CN108828547B - Meter-wave radar low elevation height measurement method based on deep neural network - Google Patents

Meter-wave radar low elevation height measurement method based on deep neural network Download PDF

Info

Publication number
CN108828547B
CN108828547B CN201810650951.1A CN201810650951A CN108828547B CN 108828547 B CN108828547 B CN 108828547B CN 201810650951 A CN201810650951 A CN 201810650951A CN 108828547 B CN108828547 B CN 108828547B
Authority
CN
China
Prior art keywords
data
network
training
characteristic
neural network
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.)
Active
Application number
CN201810650951.1A
Other languages
Chinese (zh)
Other versions
CN108828547A (en
Inventor
陈伯孝
杨婷
项厚宏
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.)
Xidian University
Original Assignee
Xidian 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 Xidian University filed Critical Xidian University
Priority to CN201810650951.1A priority Critical patent/CN108828547B/en
Publication of CN108828547A publication Critical patent/CN108828547A/en
Application granted granted Critical
Publication of CN108828547B publication Critical patent/CN108828547B/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
    • G01S7/00—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00
    • G01S7/02—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00
    • G01S7/41—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00 using analysis of echo signal for target characterisation; Target signature; Target cross-section
    • G01S7/417—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00 using analysis of echo signal for target characterisation; Target signature; Target cross-section involving the use of neural networks
    • 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
    • G01S13/00—Systems using the reflection or reradiation of radio waves, e.g. radar systems; Analogous systems using reflection or reradiation of waves whose nature or wavelength is irrelevant or unspecified
    • G01S13/88—Radar or analogous systems specially adapted for specific applications
    • G01S13/882—Radar or analogous systems specially adapted for specific applications for altimeters

Landscapes

  • Engineering & Computer Science (AREA)
  • Radar, Positioning & Navigation (AREA)
  • Remote Sensing (AREA)
  • Physics & Mathematics (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • General Physics & Mathematics (AREA)
  • Electromagnetism (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Radar Systems Or Details Thereof (AREA)

Abstract

The invention discloses a meter-wave radar low elevation height measurement method based on a deep neural network, which solves the problem of model mismatch under the condition of low elevation angle and comprises the following steps: preprocessing echo data to obtain a normalized amplitude spectrum; constructing a compression network and a decompression network with a symmetrical structure to form DNN, taking the normalized amplitude spectrum as the input of the DNN to obtain a characteristic F, decompressing to obtain decompressed data, and training and fine-tuning the weight W and the offset b of DNN parameters by using an Adam algorithm and a BP algorithm to complete the construction and training of the deep neural network; when the network is optimal, all data in the training set are input into a compression network and unitized to obtain a characteristic base set; inputting the preprocessed test set data into the trained DNN to obtain the characteristic F ', projecting the characteristic F' in the characteristic base set of the training set to determine the target incoming wave angle, and calculating the target height. The invention improves the radar height measurement accuracy, is still effective under the conditions of low elevation angle and model mismatch, and is used in the field of low elevation angle height measurement of meter wave radar.

