfy the requirements of high accuracy, high resolution, and real-time performance because of the large amount of calculation. In recent years, Compressed Sensing (CS) theory is used to solve the DOA estimation problem, and the method has better estimation performance than the conventional algorithm with a smaller number of samples and a low SNR. However, the conventional compressed sensing algorithm also has certain limitations, such as high correlation due to grid division, and the like, and still needs to be improved in terms of estimation accuracy and complexity.
Neural Networks (NN) provide a fast and accurate alternative to the powerful nonlinear approximation capability and perform the underlying mathematical operations, and are therefore well suited for determining the angular position of source signals. Most of the current research is based on an ideal model, and in order to consider the situation of array coupling, the result is either low in the accuracy of angle positioning or large in limitation because different neural networks need to be trained according to different incident signal numbers.
Disclosure of Invention
The invention provides a mutual coupling correction DOA estimation method based on Deep Learning (DL), wherein a Deep Neural Network (DNN) structure matched with an array antenna signal model is designed, the DOA of any information source can be estimated through the input and output design, and the problem that the neural network is required to be trained respectively aiming at different incident signals is solved. In addition, the neural network firstly carries out estimation and cross coupling correction of a coarse grid on the signal, and then a two-step estimation method for refining accurate estimation of the grid improves the accuracy and robustness of the algorithm.
In order to achieve the purpose, the invention provides the following technical scheme: a DOA estimation method of any source based on Deep Learning (DL), comprising the following steps:
the method comprises the following steps: defining the range of DOA and the number M of array elements, obtaining a correlation matrix from signals received by the antenna array, and respectively taking a real part and an imaginary part of partial elements of a triangle on the correlation matrix to form a real number vector as the input of the whole DNN.
Step two: and (3) collecting a large number of array signals received by the cross-coupling array antenna, and using the characteristic vector obtained after the processing in the step one as training data. The coarse and mutual coupling correction modules are realized by fully connected self-encoders, and ReLu activation functions are adopted. The decoder divides the angle of arrival into U sub-angle domains, and if there is an incoming wave in the angle range represented by the decoder, the ideal output of the decoder is equal to the eigenvector when the mutual coupling is zero, otherwise, it is 0. And training the self-encoder by taking the MSE of the ideal value and the real value as a loss function, wherein the neurons of the hidden layer are inactivated randomly according to the probability of 0.5 in the training process.
Step three: and the rough division and mutual coupling correction module is used as an input data training subdivision mapping module, so that the signal feature space R can be accurately mapped to the signal direction space theta. The direction space is discretized into L grid points, then the two grid points closest to the incident direction are 1, and the rest are 0. The ResNet is trained using the MSE of the estimated and true results as a loss function.
Step four: and inputting the characteristic vector corresponding to the receiving signal of the mutual coupling array into the trained deep neural network as a characteristic to obtain the possibility of the incidence angle falling on a certain lattice point.
Step five: and carrying out sectional weighted average processing on the obtained output data to obtain the non-meshed DOA.
The invention has the beneficial effects that: different from the traditional neural network model, the DOA estimation can be carried out without knowing the number of incident waves in advance by the model, the estimation error caused by the coupling effect among array elements is effectively reduced, and the estimation precision is improved. In addition, experiments prove that the MUSIC algorithm with the speed ratio precision of 0.1 degree is improved by 50 times, and the requirement of real-time property is met.
The method provided by the invention can estimate DOA of any information source, including information sources with different information source numbers and different signal-to-noise ratios, and can reduce the influence of array element coupling on DOA precision.
Drawings
FIG. 1 is a block diagram of a coarse and cross-coupling calibration module according to an embodiment of the present invention
FIG. 2 is a block diagram of a subdivision mapping module according to an embodiment of the present invention
FIG. 3 is a graph of simulation results for an embodiment of the present invention where ρ is 0.1
FIG. 4 is a graph of simulation results for an embodiment of the present invention where ρ is 0.4
FIG. 5 is a graph showing the comparison between DNN and MUSIC operation times according to the embodiment of the present invention
Detailed Description
The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings and embodiments.
The method comprises the following steps: defining the range of DOA and the number M of array elements, obtaining a correlation matrix from signals received by the antenna array, and respectively taking a real part and an imaginary part of partial elements of a triangle on the correlation matrix to form a real number vector as the input of the whole DNN.
Assuming that the array antenna is an equidistant linear array, K narrow-band signals are incident, and the signals received by the array antenna are Xc(t)＝AS(t)+N，A＝[C×a(θ1)C×a(θ2)…C×a(θK)]Is a steering matrix of the array antenna, and C is a mutual coupling coefficient vector
Figure BDA0003105299050000031
The Toeplitz matrix of (1), ε is the coupling coefficient, ρ is the tuning parameter, S (t) is the incident signal, and N is white Gaussian noise. Since the dimension of the signal correlation matrix is only related to the number of array elements, the invention uses the signal correlation matrix RcAs an input to reduce the variability of the DNN input.
Figure BDA0003105299050000032
N denotes the number of fast beats acquired, X (N) is for Xc(t) sampling. Since the correlation matrix contains all the information of the incident signal, and RcIs Hermite (Hermite) array, element Rc(i, j) and RcThe information of (j, i) is the same, and the diagonal elements do not contain signal direction information, so considering only the elements of the upper triangular part, one vector is obtained:
Figure BDA0003105299050000033
Figure BDA0003105299050000034
to pair
Figure BDA0003105299050000035
Taking coefficients of a real part and an imaginary part respectively and combining the coefficients into a real number vector
Figure BDA0003105299050000036
| represents the euclidean norm. The number of input neurons for this module is M (M-1).
Step two: and (3) collecting a large number of array signals received by the cross-coupling array antenna, and using the characteristic vector obtained after the processing in the step one as training data. The coarse and mutual coupling correction module is realized by a fully connected self-encoder, the self-encoder firstly compresses an input vector to a lower dimension by using an encoder to extract useful components of the input vector, then restores the useful components to the original dimension by using decoders, different decoders are respectively used for restoring signal components of different angle domains and correcting mutual coupling, the structure of the mutual coupling correction module is shown in figure 1, and the network model calculation formula is as follows:
Figure BDA0003105299050000037
in the formula (I), the compound is shown in the specification,
Figure BDA0003105299050000038
representing neural network1The output vector of the layer is then calculated,
Figure BDA0003105299050000039
representing the random inactivation of neurons of the hidden layer according to the probability P in the training process,this solves the problem of overfitting.