Description

Meter-wave radar low elevation height measurement method based on deep neural network
Technical Field
The invention belongs to the technical field of radars, particularly relates to radar low elevation height measurement, and particularly relates to a meter-wave radar low elevation height measurement method based on deep learning.
Background
The meter-wave radar plays an important role in anti-stealth, anti-radiation missile and remote early warning. However, the meter-wave radar has a long wavelength, a wide main lobe and a limited array aperture, so that the angular resolution is not high, and the high-measurement performance is directly influenced. Especially under the condition of low elevation angle, due to the fact that the wave beam hits the ground, the target multipath signal is particularly complex, and the multipath signal and the direct wave signal belong to strong coherent signals, and the coherent signals are difficult to distinguish in a space domain, a time domain and a frequency domain.
In order to solve the problem, a great number of scholars at home and abroad carry out a great deal of research, wherein the main research contents comprise two aspects:
on one hand, based on the classical reflection model, the purpose of estimating the Direction of Arrival (DOA) is achieved by performing de-coherence on the direct wave signal and the multipath signal, such as the classical multi-signal classification algorithm and the rotation invariant subspace algorithm. When the direct wave and the multipath reflected wave are completely coherent, the rank of the covariance matrix of the array receiving data is reduced to 1, and the rank of the covariance matrix is restored through a de-coherence technique, so that the DOA estimation purpose is achieved. The spatial smoothing algorithm and the matrix reconstruction algorithm achieve rank recovery by reducing the degree of freedom of the system. The disadvantage of such algorithms is the loss of array aperture, and the Toeplitz preprocessing method takes into account the Toeplitz nature of the true data covariance matrix, but the accuracy is not high.
On the other hand, by constructing a more accurate multipath reflection model, the model is close to the real situation as much as possible, and the angle measurement performance is improved. The maximum likelihood algorithm based on the accurate multipath reflection model adds prior information such as radar erection height, earth curvature and the like to the traditional multipath model, however, the algorithm does not consider the influence of the terrain reflection height of the actual complex position and is only applicable to the condition that the terrain of the position is relatively flat. The influence of actual terrain height information on the algorithm is considered by the accurate terrain matching-based synthesis guidance algorithm. The disturbing multipath signal model and the inversion method thereof estimate the height of the reflecting surface, but the algorithm only depends on the characteristics of the signal model, and once the terrain characteristics are mismatched with the model, the algorithm performance is poor.
The neural network technology has the advantages of self-adaptive learning capability, nonlinear characteristic, rapid convergence characteristic and the like, and has been successfully applied to the field of signal processing. Neural networks have successfully implemented classification of complex data and function approximation problems. Many scholars at home and abroad intensively research a Radial Basis Function (RBFNN) and successfully apply the RBFNN to the DOA estimation problem. According to the method proposed by Zooghby et al in pages 1611 to 1617 of volume 11 of the journal IEEE Transactions on Antennas and Propagation of 1997 and the method proposed by Dukrin et al in pages 139 to 144 of the IEEE International Conference on Communication Conference paper of 2002, the DOA estimation problem is reduced to a complex mapping relationship of input and output, the DOA estimation is successfully carried out on coherent sources and non-coherent sources by training RBFNN, and the method has smaller calculation amount compared with the traditional SSMUSIC algorithm and greatly reduces the time consumption of DOA estimation. Sallam et al, in order to obtain the optimal weight for beam forming in the method proposed in the 2016, pages 5095 to 5104 of volume 9 of the journal of IEEE Transactions on Geoscience and Remote Sensing, trained the input and output relationship of the neural network, obtained the optimal wiener solution of a set of network weights, and verified that the network has better angle measurement performance than the traditional Capon beam forming and the like through comparison of experimental and simulation data. The method proposed by Liuyu et al in pages 5558 to 5579 of volume 7 in 2009 of publications successfully applies the artificial neural network technology to the cloud classification problem, and the performance is greatly improved compared with a Principal Component Analysis (PCA) and a Support Vector Machine (SVM). The method proposed by Blackwell et al, at pages 2535 to 2546 of volume 11 of the IEEE Transactions on Geoscience and Remote Sensing journal, performs dimensionality reduction on data by projection Principal component analysis (PPC), and then uses Deep Neural Networks (DNN) to estimate atmospheric temperature and wet profiles. Wang Runkun et al proposed a DOA estimation method for successfully implementing HF radar by using support vector regression algorithm from page 374 to page 678 in volume 5 of 2018 of the journal of IEEE Transactions on Geoscience and Remote Sensing.
The method based on de-coherence of direct wave signals and multipath signals has the defects of aperture loss and poor precision, and the performance of the method based on accurate construction of the multipath reflection model is rapidly deteriorated when the terrain features are mismatched with the model. The DOA estimation method based on RBFNN applies the neural network technology to angle measurement, but the adopted neural network structure is a single hidden layer, and the learning capability of the target characteristics is limited.
Disclosure of Invention
In order to overcome the defects of the prior art, the invention provides the low elevation meter wave radar height measurement method based on the deep neural network, which has higher precision and higher operation speed and can still normally work under the condition of model mismatch.
The invention relates to a meter-wave radar low elevation height measurement method based on a deep neural network, which is characterized by comprising the following steps of:
step 1, echo data preprocessing: dividing the acquired data into training set data and test set data and marking respectively; carrying out frequency domain transformation on the echo data X (m) of the labeled training set to obtain a magnitude spectrum X (K) of the data, wherein K is 1,2 … K and K is the total dimension of the data, carrying out normalization processing on the magnitude spectrum X (K) to obtain a normalized magnitude spectrum X: x ═ X (X-min (X))/(max (X)) -min (X));
step 2, training a multilayer deep neural network DNN: training labeled training set data x (m) in batches, and constructing a compression network, wherein the output dimension of the compression network is smaller than the input dimension: taking the normalized amplitude spectrum X as the input of a compression network and carrying out feature extraction on input data to obtain a feature F; constructing a decompression network with symmetrical structure, wherein the output dimension of the decompression network is equal to the input dimension of the compression network: taking the compressed characteristic F as an input of a decompression network, and performing decompression processing to obtain decompressed data X' (K), wherein K is 1,2 … K; calculating the mean square error X of the amplitude spectrum data X (k) and the decompression data X' (k)mseTraining the weight W and the bias b of the network parameters by adopting an Adaptive Moment Estimation algorithm (Adam), and finely adjusting the weight W and the bias b of the network parameters by adopting a Back Propagation (BP) algorithm after each iteration until the error of a data set is converged, namely determining the network parameters W and b, and completing the training and construction of the multilayer deep neural network;
and 3, extracting the characteristic base of the training set data: when the network parameters are optimal in the DNN training process, sequentially inputting single data in a training set, acquiring the characteristics of all the single data and unitizing to obtain a characteristic base FiI is 1,2 … n, n is the number of characteristics, and constitutes the characteristic base set { F1,F2…FnAnd each characteristic base in the characteristic base set corresponds to an elevation angle thetaiI is 1,2 … n, and the elevation angle set corresponding to the characteristic base set is { theta [ [ theta ] ]1,θ2…θn}；
Step 4, estimating the DOA of the data incoming wave direction of the test set: carrying out data preprocessing on the test set data: normalizing the test set data after the frequency domain transformation by using the maximum value and the minimum value of the training set data to obtain a magnitude spectrum after normalization; will be normalizedInputting the amplitude spectrum into a trained multilayer DNN network, compressing to obtain a test set data characteristic F ', projecting the test set data characteristic F' into a characteristic basis set of training set data, and determining an incoming wave angle:
Figure BDA0001704784150000031
and calculating the height of the low elevation target to complete the height measurement of the low elevation target of the meter-wave radar.
The invention solves the problems of model mismatch and the like in the meter-wave radar engineering in practice, deeply explores the relation between the radar receiving data characteristics and the environmental characteristics after researching the action process of the environment on the information source, and introduces a deep learning technology on the basis of the traditional height measurement model and algorithm to carry out DOA estimation on the meter-wave radar.
Compared with the prior art, the invention has the following advantages:
(1) the precision is higher: under the conditions of low signal-to-noise ratio and small snapshot number, the classical SSMUSIC algorithm mainly aims at the DOA estimation problem of a coherent source although rank recovery is performed on a matrix by a spatial smoothing method. In the multipath environment, the multipath signal is not completely coherent with the direct wave signal, and even only part of the array elements of the receiving antenna are interfered by the multipath signal. Therefore, the SSMUSIC can not effectively solve the DOA estimation problem of the low elevation angle target, and the DOA effect is poor; the invention inputs the received direct wave signal and the multipath signal into the deep neural network, makes full use of the target information contained in the echo signal, and accurately carries out DOA estimation with higher precision.
(2) Solving the problem of model mismatch: the model mismatch is caused by serious multipath effect when the array environment is severe and the target elevation angle is low, the performance of the classical SSMUSIC algorithm is greatly reduced, even the SSMUSIC algorithm is completely invalid, and the actual requirement cannot be met; a DOA estimation method based on RBFNN adopts a single hidden layer neural network for learning, and the network learning capability is limited, so that the feature extraction of a target is insufficient. The invention adopts the multilayer neural network DNN, can fully and effectively extract the characteristics of the echo signals, directly utilizes the signal characteristics to invert the target elevation angle, solves the problem of model mismatch, and controls the height measurement error in a reasonable range.
(3) The operation speed is faster: the classical SSMUSIC algorithm needs to carry out eigenvalue decomposition on a data covariance matrix in the operation process, and the operation amount is very large; the method directly utilizes the trained deep neural network to project the data of the test set into the training set, determines the angle of the incoming wave through spectral peak search, has high operation speed and better meets the requirement of real-time processing.
Drawings
The invention is described in further detail below with reference to the following description of the drawings and the detailed description.
FIG. 1 is a flow chart of the present invention;
FIG. 2 is a schematic diagram of a deep neural network;
FIG. 3 is a diagram showing the relationship between the source angle and the root mean square error of the angle measurement of the present invention and the classical SSMUSIC method;
FIG. 4 is a three-dimensional spectrum of the present invention at an angle of incidence of (-2, 2);
FIG. 5 is a plot of signal-to-noise ratio versus root mean square error for the classical SSMUSIC method of the present invention;
FIG. 6 is a test set track diagram;
FIG. 7 is a graph of the angle measurement results of the present invention and the SSMUSIC method;
FIG. 8 is a graph of the height measurements of the present invention and SSMUSIC method.
Detailed Description
The present invention will be described in detail below with reference to the accompanying drawings.
Example 1
When the target is at a low elevation angle, the multipath signal is strong, the direct echo signal of the target is submerged in the multipath signal, the problem of model mismatch can be generated by adopting the traditional method, and the method has the defects of poor precision, aperture loss and the like. The DOA estimation method based on RBFNN applies the neural network technology to angle measurement, but the adopted neural network structure is a single hidden layer, the learning capability is limited, and the feature extraction of the target is not sufficient. Therefore, the invention develops research and provides a meter-wave radar low elevation height measurement method based on a deep neural network, and the method is shown in figure 1 and comprises the following steps:
step 1, echo data preprocessing: dividing the acquired data into training set data and test set data and marking respectively; carrying out frequency domain transformation on echo data X (m) of the labeled training set to obtain an amplitude spectrum X (K) of the data, wherein K is a dimension variable, K is 1,2 … K, and K is a total dimension of the echo data, and in order to avoid that a large value exists in X (K) to inactivate neurons and influence the mapping effect of a deep neural network, normalization processing needs to be carried out on the amplitude spectrum X (K) to obtain the amplitude spectrum X after the normalization processing: x ═ X (X-min (X))/(max (X)) -min (X)).
Step 2, training a multilayer deep neural network DNN: to speed up the training speed, the tagged training set data x (m) is trained in batches. Constructing a compression network, the output dimension of which is smaller than the input dimension: and taking the normalized amplitude spectrum X as the input of the compression network and extracting the features of the input data to obtain a feature F, wherein the dimension of the feature F is smaller than that of the amplitude spectrum X (k), namely, the compression network performs dimension reduction on the amplitude spectrum data X (k). Constructing a decompression network symmetrical to the structure of the compression network, the output dimension of the decompression network being equal to the input dimension of the compression network: taking the compressed characteristic F as an input of a decompression network, and performing decompression processing to obtain decompressed data X' (K), wherein K is 1,2 … K; and the decompression network performs dimension increasing processing on the characteristic F to obtain decompressed data X' (k) with the same dimension as that of the amplitude spectrum data X (k), namely, the depth neural network is used for fitting the input amplitude spectrum data X (k). Initializing the weight W and the bias b of the network with a smaller value, and calculating the mean square error X of the amplitude spectrum data X (k) and the decompression data X' (k)mseTraining the weight W and the bias b of the network parameters by adopting an adaptive moment estimation Adam algorithm, and finely adjusting the weight W and the bias b by adopting a Back Propagation (BP) algorithm after each iteration until the mean square error X of the data setmseAnd (4) converging, namely determining the parameter weight W and the bias b of the network, and finishing the training and construction of the multilayer deep neural network.
The deep neural network adopted by the invention has better representation capability on input amplitude spectrum data X (k) by a hidden layer structure, and the data is represented more abstractly by layer-by-layer feature extraction.
And 3, extracting the characteristic base of the training set data: when the network parameters are optimal in the DNN training process, sequentially inputting single data in a training set, acquiring the characteristics of all the single data and unitizing to obtain a characteristic base FiI is 1,2 … n, n is the number of characteristics, and constitutes the characteristic base set { F1,F2…Fn}. The characteristic base set comprises angle information of different angle point sources in the space domain, namely each characteristic base in the characteristic base set corresponds to an elevation angle thetaiI is 1,2 … n, and the elevation angle set corresponding to the characteristic base set is { theta [ [ theta ] ]1,θ2…θn}. Elevation angle theta with concentrated elevation anglesiCorresponding to point sources at different angles in the airspace, the target is positioned at a certain elevation angle in the elevation angle set. The invention effectively extracts the data characteristic F and extracts the angle information of the target from the echo signal.
Step 4, estimating the DOA of the data incoming wave direction of the test set: carrying out data preprocessing on the test set data: for the test set data, in order to ensure the training set and the test set input value interval matching, the maximum value and the minimum value of the training set data are used for carrying out normalization processing on the test set data after frequency domain transformation to obtain a magnitude spectrum after the normalization processing; inputting the normalized amplitude spectrum into a trained multilayer DNN network, compressing to obtain a test set data characteristic F ', projecting the test set data characteristic F' into a characteristic basis set of training set data, and determining an incoming wave angle:
Figure BDA0001704784150000061
and calculating the height of the low elevation target to complete the height measurement of the low elevation target of the meter-wave radar. According to the method, after the trained multi-layer DNN network is used for extracting the data characteristics of the test set, the data characteristics of the test set are projected into a characteristic basis set F' of the training set data, the incoming wave angle is determined, and the low elevation target height is calculated.