Figure BDA00031052990500000310
Denotes the l1The layer activation function, ReLu is used in the present invention.
Figure BDA0003105299050000041
Denotes from the l1Layer 1 to layer l1The weight of a layer is determined by the weight of the layer,
Figure BDA0003105299050000042
is the first1Layer additive offset vectors.
The U decoders average the angle of arrival into U sub-angle domains, and if there is an incoming wave in the angle range represented by the decoder, the ideal output of the decoder is equal to the eigenvector when the mutual coupling is zero, otherwise it is 0. The method for calculating the eigenvector when the mutual coupling is zero is the same as the first step, and C is the identity matrix. Obtaining the expected output of the entire auto-encoder:
Figure BDA0003105299050000043
then, the training set and the label set of the rough-separation and mutual-coupling correction module are respectively:
Figure BDA0003105299050000044
Figure BDA0003105299050000045
and training the self-encoder by taking the MSE of the label value and the real value as a loss function.
Step three: the output of the rough-division and mutual-coupling correction module is a training subdivision mapping module, so that the signal feature space R can be accurately mapped to the signal direction space theta. The module is realized by ResNet, the structure of which is shown in FIG. 2, and U parallel ResNet are constructed in total, wherein the input of the U-th ResNet is the output of the U-th decoder of the coarse separation and mutual coupling correction module. The network model calculation formula is as follows:
Figure BDA0003105299050000046
superscript (·)(u)Parameter representing the u-th ResNet, definition
Figure BDA0003105299050000047
The outputs of the U ResNet are combined to obtain the total output:
Figure BDA0003105299050000048
in particular, the amount of the solvent to be used,
Figure BDA0003105299050000049
is calculated by the formula
Figure BDA00031052990500000410
In the formula [ theta ]jFor the angle represented by the jth grid,
Figure BDA00031052990500000411
i.e., the direction space is discretized into J grid points, then one or two grid points closest to the incident direction are 1, and the rest are 0. Then, the training set and the label set of the subdivision mapping module are respectively:
Figure BDA0003105299050000051
step four: and inputting the characteristic vector corresponding to the receiving signal of the mutual coupling array into the trained deep neural network as a characteristic to obtain the possibility of the incidence angle falling on a certain lattice point.
Step five: and carrying out sectional weighted average processing on the obtained output data to obtain the non-meshed DOA. And taking all the angle grids with the output near the maximum value larger than a certain threshold value into consideration, carrying out weighted average processing on the angle grids, and weighting the angle grids to obtain the output value of the grid, thereby finally obtaining the non-grid precise DOA.
Verification experiment
A phased array antenna with 12 array elements is an example. The range of possible incoming wave directions of-60 °,60 ° is evenly divided into 4 sub-angular domains, i.e., -60 °, -30 °,0 °, [0 °,30 °) and [30 °,60 °). The coarse and cross-coupling correction modules have 132 input neurons and 528 output neurons. The subdivision mapping module contains 4 ResNet, each ResNet having 132 input neurons and 30 output neurons. The array antenna receives incident signals with incident angles of-20 deg. and 10 deg. respectively,
Figure BDA0003105299050000052
Figure BDA0003105299050000053
rho in the vector is 0.1 and 0.4, the results of 100 simulation experiments are respectively shown in fig. 3 and fig. 4, the prediction accuracy is high, and the influence of the coupling degree change is small, so that the DOA estimation can be well carried out on the antenna array with array elements mutually coupled. Fig. 5 shows that the speed is increased by about 50 times by comparing the operation time of the MUSIC algorithm with the accuracy of 0.1 °.