The invention utilizes the deep neural network technology to extract data characteristics (environmental characteristics), trains a group of new characteristic bases, and projects the characteristics of the test set data onto the characteristic base set, thereby achieving the purpose of inverting DOA.
Under the conditions of severe array environment and low target elevation angle, the model mismatch is caused by serious multipath effect, the performance of the classical SSMUSIC algorithm is greatly reduced, even completely fails, and the actual requirement cannot be met; the DOA estimation method based on RBFNN adopts a single hidden layer neural network to process echo data, and the feature extraction of the target is not sufficient. The invention adopts a multilayer deep neural network, can fully and effectively extract the characteristics of echo signals, directly utilizes the signal characteristics to invert DOA, solves the problem of model mismatch and controls the height measurement error in a reasonable range.
Example 2
The method for measuring height of meter-wave radar at low elevation angle based on the deep neural network is the same as that in embodiment 1, referring to fig. 2, and the step 2 of DNN training of the multilayer deep neural network comprises the following steps:
2.1, constructing a deep neural network: and taking the normalized magnitude spectrum X as the input of a compression network, and performing feature extraction on input magnitude spectrum data to obtain a feature F, wherein the feature F comprises the azimuth, distance and elevation information of any point source in an airspace. And meanwhile, constructing another decompression network with a structure symmetrical to the compression network, decompressing the compressed characteristic F to obtain decompressed data X' (k), and fitting the input normalized amplitude spectrum X by using a deep neural network formed by the compression network and the decompression network.
2.2 training of deep neural network: because the composition of the multipath signal is relatively complex, the weight W and the bias b of the network parameter are trained by adopting a self-adaptive time estimation algorithm so as to meet different multipath conditions. A corrected Linear Unit (ReLU) is adopted as a network activation function, and a nonlinear factor is added, so that the Linear fitting condition is avoided, and the fitting effect is not influenced. The network activation function ReLU is defined as follows:
Figure BDA0001704784150000071
the ReLU function is non-linear, easily propagates errors backwards and does not activate all neurons simultaneously. If the input value is negative, the ReLU function will convert the input value to 0 and the neurons will not be activated, i.e. only a small number of neurons will be activated for a period of time, so that the neural network is sparse and thus efficient and easy to compute.
Calculating the mean square error X of the decompressed data X' (k) and the amplitude spectrum data X (k)mse：
Figure BDA0001704784150000072
The smaller the error, the better the fit to the magnitude spectral data x (k). Fine-tuning the network weight W and the bias b by using an error back propagation BP algorithm after each iteration, updating the network weight W and the bias b from an output layer to a hidden layer and then to an input layer according to a chain derivation rule after the BP algorithm calculates the weight W and the bias b of each layer by forward propagation, optimizing the network weight after multiple times of training, decompressing the mean square error X of data X' (k) and amplitude spectrum data X (k)mseThe minimum, the best fitting effect. When the iteration is completed, the dimension of the feature F obtained by compression is usually smaller than the dimension of the amplitude spectrum data x (k), i.e. the amplitude spectrum data x (k) of the echo data is subjected to dimension reduction processing, see fig. 2. Assuming that feature F has dimension q, feature F retains the q-dimensional features of the magnitude spectral data x (k).
2.3 hidden layer adjustment of deep neural network: the deep neural network structure comprises an input layer, an output layer and a hidden layer; the hidden layer structure has better representation capability on input data, the data are represented more abstractly through layer-by-layer feature extraction, and data features are effectively extracted. The number of layers and the number of nodes of the hidden layer are adjusted according to the training result.
2.4 batch processing of training set data: in order to accelerate the training process, the training set data is processed in batches, each batch of data generates respective gradient descending direction, the gradient search range in the optimization process is reduced, the optimal weight W and the bias b can be quickly found, and the training and construction of the multilayer deep neural network are completed.
The method comprises the steps of taking a normalized amplitude spectrum X as input of a compression network, extracting characteristics of data to obtain characteristics F, constructing another decompression network with a symmetrical structure, decompressing the compressed characteristics F to obtain decompressed data X '(k), comparing differences between the amplitude spectrum data X (k) and the decompressed data X' (k), training a weight W and an offset b of network parameters by adopting a self-adaptive time estimation algorithm, and finely adjusting the network weight W and the offset b by utilizing a BP algorithm after each iteration. The invention uses multilayer neural network, makes full use of data characteristics contained in multipath signals, and effectively extracts target information in echo signals.
Example 3
The method for measuring height of a meter-wave radar at a low elevation angle based on a deep neural network is the same as that in the embodiment 1-2, and the step 3 of extracting the training set data feature base specifically comprises the following steps:
when the network parameters are optimal in the DNN training process, sequentially inputting single data in a training set, acquiring the characteristics of all the single data and unitizing to obtain a characteristic base FiI is 1,2 … n, n is the number of characteristics, and constitutes the characteristic base set { F1,F2…FnAnd constructing a characteristic space corresponding to point sources at different elevation angles in the space domain. The characteristic base set comprises angle information of point sources with different elevation angles in the space domain, namely each characteristic base in the characteristic base set corresponds to an elevation angle thetaiI is 1,2 … n, and the elevation angle set corresponding to the characteristic base set is { theta [ [ theta ] ]1,θ2…θn}. Elevation angle theta with concentrated elevation anglesiCorresponding to point sources at different angles in the airspace, the target is positioned at a certain elevation angle in the elevation angle set. The invention effectively extracts the data characteristic F from the echo signal and constructs the characteristic space.
Example 4
The method for measuring height of a meter-wave radar at a low elevation angle based on a deep neural network is the same as that in the embodiment 1-3, and the step 4 of calculating the height of the target at the low elevation angle specifically comprises the following steps:
for test set data, in order to ensure that the input values of a training set and a test set are matched, the maximum value and the minimum value of the training set are required to be subjected to normalization processing and then input into a trained network; as the sample in the test set only needs to extract the characteristics of the sample and does not need to decompress the characteristics, the method has the advantages of simple process, low cost and the likeOnly half of the deep neural network is needed, the test set data characteristics F 'are directly output in the compression network, and the test set data characteristics F' are projected in the characteristic space of the training set to determine the incoming wave angle
Figure BDA0001704784150000085
Figure BDA0001704784150000081
The target height h is calculated. In the formula FiIs a characteristic base of training set data, n is the number of characteristics of the training set data, theta is the angle of a point source in the space domain,
Figure BDA0001704784150000082
representing the data characteristics F' of the test set in the data characteristic base F of the training setiProjection of (2).
Obtain the angle of the incoming wave
Figure BDA0001704784150000083
Calculating the corresponding target height h:
Figure BDA0001704784150000084
in the formula reIs 4/3 equivalent earth radius, haIs the radar altitude, rdIs the true distance of the target from the radar.
Example 5
The method for measuring height of a meter-wave radar at a low elevation angle based on a deep neural network is the same as that in embodiments 1-4, and the number of layers and the number of nodes of the hidden layer in step 2.3 are adjusted according to a training result, and specifically comprises the following steps:
and adjusting the number of layers and the number of nodes of the hidden layer of the deep neural network according to the training result. When the number of layers of the hidden layer is too many and the number of nodes is too many, the deep neural network overfitts the input data, only the training set data is valid but the test set data is invalid, namely the DNN cannot effectively extract the characteristics of the input test set data, and the number of layers and the number of nodes of the hidden layer need to be reduced; when the number of layers of the hidden layer is too small and the number of nodes is too small, the learning of the deep neural network is limited, the characteristics of input data cannot be effectively learned, the training error of a training set or a test set is large, the problem of under-fitting is caused, and the number of layers and the number of nodes of the hidden layer need to be increased. The number of layers and the number of nodes of the hidden layer of the deep neural network are adjusted, so that the fitting effect of the decompressed data X' (k) and the amplitude spectrum data X (k) is good. In this example, it is assumed that the total number of layers of the network is 5, and the number of nodes is 256 × 128 × 64 × 128 × 256, where the hidden layer with the number of nodes of 64 is the feature output layer, i.e., the dimension q of the feature F is 64.
A more detailed example is given below to further illustrate the present invention.
Example 6
The meter-wave radar low elevation height measurement method based on the deep neural network is the same as the embodiment 1-5.
Referring to fig. 1, the low elevation radar height finding method based on deep learning of the invention comprises the following specific steps:
step 1, assuming that the receiving array is a uniform linear array with M array elements, and the fast beat number is L, the array receiving signal y (t) is: x (t) ═ as (t) + n (t), where x (t) ═ x (t)1(t),x2(t),…,xM(t)]TReceiving a data vector for the array, N (t) ═ n1(t),n2(t),…,nM(t)]TFor noisy data vectors, S (t) ═ s1(t),s2(t),…,sM(t)]TIs a source data vector.
Figure BDA0001704784150000091
Steering the vector for the array; carrying out frequency domain transformation on the labeled data X (m) to obtain a magnitude spectrum X (k) of the data, wherein in order to avoid that a larger or smaller numerical value in X (k) influences the mapping effect of the deep neural network, normalization processing needs to be carried out on X (k) to obtain a normalized magnitude spectrum X: x ═ X (X-min (X))/(max (X)) -min (X)).
Step 2, taking the normalized frequency spectrum as the input of the compression network, extracting the characteristics of the data to obtain characteristics F, constructing another decompression network with symmetrical structure at the same time, decompressing the compressed characteristic data to obtain data X' (k), and calculating decompressed dataMean square error X of X' (k) and amplitude spectrum data X (k)mseAfter each iteration, the BP algorithm is needed to be utilized to finely adjust the network weight and the bias each time, so that the network weight is optimal, and the error X ismseAnd minimum. When the iteration is complete, the resulting compressed data F is typically smaller in dimension than x (k), assuming that F has dimension q, then F preserves the q-dimensional characteristics of the input data x (k). Assuming that the total number of deep neural networks is 5 layers and the number of nodes is 256 × 128 × 64 × 128 × 256, the network structure is shown in fig. 2.
The hidden layer with the node number of 64 is the feature output layer, i.e. q is 64. The network activation function adopts a ReLU function, and the ReLU function is defined as follows:
Figure BDA0001704784150000101
the network optimization algorithm adopts a self-adaptive moment estimation Adam algorithm to train the weight W and the bias b of the network parameters, and adopts a BP algorithm to finely adjust the network weight. In order to accelerate the training process, the training set data can be divided into a plurality of small batches of data, and the training and construction of the deep neural network are completed.
Step 3, when the network parameters are optimal in the DNN training process, sequentially inputting single data in the training set, acquiring the characteristics of all the single data and unitizing to obtain a characteristic base FiI is 1,2 … n, n is the number of characteristics, and constitutes the characteristic base set { F1,F2…FnAnd each characteristic base in the characteristic base set corresponds to an elevation angle thetaiI is 1,2 … n, and the elevation angle set corresponding to the characteristic base set is { theta [ [ theta ] ]1,θ2…θn}。
And 4, for the data of the test set, in order to ensure that the input values of the training set and the test set are matched, the maximum value and the minimum value of the training set are required to be subjected to normalization processing, and then the normalized data are input into the network with trained parameters. Because the sample of the test set only needs to extract the characteristics of the sample and does not need to decompress the characteristics, only half of the network is needed, for example, the network structure is 256 × 128 × 64, the characteristics F 'are directly output, and the F' is projected to the characteristic space of the training setDetermining an incoming wave angle in the middle:
Figure BDA0001704784150000102
referring to fig. 4, the peak value of the spectral peak is searched, the closer the search angle is to the real angle of the target, the stronger the characteristic amplitude is, and otherwise, the weaker the search angle is, and the target height h is calculated. In the formula FiThe method is characterized in that the method is a feature base of training set data, n is the number of features of the training set data, and theta is an angle of a point source in the space domain.
Obtain the angle of the incoming wave
Figure BDA0001704784150000103
Calculating the corresponding target height h:
Figure BDA0001704784150000104
in the formula reIs 4/3 equivalent earth radius, haIs the radar altitude, rdIs the true distance of the target from the radar. And (4) completing accurate DOA estimation through searching and calculation.
The invention inputs the echo signal into the deep neural network, makes full use of the target information contained in the echo signal, accurately performs DOA estimation, and has higher precision and higher operation speed.
The model mismatch can be caused by serious multipath effect when the array environment is severe and the target elevation angle is low, the performance of the classical SSMUSIC algorithm is greatly reduced, even the SSMUSIC algorithm is completely invalid, and the actual requirement cannot be met; the DOA estimation method based on RBFNN adopts a single-layer neural network to process echo data, and the feature extraction of the target is not sufficient. The invention adopts the multi-layer deep neural network DNN, can fully and effectively extract the characteristics of the echo signals, directly utilizes the signal characteristics to invert DOA, solves the problem of model mismatch, and controls the height measurement error in a reasonable range.
The effect of the present invention can be further illustrated by the following simulation experiments:
example 7
The meter-wave radar low elevation height measurement method based on the deep neural network is the same as the embodiment 1-6.
1) Simulation conditions are as follows: the array structure is set to be a 24-array-element uniform linear array, the wavelength is 1m, and the array element interval is 0.5 m. Two cases were simulated. Data generation and processing of the experiment was done on MATLAB2017a and the neural network training part was done on python 3.5.
2) Simulation content:
simulation 1: setting the angle of incidence theta in consideration of the two fully coherent sources1∈[-1°:0.1:-3°]，θ2∈[1°:0.1:3°]The noise is white gaussian noise, and 100 sets of data are generated, wherein 10 sets of data are randomly extracted as test samples. Compared with the classical SSMUSIC algorithm, the angle measurement root mean square error caused by the coherent information source included angle is compared with the information source included angle theta1-θ2The analysis result of the root mean square error of the | and the angle measurement is shown in fig. 3, the curve with the asterisk is the root mean square error curve of the angle measurement caused by the included angle of the coherent information source in the SSMUSIC algorithm, and the curve with the circle is the root mean square error curve of the angle measurement caused by the included angle of the coherent information source in the invention. When the incident angle is (-2 deg., 2 deg.), the three-dimensional spectrum of the present invention is shown in fig. 4.
3) And (3) simulation result analysis:
as can be seen from fig. 3, considering that the 3dB beamwidth of the array is about 4 °, when the coherent source angle is smaller than 4 °, the SSMUSIC algorithm has completely failed, but the present invention can still effectively perform DOA estimation when the coherent source angle is smaller than the 3dB beamwidth of the array. The angle measurement precision obtained by deep neural network inversion is gradually reduced along with the increase of the included angle of the information source, and the performance of the angle measurement precision is far superior to that of an SSMUSIC algorithm, so that the angle measurement precision is proved to be effectively applied under the condition of coherent information sources.
Example 8
The method for measuring height of the meter-wave radar at low elevation angle based on the deep neural network is the same as that in the embodiments 1-5, and the simulation conditions are the same as that in the embodiment 7.
FIG. 4 is a three-dimensional spectrogram of the DOA estimation method under the condition of the coherent source, wherein when the incident angle is (-2 degrees and 2 degrees), the three-dimensional spectrogram is analyzed from a three-dimensional spectrogram of a point with the incident angle of (-2 degrees and 2 degrees, the real angle can be effectively inverted through characteristic inversion, the spectral peak is very obvious, and the DOA estimation effectiveness of the DOA estimation method under the condition of the coherent source is verified. And analyzing on the spectrogram of the point trace, wherein the closer the search angle is to the real angle of the target, the stronger the characteristic amplitude is, and the weaker the characteristic amplitude is otherwise. The purpose of estimating the DOA in the incoming wave direction can be achieved by searching the peak value of the spectrum peak.
Example 9
The method for measuring height of the meter-wave radar at low elevation angle based on the deep neural network is the same as that in the embodiments 1-5, and the simulation conditions are the same as that in the embodiment 7.
Simulation 2: fast beat number is 10, SNR ∈ [ -10:5:10 [)]dB, consider the case of two fully coherent sources, the angle of incidence θ1＝-2°，θ2The noise is white Gaussian noise, 400 groups of data are respectively generated, wherein 100 groups of data are randomly extracted as test samples, the angle measurement root mean square error under different signal-to-noise ratios is compared with the SSMUSIC algorithm, the simulation result is shown in figure 5, the curve with the asterisk is the angle measurement root mean square error curve of the SSMUSIC algorithm under different signal-to-noise ratios, and the curve with the circle is the angle measurement root mean square error curve of the invention under different signal-to-noise ratios.
As can be seen from fig. 5, when the coherent source included angle is approximately equal to 3dB in width, the classical ssmusci algorithm performs rank recovery on the matrix, but its DOA effect is still poor, and when the signal-to-noise ratio is 5dB, the root mean square error of angle measurement is 2.112 °, but the present invention can still perform DOA estimation accurately, and when the signal-to-noise ratio is 5dB, the root mean square error of angle measurement is 0.5075 °, which is significantly smaller than that of the ssmusci algorithm. Because DNN precision is restricted by the minimum training batch number and training times, if the minimum training batch number is larger, although convergence is fast, the error is larger; on the contrary, if the minimum training batch number is smaller, although the error is smaller, the training time is longer, so the angle measurement root mean square error can not be rapidly reduced along with the increase of the signal-to-noise ratio, but the time problem does not need to be considered during the training, so the network structure can be further adjusted to enable the algorithm performance to be optimal.
Example 9
The meter-wave radar low elevation height measurement method based on the deep neural network is the same as the embodiment 1-5.
Field experiment: in order to verify the practicability of the invention, the measured data of a certain array of the meter wave radar is processed. In order to ensure the mutual exclusivity of the training set and the test set, the training set adopts a plurality of routes with the same course and known truth values, and the data of the training set is subjected to feature extraction and a feature basis set is constructed. The test set is data of another flight, a track diagram of the flight is shown in fig. 6, the flight stably flies from south to north, the real height of the flight is about 9.5km, the flight distance radar is about 80-250 km, the elevation angle is less than 6 degrees, the beam width of the radar 3dB is about 4 degrees, the position environment of the target is very bad, and more objects such as high buildings, trees, mountains and the like exist, and the type belongs to a complex position low elevation height measurement type. Compared with the classical SSMUSIC algorithm, the angle measurement result and the height measurement result are respectively shown in FIG. 7 and FIG. 8, the solid lines in FIG. 7 and FIG. 8 are true values, the asterisks indicate the result of the SSMUSIC algorithm, and the circles indicate the result of the method.
Analyzing fig. 7 and 8, comparing the angle measurement result and the height measurement result of the classical ssmuscic algorithm and the angle measurement result and the height measurement result of the invention with the true value respectively, it is easy to find that the angle measurement result and the angle measurement result of the classical ssmuscic algorithm have larger difference with the true value. Referring to fig. 7, when the distance is 137.2km, the true value of the target angle is 3.498 °, the angle measurement result of the SSMUSIC algorithm is 3.1 ° with a difference of 0.398 °, the angle measurement result of the present invention is 3.432 ° with a difference of 0.066 °. The angle measurement result of the SSMUSIC algorithm always differs from the true value of the target angle by about 0.3 degrees, and the error between the angle measurement result and the true value of the target angle is about 0.06 degrees. As can be seen from FIG. 7, the error of the angle measurement result of the method of the present invention is small, while the error of the angle measurement result of the SSMUSIC method is large.
Referring to fig. 8, when the distance is 137.2km, the true value of the target height is 9.479km, the height measurement result of the SSMUSIC method is 8.528km with a difference of 0.9510km, and the height measurement result of the present invention is 9.32km with a difference of 0.159 km. The difference between the height measurement result of the SSMUSIC algorithm and the true value of the target height is more than 0.9km, namely more than 900m, and the difference between the height measurement result of the SSMUSIC algorithm and the true value of the target height is not more than 500 m. As can be seen from FIG. 8, the error of the height measurement result of the method of the present invention is small, while the error of the height measurement result of the SSMUSIC method is large. The precision of the SSMUSIC algorithm is influenced by the constraint of array aperture and the array environment, so that effective decorrelation and accurate DOA estimation are difficult to perform when the array environment is complex, and the caused DOA estimation error is very large. The invention takes data characteristics (namely environment characteristics) as a research object, considers that the information sources at the adjacent positions of an airspace always have similar characteristics, and even if the position environment is severe and the target is at a low elevation angle, the height measurement error of the invention is not higher than 500m, and the invention completely meets the requirement of low elevation angle height measurement precision in the actual engineering, so the invention also has high practical engineering property.
In short, the meter-wave radar low elevation height measurement method based on the deep neural network disclosed by the invention solves the problem of model mismatch under the condition of low elevation angle, and comprises the following steps: echo data preprocessing: carrying out frequency domain transformation on the echo data X (m) with the label to obtain an amplitude spectrum X (k) of the echo data, and carrying out normalization processing on the X (k) to obtain a normalized amplitude spectrum X in order to avoid that a larger or smaller numerical value in the X (k) influences the mapping effect of the deep neural network; training of the deep neural network: taking the normalized amplitude spectrum X as the input of a compression network to extract the characteristics to obtain the characteristics F, meanwhile, constructing another decompression network which is symmetrical to the compression network structure to decompress the compressed characteristics F to obtain decompressed data X '(k), and calculating the mean square error X (k) of the amplitude spectrum data X (k) and the decompressed data X' (k)mseInitializing a weight W and a bias b by a smaller value, training network parameters W and b by adopting a self-adaptive time estimation algorithm, and finely adjusting the network weight W by adopting an error back propagation BP algorithm to optimize the network weight W and obtain a mean square error XmseTraining label samples in batches, determining network parameters W and b, and completing construction and training of the deep neural network; extracting feature bases of training set data: when the network weight W is optimal, inputting single data in a training set, obtaining the characteristics of the single data, and performing unitization processing on the characteristics to obtain characteristic bases, wherein all the characteristic bases form a characteristic base set, each characteristic base corresponds to an elevation angle, and the elevation angles corresponding to all the characteristic bases form an elevation angle set; estimating the incoming wave direction of the test set data: and after normalization processing is carried out on the test set data, inputting the trained network to obtain a characteristic F ', projecting the characteristic F' in a characteristic base set of the training set to determine a target incoming wave angle, and calculating the target height. The invention improves the height measurement precision of the meter wave radarThe method has the advantages of accelerating the operation speed and still being capable of being used under the conditions of low elevation angle and model mismatch. The invention is used for height measurement of the meter wave radar.

Claims (5)

1. A meter-wave radar low elevation height measurement method based on a deep neural network is characterized by comprising the following steps:
step 1, echo data preprocessing: dividing the acquired data into training set data and test set data and respectively carrying out label marking; carrying out frequency domain transformation on the echo data X (m) of the labeled training set to obtain a magnitude spectrum X (K) of the data, wherein K is 1,2 … K and K is a data dimension, carrying out normalization processing on the magnitude spectrum X (K) to obtain a magnitude spectrum X after the normalization processing: x ═ X (X-min (X))/(max (X)) -min (X));
step 2, training a multilayer deep neural network DNN: training labeled training set data x (m) in batches, and constructing a compression network, wherein the output dimension of the compression network is smaller than the input dimension: taking the normalized amplitude spectrum X as the input of a compression network and carrying out feature extraction on input data to obtain a feature F; constructing a decompression network symmetrical to the structure of the compression network, the output dimension of the decompression network being equal to the input dimension of the compression network: taking the compressed characteristic F as an input of a decompression network, and performing decompression processing to obtain decompressed data X' (K), wherein K is 1,2 … K; calculating an error X of the amplitude spectrum data X (k) and the decompressed data X' (k)mseTraining the weight W and the bias b of the network parameters by adopting a self-adaptive time estimation algorithm, and finely adjusting the weight W and the bias b of the network by adopting an error back propagation algorithm after each iteration until the error of the data set is converged, namely determining the network parameters W and b, and completing the training and construction of the multilayer deep neural network;
and 3, extracting the characteristic base of the training set data: when the network parameters are optimal in the DNN training process, sequentially inputting single data in a training set, acquiring the characteristics of all the single data and unitizing to obtain a characteristic base FiI is 1,2 … n, n is the number of characteristics, and constitutes the characteristic base set { F1,F2…FnAnd each characteristic base in the characteristic base set corresponds to an elevation angle thetaiI is 1,2 … n, characteristic radicalThe elevation angle set corresponding to the set is { theta1,θ2…θn}；
Step 4, estimating the incoming wave direction of the test set data: carrying out data preprocessing on the test set data: normalizing the test set data after the frequency domain transformation by using the maximum value and the minimum value of the training set data to obtain a magnitude spectrum after normalization; inputting the normalized amplitude spectrum into a trained multilayer DNN network, compressing to obtain a test set data characteristic F ', projecting the test set data characteristic F' into a characteristic basis set of training set data, and determining an incoming wave angle:
Figure FDA0003508336520000011
wherein
Figure FDA0003508336520000012
Representing the feature F' of the test set in the data feature base F of the training setiAnd (4) calculating the height of the low elevation target by the upper projection, and completing the height measurement of the low elevation target of the meter-wave radar.
2. The meter-wave radar low elevation height measurement method based on the deep neural network as claimed in claim 1, wherein the multi-layer DNN training of step 2 comprises the following steps:
2.1, taking the normalized amplitude spectrum X as the input of a compression network, performing feature extraction on the input amplitude spectrum data to obtain a feature F, and simultaneously constructing another decompression network with a symmetrical structure to perform decompression processing on the feature F obtained by compression to obtain decompressed data X' (k);
2.2 training the weight W and the bias b of the network parameters by adopting a self-adaptive time estimation algorithm, and adopting a modified linear unit ReLU as a network activation function, wherein the network activation function ReLU is defined as follows:
Figure FDA0003508336520000021
calculating the average of the decompressed data X' (k) and the amplitude spectrum data X (k)Square error Xmse：
Figure FDA0003508336520000022
And after each iteration, the BP algorithm is utilized to finely adjust the network weight W and the bias b, so that the network weight is optimal, and the error X ismseMinimum; when iteration is completed, the dimension of the feature F obtained by compression is usually smaller than that of the amplitude spectrum data X (k), and the feature F reserves the q-dimensional feature of the amplitude spectrum data X (k) on the assumption that the dimension of the feature F is q;
2.3 the deep neural network structure comprises an input layer, an output layer and a hidden layer, wherein the number of layers and the number of nodes of the hidden layer are adjusted according to a training result;
and 2.4, accelerating the training process, and processing the training set data in batches to finish the training and construction of the multilayer deep neural network.
3. The meter-wave radar low elevation height measurement method based on the deep neural network as claimed in claim 1, wherein the extracting of the training set data feature basis in step 3 specifically comprises:
when the weight W and the bias b of the DNN reach the optimum, sequentially inputting single data in a training set, acquiring the characteristics of all the single data and unitizing to obtain a characteristic base FiI is 1,2 … n, n is the number of characteristics, and constitutes the characteristic base set { F1,F2…FnConstructing a characteristic space of point sources at different elevation angles in the space domain, wherein each characteristic base in the characteristic base set corresponds to an elevation angle thetaiI is 1,2 … n, and the elevation angle set corresponding to the characteristic base set is { theta [ [ theta ] ]1,θ2…θn}。
4. The meter-wave radar low elevation height measurement method based on the deep neural network as claimed in claim 1, wherein the step 4 of calculating the low elevation target height specifically comprises:
the test set sample only needs to extract the characteristics of the sample, and does not need to decompress the characteristics, so only half of the network is needed to directly output the characteristics F ', and the F' is projected on the trainingDetermining an incoming wave angle in a feature space of a set:
Figure FDA0003508336520000031
the target height h is calculated.
5. The meter-wave radar low elevation height measurement method based on the deep neural network as claimed in claim 2, wherein the number of layers and the number of nodes of the hidden layer in step 2.3 are adjusted according to a training result, specifically:
adjusting the number of layers and the number of nodes of a hidden layer of the deep neural network according to a training result, and reducing the number of layers and the number of nodes of the hidden layer when the number of layers of the hidden layer is too large and the number of nodes is too large; when the number of layers of the hidden layer is too small and the number of nodes is too small, the number of layers and the number of nodes of the hidden layer need to be increased; the number of layers and the number of nodes of the hidden layer of the neural network are adjusted to ensure that the fitting effect of the decompressed data X' (k) and the amplitude spectrum data X (k) is the best: assuming that the total number of layers of the network is 5, the number of nodes is 256 × 128 × 64 × 128 × 256, wherein the hidden layer with the number of nodes 64 is the feature output layer, i.e., the dimension q of the feature F is 64.
CN201810650951.1A 2018-06-22 2018-06-22 Meter-wave radar low elevation height measurement method based on deep neural network Active CN108828547B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201810650951.1A CN108828547B (en) 2018-06-22 2018-06-22 Meter-wave radar low elevation height measurement method based on deep neural network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201810650951.1A CN108828547B (en) 2018-06-22 2018-06-22 Meter-wave radar low elevation height measurement method based on deep neural network

Publications (2)

Publication Number Publication Date
CN108828547A CN108828547A (en) 2018-11-16
CN108828547B true CN108828547B (en) 2022-04-29

Family

ID=64137452

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201810650951.1A Active CN108828547B (en) 2018-06-22 2018-06-22 Meter-wave radar low elevation height measurement method based on deep neural network

Country Status (1)

Country Link
CN (1) CN108828547B (en)

Families Citing this family (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109597046B (en) * 2018-11-29 2022-11-18 西安电子科技大学 Metric wave radar DOA estimation method based on one-dimensional convolutional neural network
CN109597047B (en) * 2018-11-29 2022-11-18 西安电子科技大学 Meter wave radar DOA estimation method based on supervised deep neural network
CN109597048B (en) * 2018-11-29 2023-08-11 西安电子科技大学 Meter wave radar DOA estimation method based on two-dimensional convolutional neural network
US20200278423A1 (en) * 2019-03-01 2020-09-03 GM Global Technology Operations LLC Removing false alarms at the beamforming stage for sensing radars using a deep neural network
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
CN110197666B (en) * 2019-05-30 2022-05-10 广东工业大学 Voice recognition method and device based on neural network
CN111310889B (en) * 2020-01-16 2022-04-12 西北工业大学 Evaporation waveguide profile estimation method based on deep neural network
CN111175693A (en) * 2020-01-19 2020-05-19 河北科技大学 Direction-of-arrival estimation method and direction-of-arrival estimation device
CN111929674B (en) * 2020-07-10 2022-10-04 西安电子科技大学 Intelligent amplitude comparison angle measurement method based on neural network, storage medium and equipment
CN112711909A (en) * 2020-12-30 2021-04-27 浙江大学 Method for rapidly calculating scattering amplitude function of basic three-dimensional shape body
CN115327505B (en) * 2022-10-12 2023-01-24 深圳大学 Low-altitude target angle estimation method in multipath environment and related equipment
IL300197B1 (en) * 2023-01-25 2024-04-01 Rafael Advanced Defense Systems Ltd Novel method and system for target height estimatiom using multipath corrupted radar measurements

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5768477A (en) * 1996-09-10 1998-06-16 Southwest Research Institute Radio direction finding system for narrowband multiple signals
DE59810720D1 (en) * 1998-11-09 2004-03-11 Atlas Elektronik Gmbh Device for determining the direction of incidence of temporally limited sound waves
CN102012505B (en) * 2010-10-15 2012-12-05 西安电子科技大学 Method for estimating direction of arrival of radar low-elevation target
CN108169708B (en) * 2017-12-27 2020-03-13 中国人民解放军战略支援部队信息工程大学 Direct positioning method of modular neural network

Also Published As

Publication number Publication date
CN108828547A (en) 2018-11-16

Similar Documents

Publication Publication Date Title
CN108828547B (en) Meter-wave radar low elevation height measurement method based on deep neural network
CN108549059B (en) Low-altitude target elevation angle estimation method under complex terrain condition
Xiang et al. Improved de-multipath neural network models with self-paced feature-to-feature learning for DOA estimation in multipath environment
CN106443621B (en) Coherent dynamic DOA method for tracing based on the orthogonal sparse reconstruct of matching
CN108872926B (en) Amplitude-phase error correction and DOA estimation method based on convex optimization
CN109597046B (en) Metric wave radar DOA estimation method based on one-dimensional convolutional neural network
CN109188344A (en) Based on mutually circulation correlation MUSIC algorithm information source number and arrival bearing's angular estimation method under impulse noise environment
CN109061554A (en) A kind of target angle of arrival estimation method updated based on spatial spreading grid dynamic
CN107991659B (en) Method for measuring height of low-elevation target of meter wave radar based on dictionary learning
CN108089147B (en) Improved short-wave single-station positioning method
Zhu et al. A deep learning architecture for broadband DOA estimation
CN110888105A (en) DOA estimation method based on convolutional neural network and received signal strength
CN108983169B (en) Meter wave radar terrain correction method based on digital elevation model
CN112612006A (en) Airborne radar non-uniform clutter suppression method based on deep learning
CN108089146B (en) High-resolution broadband direction-of-arrival estimation method for pre-estimated angle error robustness
CN113671485B (en) ADMM-based two-dimensional DOA estimation method for meter wave area array radar
Ma et al. A novel DOA estimation for low-elevation target method based on multiscattering center equivalent model
CN113466782A (en) Deep Learning (DL) -based cross-coupling correction D O A estimation method
Liu et al. Single snapshot DOA estimation based on spatial smoothing MUSIC and CNN
CN111239682B (en) Electromagnetic emission source positioning system and method
CN114791583A (en) Low elevation angle target DOA estimation method based on sum/difference mode of deep neural network
CN109633635B (en) Meter wave radar height measurement method based on structured recursive least squares
Hu et al. Off-grid DoA estimation based on compressed sensing on multipath environment
CN116539643B (en) Method and system for identifying coal rock data by using radar
Xu et al. A low computational complexity DOA estimation using sum/difference pattern based on DNN

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