Claims (4)

1. A D O A estimation method of any source based on Deep Learning (DL), which is characterized by comprising the following steps:
the method comprises the following steps: defining the range of DOA and the number M of array elements, obtaining a correlation matrix from signals received by an antenna array, and respectively taking a real part and an imaginary part of partial elements of a triangle on the correlation matrix to form a real number vector as the input of the whole DNN;
step two: collecting array signals received by a large number of cross-coupling array antennas, and using the characteristic vectors obtained after the processing in the first step as training data; the coarse and mutual coupling correction modules are realized by fully connected self-encoders, and ReLu activation functions are adopted; the decoder divides the arrival angle into U sub-angle domains, if the angle range represented by the decoder has an incoming wave, the ideal output of the decoder is equal to the characteristic vector when the mutual coupling is zero, otherwise, the ideal output is 0; training a self-encoder by taking MSE of an ideal value and an actual value as a loss function, wherein neurons of a hidden layer are inactivated randomly according to the probability of 0.5 in the training process;
step three: the rough-division and cross-coupling correction module is used as an input data training subdivision mapping module, so that the signal characteristic space R is mapped to a signal direction space theta; discretizing the direction space into L grid points, wherein two grid points closest to the incident direction are 1, and the rest are 0; training ResNet by taking the MSE of the estimated result and the real result as a loss function;
step four: inputting a characteristic vector corresponding to a mutual coupling array receiving signal into a trained deep neural network as a characteristic to obtain the possibility that an incident angle falls on a certain lattice point;
step five: and carrying out sectional weighted average processing on the obtained output data to obtain the non-meshed DOA.
2. The method of claim 1, characterized by the steps of:
assuming that the array antenna is an equidistant linear array, K narrow-band signals are incident, and the signals received by the array antenna are Xc(t)＝AS(t)+N，A＝[C×a(θ1) C×a(θ2) … C×a(θK)]Is a steering matrix of the array antenna, and C is a mutual coupling coefficient vector
Figure FDA0003105299040000011
The Toeplitz matrix of (1), epsilon is a coupling coefficient, rho is an adjustment parameter, S (t) is an incident signal, and N is white Gaussian noise; using a signal correlation matrix RcAs an input to reduce the variability of the DNN input;
Figure FDA0003105299040000012
n denotes the number of fast beats acquired, X (N) is for Xc(t) sampling; since the correlation matrix contains all the information of the incident signal, and RcIs Hermite (Hermite) array, element Rc(i, j) and RcThe information of (j, i) is the same, and the diagonal elements do not contain signal direction information, so considering only the elements of the upper triangular part, one vector is obtained:
Figure FDA0003105299040000013
to pair
Figure FDA0003105299040000021
Taking coefficients of a real part and an imaginary part respectively and combining the coefficients into a real number vector
Figure FDA0003105299040000022
Figure FDA0003105299040000023
| represents the euclidean norm; the number of input neurons for this module is M (M-1).
3. The method of claim 1, characterized by the steps of:
collecting array signals received by a large number of cross-coupling array antennas, and using the characteristic vectors obtained after the processing in the first step as training data; the coarse and mutual coupling correction module is realized by a fully connected self-encoder, the self-encoder firstly compresses an input vector to a lower dimension by using an encoder to extract useful components of the input vector, then restores the useful components to the original dimension by using decoders, different decoders are respectively used for restoring signal components of different angle domains and correcting mutual coupling, and a network model calculation formula is as follows:
Figure FDA0003105299040000024
in the formula (I), the compound is shown in the specification,
Figure FDA0003105299040000025
representing neural network1The output vector of the layer is then calculated,
Figure FDA0003105299040000026
representing that neurons of a hidden layer are inactivated randomly according to the probability P in the training process;
Figure FDA0003105299040000027
denotes the l1A layer activation function, ReLu;
Figure FDA0003105299040000028
denotes from the l1Layer 1 to layer l1The weight of a layer is determined by the weight of the layer,
Figure FDA0003105299040000029
is the first1A layer additive offset vector;
the U decoders divide the arrival angle into U sub-angle domains, if the angle range represented by the decoders has the arrival wave, the ideal output of the decoders is equal to the characteristic vector when the mutual coupling is zero, and otherwise, the ideal output is 0; the calculation method of the characteristic vector when the mutual coupling is zero is the same as the first step, and C is an identity matrix; obtaining the expected output of the entire auto-encoder:
Figure FDA00031052990400000210
then, the training set and the label set of the rough-separation and mutual-coupling correction module are respectively:
Figure FDA00031052990400000211
Figure FDA00031052990400000212
and training the self-encoder by taking the MSE of the label value and the real value as a loss function.
4. The method of claim 1, characterized by the steps of three:
the network model calculation formula is as follows:
Figure FDA00031052990400000213
superscript (·)(u)Parameter representing the u-th ResNet, definition
Figure FDA00031052990400000214
The outputs of the U ResNet are combined to obtain the total output:
Figure FDA0003105299040000031
in particular, the amount of the solvent to be used,
Figure FDA0003105299040000032
is calculated by the formula
Figure FDA0003105299040000033
In the formula [ theta ]jFor the angle represented by the jth grid,
Figure FDA0003105299040000034
namely, the direction space is discretized into J grid points, one or two grid points closest to the incident direction are 1, and the rest are 0;
then, the training set and the label set of the subdivision mapping module are respectively:
Figure FDA0003105299040000035
Figure FDA0003105299040000036
CN202110636128.7A 2021-06-08 2021-06-08 Mutual coupling correction DOA estimation method based on Deep Learning (DL) Active CN113466782B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110636128.7A CN113466782B (en) 2021-06-08 2021-06-08 Mutual coupling correction DOA estimation method based on Deep Learning (DL)

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110636128.7A CN113466782B (en) 2021-06-08 2021-06-08 Mutual coupling correction DOA estimation method based on Deep Learning (DL)

Publications (2)

Publication Number Publication Date
CN113466782A true CN113466782A (en) 2021-10-01
CN113466782B CN113466782B (en) 2022-09-13

Family

ID=77868702

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110636128.7A Active CN113466782B (en) 2021-06-08 2021-06-08 Mutual coupling correction DOA estimation method based on Deep Learning (DL)

Country Status (1)

Country Link
CN (1) CN113466782B (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2023070994A1 (en) * 2021-10-27 2023-05-04 东南大学 Array super-resolution direction of arrival estimation method

Citations (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
FR1449700A (en) * 1965-05-21 1966-05-06 Thomson Houston Comp Francaise Improvements to reception systems for very weak radio signals
US20120100136A1 (en) * 2009-05-01 2012-04-26 Ophthotech Corporation Methods for treating or preventing ophthalmological diseases
CN103914431A (en) * 2014-01-16 2014-07-09 同济大学 Mesh-less method for calculating anisotropic structure radar cross section
CN110007265A (en) * 2019-04-30 2019-07-12 哈尔滨工业大学 A kind of Wave arrival direction estimating method based on deep neural network
CN110212966A (en) * 2019-06-11 2019-09-06 哈尔滨工业大学 Mutual coupling of antenna bearing calibration based on importance resampling under the conditions of a kind of coherent source
CN110967665A (en) * 2019-10-07 2020-04-07 西安电子科技大学 DOA estimation method of moving target echoes under multiple external radiation sources
CN111767791A (en) * 2020-05-21 2020-10-13 中国地质大学（武汉） Arrival angle estimation method based on anti-regularization deep neural network
CN112379327A (en) * 2020-12-01 2021-02-19 北京工业大学 Two-dimensional DOA estimation and cross coupling correction method based on rank loss estimation
CN112699735A (en) * 2020-12-07 2021-04-23 中国人民解放军国防科技大学 Near-field source positioning method based on deep learning
KR102247740B1 (en) * 2020-01-10 2021-05-04 홍익대학교 산학협력단 Apparatus for estimating direction-of-arrival based on artificial neural network and method thereof
US20210159964A1 (en) * 2019-11-25 2021-05-27 Yangtze University Direction-of-arrival estimation and mutual coupling calibration method and system with arbitrary sensor geometry and unknown mutual coupling
CN112881972A (en) * 2021-01-15 2021-06-01 电子科技大学 Direction-of-arrival estimation method based on neural network under array model error

Patent Citations (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
FR1449700A (en) * 1965-05-21 1966-05-06 Thomson Houston Comp Francaise Improvements to reception systems for very weak radio signals
US20120100136A1 (en) * 2009-05-01 2012-04-26 Ophthotech Corporation Methods for treating or preventing ophthalmological diseases
CN103914431A (en) * 2014-01-16 2014-07-09 同济大学 Mesh-less method for calculating anisotropic structure radar cross section
CN110007265A (en) * 2019-04-30 2019-07-12 哈尔滨工业大学 A kind of Wave arrival direction estimating method based on deep neural network
CN110212966A (en) * 2019-06-11 2019-09-06 哈尔滨工业大学 Mutual coupling of antenna bearing calibration based on importance resampling under the conditions of a kind of coherent source
CN110967665A (en) * 2019-10-07 2020-04-07 西安电子科技大学 DOA estimation method of moving target echoes under multiple external radiation sources
US20210159964A1 (en) * 2019-11-25 2021-05-27 Yangtze University Direction-of-arrival estimation and mutual coupling calibration method and system with arbitrary sensor geometry and unknown mutual coupling
KR102247740B1 (en) * 2020-01-10 2021-05-04 홍익대학교 산학협력단 Apparatus for estimating direction-of-arrival based on artificial neural network and method thereof
CN111767791A (en) * 2020-05-21 2020-10-13 中国地质大学（武汉） Arrival angle estimation method based on anti-regularization deep neural network
CN112379327A (en) * 2020-12-01 2021-02-19 北京工业大学 Two-dimensional DOA estimation and cross coupling correction method based on rank loss estimation
CN112699735A (en) * 2020-12-07 2021-04-23 中国人民解放军国防科技大学 Near-field source positioning method based on deep learning
CN112881972A (en) * 2021-01-15 2021-06-01 电子科技大学 Direction-of-arrival estimation method based on neural network under array model error

Non-Patent Citations (8)

* Cited by examiner, † Cited by third party
Title
AYA MOSTAFA AHMED 等: "Deep Autoencoders for DOA Estimation of Coherent Sources using Imperfect Antenna Array", 《2020 THIRD INTERNATIONAL WORKSHOP ON MOBILE TERAHERTZ SYSTEMS (IWMTS)》 *
CHEN JING ZHANG 等: "An Improved Genetic Algorithm for Multiband Microstrip Antenna Design", 《2019 PHOTONICS & ELECTROMAGNETICS RESEARCH SYMPOSIUM - FALL (PIERS - FALL)》 *
PING LANG 等: "A Comprehensive Survey of Machine Learning Applied to Radar Signal Processing", 《JOURNAL OF LATEX CLASS FILES》 *
ZHANG-MENG LIU 等: "Direction-of-Arrival Estimation Based on Deep Neural Networks With Robustness to Array Imperfections", 《IEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION》 *
张志峰 等: "集成电路设计平台构建及其教学科研应用", 《工业和信息化教育》 *
张永皓 等: "一种基于深度学习的环形阵列混合波束成型大规模MIMO系统DOA估计算法", 《微型电脑应用》 *
葛晓凯 等: "利用深度学习方法的相干源DOA估计", 《信号处理》 *
郭业才 等: "基于稀疏堆叠降噪自编码器-深层神经网络的语音DOA估计算法", 《实验室研究与探索》 *

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2023070994A1 (en) * 2021-10-27 2023-05-04 东南大学 Array super-resolution direction of arrival estimation method

Also Published As

Publication number Publication date
CN113466782B (en) 2022-09-13

Similar Documents

Publication Publication Date Title
Xiang et al. Improved de-multipath neural network models with self-paced feature-to-feature learning for DOA estimation in multipath environment
CN109061554B (en) Target arrival angle estimation method based on dynamic update of spatial discrete grid
CN109298383B (en) Mutual-prime array direction-of-arrival estimation method based on variational Bayes inference
CN109375154B (en) Coherent signal parameter estimation method based on uniform circular array in impact noise environment
CN108872926B (en) Amplitude-phase error correction and DOA estimation method based on convex optimization
Reddy et al. Reduced-complexity super-resolution DOA estimation with unknown number of sources
Xiang et al. Altitude measurement based on characteristics reversal by deep neural network for VHF radar
CN112881972B (en) Direction-of-arrival estimation method based on neural network under array model error
Zhou et al. Radar coincidence imaging for off-grid target using frequency-hopping waveforms
CN113466782B (en) Mutual coupling correction DOA estimation method based on Deep Learning (DL)
Hu et al. A self-calibration algorithm for cross array in the presence of mutual coupling
Elvander et al. Tracking and sensor fusion in direction of arrival estimation using optimal mass transport
Reaz et al. A comprehensive analysis and performance evaluation of different direction of arrival estimation algorithms
Kwizera et al. Direction of arrival estimation based on MUSIC algorithm using uniform and non-uniform linear arrays
Qi et al. DOA estimation and self-calibration algorithm for multiple subarrays in the presence of mutual coupling
CN109507634B (en) Blind far-field signal direction-of-arrival estimation method based on propagation operator under any sensor array
Zhang et al. Direction finding in partly calibrated arrays exploiting the whole array aperture
Bhargav et al. Compressive sensing based doa estimation for multi-path environment
Ning et al. A velocity independent MUSIC algorithm for DOA estimation
CN115248413A (en) Off-grid signal direction-of-arrival estimation method suitable for non-uniform linear array
Rajani et al. Direction of arrival estimation by using artificial neural networks
Liu et al. DOA estimation with sparse array under unknown mutual coupling
Yin et al. Performance study of different parameters for arbitrary arrays and a novel dimension reduction-MUSIC algorithm
Zheng et al. Deep Learning-Based DOA Estimation
Guo et al. Deep Learning Based DOA Estimation With Trainable-Step-Size LMS Algorithm

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant