CN111313943A - Three-dimensional positioning method and device under deep learning assisted large-scale antenna array - Google Patents
Three-dimensional positioning method and device under deep learning assisted large-scale antenna array Download PDFInfo
- Publication number
- CN111313943A CN111313943A CN202010105977.5A CN202010105977A CN111313943A CN 111313943 A CN111313943 A CN 111313943A CN 202010105977 A CN202010105977 A CN 202010105977A CN 111313943 A CN111313943 A CN 111313943A
- Authority
- CN
- China
- Prior art keywords
- positioning
- dimensional
- information
- angle
- time delay
- 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.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 50
- 238000013135 deep learning Methods 0.000 title claims abstract description 30
- 239000011159 matrix material Substances 0.000 claims abstract description 51
- 238000013528 artificial neural network Methods 0.000 claims abstract description 43
- 238000000605 extraction Methods 0.000 claims abstract description 40
- 238000013507 mapping Methods 0.000 claims abstract description 7
- 238000011176 pooling Methods 0.000 claims description 20
- 238000012549 training Methods 0.000 claims description 18
- 230000004044 response Effects 0.000 claims description 17
- 230000008569 process Effects 0.000 claims description 7
- 238000004590 computer program Methods 0.000 claims description 6
- 238000012935 Averaging Methods 0.000 claims description 3
- 230000004807 localization Effects 0.000 claims description 2
- 238000012545 processing Methods 0.000 abstract description 5
- 238000010586 diagram Methods 0.000 description 6
- 238000003491 array Methods 0.000 description 3
- 238000004364 calculation method Methods 0.000 description 3
- 230000006870 function Effects 0.000 description 3
- 230000004913 activation Effects 0.000 description 2
- 230000005540 biological transmission Effects 0.000 description 2
- 125000004122 cyclic group Chemical group 0.000 description 2
- 238000013461 design Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000010606 normalization Methods 0.000 description 2
- 230000009466 transformation Effects 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004422 calculation algorithm Methods 0.000 description 1
- 238000004891 communication Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 238000001914 filtration Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 238000010295 mobile communication Methods 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
Images
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04B—TRANSMISSION
- H04B7/00—Radio transmission systems, i.e. using radiation field
- H04B7/02—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas
- H04B7/04—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas
- H04B7/0413—MIMO systems
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04W—WIRELESS COMMUNICATION NETWORKS
- H04W64/00—Locating users or terminals or network equipment for network management purposes, e.g. mobility management
Landscapes
- Engineering & Computer Science (AREA)
- Computer Networks & Wireless Communication (AREA)
- Signal Processing (AREA)
- Mobile Radio Communication Systems (AREA)
- Position Fixing By Use Of Radio Waves (AREA)
Abstract
The invention discloses a three-dimensional positioning method and a three-dimensional positioning device under a large-scale antenna array assisted by deep learning. The invention utilizes the characteristics of high time delay and high angular resolution of a large-scale MIMO-OFDM system to construct an angle time delay domain channel energy matrix as the positioned position fingerprint information through signal processing. Then, a positioning neural network based on a three-dimensional convolution kernel is designed to realize positioning based on fingerprints, and the positioning neural network comprises a basic feature extraction module, a high-order feature extraction module and a regression module, so that mapping from position fingerprint information to three-dimensional coordinates of a user is realized. In an off-line stage, the position fingerprint information and the position information under the high signal-to-noise ratio or ideal environment are utilized to train a positioning neural network; in an online stage, the sparsity of an angle time delay domain channel energy matrix is utilized to remove noise from position fingerprint information, and then the position fingerprint information is input into a trained positioning neural network to estimate the three-dimensional coordinates of a user. The invention has the advantages of high positioning precision, low complexity, easy realization and the like.
Description
Technical Field
The invention relates to a three-dimensional positioning method under a large-scale antenna array assisted by deep learning, in particular to a three-dimensional positioning method and a three-dimensional positioning device based on a three-dimensional convolution neural network and by utilizing the channel characteristics of the large-scale antenna array.
Background
Services based on geographical location, as well as a wide arrangement of devices, place demands on accurate positioning. Traditional global positioning systems (such as GPS, beidou, galileo, etc.) can realize relatively accurate positioning under the condition of line of sight. However, in the environment with many obstacles such as urban buildings, the traditional global positioning system will generate serious errors due to the shielding and reflection caused by buildings, vehicles, pedestrians, etc.
In non-line-of-sight environments, fingerprint positioning methods using position fingerprint information have received much attention. Fingerprint positioning is a matching-positioning method, which includes off-line and on-line stages. In the off-line phase, the location reference point is selected, and the fingerprint information and the coordinate information are collected in the database. In the online phase, the location of the user is estimated by matching the fingerprint of the user's mobile terminal with the fingerprints in the database. Most fingerprinting methods use Received Signal Strength (RSS) based measurement information, which is affected by the heterogeneity of hardware and environmental changes. The channel state information has the potential of improving the accuracy of the fingerprint positioning method, but in some scenarios, such as a wireless sensor network and a WiFi network, the channel state information cannot capture sufficient multipath information, and therefore, the improvement of the performance of the fingerprint positioning method is limited.
Massive MIMO is a key technology in fifth-generation mobile communication technology, and due to the configuration of massive antenna arrays and the use of wide bandwidth, channel state information can capture rich multi-path information such as energy, angle and time delay. Meanwhile, a one-dimensional or two-dimensional large-scale antenna array is arranged on the base station side, so that high angular resolution can be realized in the vertical direction and the horizontal direction, and three-dimensional positioning in space is facilitated.
When channel information in a broadband large-scale MIMO system is used as a position fingerprint, the existing wireless positioning method based on the fingerprint information has the problems of high computational complexity and storage overhead. Meanwhile, the noise factor also affects the accuracy of the existing positioning method.
Disclosure of Invention
The purpose of the invention is as follows: the invention aims to provide a three-dimensional positioning method and a three-dimensional positioning device under a large-scale antenna array assisted by deep learning, which can overcome the defects in the prior art.
The technical scheme is as follows: in order to achieve the purpose, the invention adopts the following technical scheme:
the invention relates to a three-dimensional positioning method under a deep learning assisted large-scale antenna array, which comprises the following steps:
establishing a position fingerprint database, wherein the database stores position fingerprint information of a positioning reference point and corresponding three-dimensional coordinate information; the position fingerprint information is an angle delay domain channel energy matrix of a statistical channel, subscripts of elements of the position fingerprint information respectively represent angle and time delay, and values of the elements represent the magnitude of channel energy on the corresponding angle and time delay;
constructing a positioning neural network based on deep learning, wherein the positioning neural network is based on a three-dimensional convolution kernel and is formed by sequentially cascading a basic feature extraction module, a plurality of high-order feature extraction modules and a regression module; the input is a three-dimensional angle delay domain channel energy tensor processed by an angle delay domain channel energy matrix, the three dimensions of the tensor respectively represent a vertical angle, a horizontal angle and time delay, and the tensor is output as a three-dimensional coordinate;
training a positioning neural network, specifically, using position fingerprint information of a positioning reference point stored in a position fingerprint database as input, using coordinate information of the stored positioning reference point as a label, and training the network until convergence and storing a trained parameter;
predicting the position information of the user, specifically, collecting the current position fingerprint information of the user, removing noise by using the sparsity of the fingerprint, inputting the information into a trained positioning neural network, and outputting the result, namely, the estimated value of the three-dimensional coordinate of the position of the user.
Further, the acquiring process of the position fingerprint information is as follows: considering the uplink of a user to a base station, acquiring channel state information by channel estimation at the base station side by transmitting a pilot signal at a user side; acquiring a channel response matrix of a spatial frequency domain from the channel state information, and converting the channel response matrix of the spatial frequency domain into a channel response matrix of an angle time delay domain through inverse discrete Fourier transform; and multiplying each element of the angle time delay domain channel response matrix by the conjugate of the element and obtaining the expectation of the element to obtain the energy matrix of the angle time delay domain.
Furthermore, the position fingerprint database is established at the base station side, positioning reference points of the positioning area are traversed, pilot signals and current position information are transmitted at each positioning reference point, the base station side processes the received pilot signals into position fingerprint information, and the position fingerprint information and the received corresponding position information are recorded in the database.
Further, the positioning neural network uses a Three-dimensional Convolution-Normalization-Activation (3D CNA, Three-dimensional Convolution-Normalization-Activation) unit as a basis to construct Three main types of modules: the device comprises a basic feature extraction module, a high-order feature extraction module and a regression module.
The basic feature extraction module is used for extracting a basic feature map from a three-dimensional angle time delay domain channel energy tensor. The method comprises two parallel branches, wherein each branch comprises two 3D CNA layers and a maximum pooling layer, physical information of a vertical angle-time delay dimension and a horizontal angle-time delay dimension is extracted as a feature map tensor by setting an asymmetric three-dimensional convolution kernel and a pooling step length, branch outputs are cascaded along a channel dimension at the tail end, and a feature map based on the extraction of one layer of symmetric 3D CNA layer and one layer of symmetric maximum pooling layer is used as an output.
The high-order feature extraction module is used for further extracting a higher-order feature map from the feature map output obtained by the last module; the method comprises a plurality of parallel branches formed by 3D CNA layers, wherein each branch is provided with symmetrical convolution kernels with different sizes, different feature information is extracted, and the different feature information is integrated at an output end to be used as a higher-order feature map;
and the regression module is used for mapping the output of the high-order feature extraction module into a three-dimensional coordinate. The method comprises a global pooling layer and a full-connection layer, wherein the global pooling layer is used for averaging each input feature map, so that multidimensional tensor input is converted into a multidimensional vector, and the full-connection layer is converted into a three-dimensional coordinate vector.
Further, in the training of the positioning neural network, the used position fingerprint information is small in noise component under high signal-to-noise ratio or ideally noise-free.
Further, a process of removing noise when predicting the position information of the user: setting a corresponding threshold value by utilizing the sparsity of an angle time delay domain channel energy matrix according to a specific distribution mode of noise pollution and a signal-to-noise ratio in a channel environment, and setting elements smaller than the threshold value in the collected position fingerprint information (namely the angle time delay domain channel energy matrix) of the user mobile terminal to be 0.
Based on the same inventive concept, the three-dimensional positioning device under the deep learning assisted large-scale antenna array comprises:
the position fingerprint database is used for storing the position fingerprint information of the positioning reference point and the corresponding three-dimensional coordinate information; the position fingerprint information is an angle delay domain channel energy matrix of a statistical channel, subscripts of elements of the position fingerprint information respectively represent angle and time delay, and values of the elements represent the magnitude of channel energy on the corresponding angle and time delay;
the positioning neural network is based on a three-dimensional convolution kernel and is formed by sequentially cascading a basic feature extraction module, a plurality of high-order feature extraction modules and a regression module; the input is a three-dimensional angle delay domain channel energy tensor processed by an angle delay domain channel energy matrix, the three dimensions of the tensor respectively represent a vertical angle, a horizontal angle and time delay, and the tensor is output as a three-dimensional coordinate;
the network training module is used for inputting the position fingerprint information of the positioning reference points stored in the fingerprint database, using the coordinate information of the stored positioning reference points as labels, and training the network until convergence and storing the trained parameters;
and the position prediction module is used for collecting the current position fingerprint information of the user, removing noise by using the sparsity of the fingerprint, inputting the information into the trained positioning neural network, and outputting the result which is the estimated value of the three-dimensional coordinate of the position of the user.
Based on the same inventive concept, the three-dimensional positioning device under the deep learning assisted large-scale antenna array comprises a memory, a processor and a computer program which is stored on the memory and can be run on the processor, and is characterized in that when the computer program is loaded to the processor, the three-dimensional positioning method under the deep learning assisted large-scale antenna array is realized.
Has the advantages that: the invention discloses a three-dimensional positioning method under a deep learning-assisted large-scale antenna array, which has the following beneficial effects compared with the prior art:
1) by utilizing the high resolution of the large-scale MIMO system with multiple antennas and large broadband in angle and time delay, rich multipath channel information can be obtained, and the constructed angle time delay domain channel energy matrix is used as the second-order statistic of the channel, so that the position fingerprint information with high discrimination, stability and easy acquisition is provided for the geographical position of a user.
2) Due to the sparse characteristic of the angle time delay domain channel energy matrix, the noise component can be simply separated and removed, and the robustness of the position fingerprint to noise is ensured.
3) Compared with the traditional mathematical modeling method, the deep learning-based positioning neural network can accurately train the mapping relation between the user position fingerprint and the position coordinate, and avoids errors caused by a model.
4) Compared with a search-matching positioning algorithm, the neural network is used for positioning, the positioning accuracy is improved, and meanwhile, the calculation complexity and the storage cost of positioning are greatly reduced.
5) The positioning neural network uses a basic feature extraction module and a high-order feature extraction module based on a three-dimensional convolution kernel, so that the actual physical features of the position fingerprints can be extracted and converted into high-order features for positioning, and the positioning precision is improved.
6) The position fingerprint information is obtained from the channel state information of the user mobile terminal, does not need to occupy other communication resources, and is easy to realize in an actual system.
7) All storage resources and calculation resources required by positioning exist at the base station side, and the user mobile terminal only needs to transmit a pilot signal and receive position information, so that the calculation and storage resources of the user mobile terminal are not occupied, and the base station can uniformly manage and allocate the positioning resources.
Drawings
FIG. 1 is a flow chart of a method according to an embodiment of the present invention.
Fig. 2 is a schematic diagram of signal propagation in a channel from a user to a base station according to an embodiment of the present invention.
Fig. 3 is a schematic diagram of an angle delay domain channel energy matrix according to an embodiment of the present invention.
Fig. 4 is a block diagram of a basic feature extraction module according to an embodiment of the present invention.
Fig. 5 is a block diagram of a high-order feature extraction module according to an embodiment of the present invention.
Fig. 6 is a block diagram of a deep learning based positioning neural network according to an embodiment of the present invention.
Detailed Description
The technical solution of the present invention is further described with reference to the following specific embodiments.
In order to make the technical solutions of the present invention better understood, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
As shown in fig. 1, an embodiment of the present invention discloses a deep learning assisted three-dimensional positioning method under a large-scale antenna array, which utilizes the characteristics of high time delay and high angular resolution of a large-scale MIMO-OFDM system to construct an angle time delay domain channel energy matrix as positioning position fingerprint information through signal processing. Then, a positioning neural network based on a three-dimensional convolution kernel is designed to realize positioning based on fingerprints, and the positioning neural network comprises a basic feature extraction module, a high-order feature extraction module and a regression module, so that mapping from position fingerprint information to three-dimensional coordinates of a user is realized. In an off-line stage, the position fingerprint information and the position information under the high signal-to-noise ratio or ideal environment are utilized to train a positioning neural network; in an online stage, the sparsity of an angle time delay domain channel energy matrix is utilized to remove noise from position fingerprint information, and then the position fingerprint information is input into a trained positioning neural network to estimate the three-dimensional coordinates of a user.
In this embodiment, considering a massive MIMO uplink transmission system, as shown in fig. 2, a uniform area array is configured on the base station side, the number of each row of antennas is M, the number of each row of antennas is N, and the total number of antennas is NaMN. The base station side is actually a special case where the number M of antennas is 1 when the base station configures a uniform line array. The user terminal is provided with a single omnidirectional antenna. Assuming that the total number of users is K, the number of signal propagation multi-paths is Np. The arrival angles of the pth path of the user k in the vertical direction and the horizontal direction are respectively recorded as theta and 0 ≦ thetap,k≤π,Its array response vector can be expressed as:
wherein,
Considering OFDM modulation, the number of subcarriers is NcSampling interval of TsThe OFDM cyclic prefix length is Ng. Frequency of the l-th sub-carrierThe channel frequency response of user k on the ith subcarrier is:
whereinτp,kThe time delay of the p-th path for user k.Is the complex channel gain of the p-th path, satisfies the mean of 0 and the variance ofComplex gaussian distribution. The space-frequency domain response matrix of user k to the base station is defined as:
the three-dimensional positioning method under the deep learning assisted large-scale antenna array comprises the following steps:
s1, configuring a one-dimensional or two-dimensional large-scale antenna array at the base station side, and acquiring position fingerprint information in an uplink from a user to the base station through pilot frequency transmission, channel estimation, signal processing and the like.
And S2, selecting positioning reference points in the positioning area, wherein the interval between the reference points is a fixed constant.
And S3, establishing a position fingerprint database, wherein the database consists of position fingerprint information of the positioning reference point and corresponding three-dimensional coordinate information.
And S4, constructing a positioning neural network based on deep learning, wherein the positioning neural network is based on a three-dimensional convolution kernel and is formed by sequentially cascading a basic feature extraction module, three high-order feature extraction modules and a regression module.
And S5, training a positioning neural network, specifically, using the position fingerprint information of the positioning reference points stored in the position fingerprint database as input, using the coordinate information of the stored positioning reference points as labels, and training the network until convergence and storing the trained parameters.
And S6, predicting the position information of the user, specifically, collecting the current position fingerprint information of the user, removing noise by using the sparsity of the fingerprint, inputting the information into a stored positioning neural network, and outputting an output result which is an estimated value of the three-dimensional coordinate of the position of the user.
Further, the step S1 includes the steps of:
s1.1 DefinitionsAndis a phase-shifted DFT matrix whose (m, n) th elements are respectivelyAnddefinition ofFor DFT matrixFront N ofgColumn of (m, n) th element ofThrough inverse discrete Fourier transform, the space frequency domain channel response matrix HkTransforming into an angular time delay domain channel response matrix:
S1.2, multiplying each element of the angle time delay domain channel response matrix by the conjugate of the element, and obtaining the expectation to obtain the energy matrix of the angle time delay domain:
wherein ⊙ represents the Hadamard product (.)*Representing the conjugate of the matrix. As shown in fig. 3, subscripts of the angle delay domain channel energy matrix elements respectively represent angle and delay, and values of the elements represent magnitudes of channel energy at the corresponding angle and delay.
Further, the interval of the positioning reference points in the step S2 affects the positioning accuracy and the training effect of the network, and the smaller the interval is, the higher the positioning accuracy of the fingerprint method is theoretically, but the increased reference points may cause the network to be easy to over-fit, and on the contrary, the positioning accuracy based on the neural network fingerprint method is reduced, so the appropriate interval of the reference points should be determined according to the actual accuracy requirement and the network design.
Further, the database in step S3 is established on the base station side, an automatic acquisition device is used to traverse the positioning reference points of the positioning area, a pilot signal and current position information are transmitted at each positioning reference point, and the base station side processes the received pilot signal into position fingerprint information, and records the position fingerprint information in the database together with the received corresponding position information.
Further, the step S4 includes the steps of:
to describe the structure of the positioning neural network specifically, we consider configuring massive MIMO uniform area array antennas on the base station side, where the number of column antennas M is 8, the number of row antennas N is 16, and OFDM cyclic prefix N g128 case. For the difference of network design caused by configuring uniform linear arrays on the base station side, we will explain the corresponding steps later.
S4.1 in the input stage, processing the angle time delay domain channel energy matrix into a three-dimensional angle time delay domain channel energy tensor
[Xk]m,n,j=[Ωk]mN+n,j(8)
WhereinAnd the three-dimensional angle time delay domain channel energy tensor is expressed, and the three dimensions of the three-dimensional angle time delay domain channel energy tensor respectively represent a vertical angle, a horizontal angle and time delay. When the uniform linear array is configured at the base station side, the angle time delay domain channel energy matrix is still processed into a three-dimensional angle time delay domain channel energy tensor according to the formula (8)The only difference is that its first dimension is 1.
S4.2 the localizing neural network is denoted f (-) which maps XkMapping to three-dimensional space coordinates a of a userk(x, y, z), i.e.:
and using theta to represent trainable parameters in the positioning neural network, and expressing a loss function of the positioning neural network as MSE on a predicted value and a real value on a training set:
wherein N istrainIs the number of training set samples, and λ is the second term L2Weight of regularization.
S4.3 constructing a 3D CNA unit as a base unit for a localization neural network, which comprises three parts: three-dimensional convolution, normalized transformation and activation function. Assume that the input of the 3D CNA cell is noted asWhere H, W, L, P represents the height, width, length, and number of channels, respectively, of the input. Use ofA convolution kernel representing three dimensions, where K1、K2、K3Respectively representing the height, width and length of the convolution kernel, and Q is the number of output channels. And (3) convolving the input I with a three-dimensional convolution kernel K, setting the moving step length of the convolution kernel to be 1, and keeping the height, the width and the length of the output consistent with the input by using a zero filling mode. We denote the output of the three-dimensional convolution asThe (h, w, l, q) th element is represented as:
next, the three-dimensional convolved output is normalized by Batch Normalization (BN), and finally, the normalized output is activated by a Linear rectifying Unit (ReLU), and the outputs of the above two steps can be expressed as:
[T]h,w,l,q=max(0,BN([O]h,w,l,q)) (12)
wherein, BN (·) represents a batch normalization operation, and the tensor T is the output of the 3D CNA unit. For the condition that the uniform linear array is configured at the base station side, the height of a three-dimensional convolution kernel K in the 3D CNA is set as K1=1。
And S4.4, constructing a basic feature extraction module on the basis of the 3D CNA unit, and extracting a basic feature map from the three-dimensional angle time delay domain channel energy tensor. It contains two parallel branches, each branch containing two 3D CNA layers and one maximum pooling layer, as shown in fig. 4, the numbers in the block diagram describe the parameters of convolution and maximum pooling in 3D CNA, in the form of "size x number of channels" and "size/step", respectively. By setting an asymmetric three-dimensional convolution kernel and a pooling step length, extracting physical information of a vertical angle-time delay dimension and a horizontal angle-time delay dimension into feature map tensors, cascading branch output along a channel dimension at the tail end, and extracting feature maps based on a layer of symmetric 3D CNA layer and a layer of symmetric maximum pooling layer to serve as output. And for the condition that the uniform linear array is configured at the base station side, the maximum pooling height in the basic feature extraction module and the three-dimensional convolution kernel height in the 3D CNA are both set to be 1.
And S4.5, constructing a high-order feature extraction module on the basis of the 3D CNA unit, and extracting a high-order feature map from the output of the basic feature extraction module. It contains four parallel branches of 3D CNA layers, as shown in fig. 5, where n represents the coefficient of the number of convolution kernels. Each path is provided with a symmetrical convolution kernel with different size, different feature information is extracted, and the feature information is integrated at the output end to be used as a high-order feature map. And for the condition that the uniform linear array is configured on the base station side, the average pooling height in the high-order feature extraction module and the height of the convolution kernel of the 3D CNA are both set to be 1.
And S4.6, on the basis of the 3D CNA unit, constructing a regression module for mapping the output of the high-order feature extraction module into a three-dimensional coordinate. The method comprises a global pooling layer and a full-connection layer, wherein the global pooling layer is used for removing an activation function, averaging operation is carried out on each input feature map by the global pooling layer, so that multidimensional tensor input is converted into a multidimensional vector, and the full-connection layer converts the multidimensional vector into a three-dimensional coordinate vector.
S4.7, constructing a positioning neural network, as shown in FIG. 6, which is formed by sequentially cascading a basic feature extraction module, three high-order feature extraction modules and a regression module. The input is a three-dimensional angle time delay domain channel energy tensor, and the output is a three-dimensional user coordinate. In the case of the base station side arrangement of the uniform line arrays, the heights of the maximum pooling shown in fig. 6 are each set to 1.
Further, in the training of the positioning neural network in step S5, the position fingerprint information used is small noise component at high signal-to-noise ratio or ideally noise-free.
Further, the process of removing noise in step S6: by utilizing the sparsity of the angle delay domain channel energy matrix, as shown in fig. 2, a corresponding threshold value is set according to a specific distribution mode of noise pollution and a signal-to-noise ratio in a channel environment, and an element smaller than the threshold value in the collected fingerprint information (i.e., the angle delay domain channel energy matrix) of the user mobile terminal is set to be 0, which specifically includes the following steps:
s6.1 the channel obtained by channel estimation in the actual positioning environment often has noise pollution. The space-frequency domain response matrix of the k user obtained by the channel estimation is assumed to beProcessing the channel response matrix into an expanded angle time delay domain channel response matrix through Fourier inverse transformation:
further, the extended angle delay domain channel energy matrix is obtained by the following formula:
s6.2, estimating the average energy of noise pollution by using the extended angle time delay domain channel energy matrix:
s6.3, filtering out by setting a threshold valueIn the method, the noise pollution is avoided, and an angle time delay domain channel energy matrix with the noise removed is obtainedThe following were used:
for all i ═ 0,1, …, Na-1 and j ═ 0,1, …, Ng-1,
Wherein,β are threshold coefficients depending on the particular distribution of noise pollution and the ambient signal-to-noise ratio.
The embodiment of the invention discloses a three-dimensional positioning device under a large-scale antenna array assisted by deep learning, which comprises: the position fingerprint database is used for storing the position fingerprint information of the positioning reference point and the corresponding three-dimensional coordinate information; the position fingerprint information is an angle delay domain channel energy matrix of a statistical channel, subscripts of elements of the position fingerprint information respectively represent angle and time delay, and values of the elements represent the magnitude of channel energy on the corresponding angle and time delay; the positioning neural network is based on a three-dimensional convolution kernel and is formed by sequentially cascading a basic feature extraction module, a plurality of high-order feature extraction modules and a regression module; the input is a three-dimensional angle delay domain channel energy tensor processed by an angle delay domain channel energy matrix, the three dimensions of the tensor respectively represent a vertical angle, a horizontal angle and time delay, and the tensor is output as a three-dimensional coordinate; the network training module is used for inputting the position fingerprint information of the positioning reference points stored in the fingerprint database, using the coordinate information of the stored positioning reference points as labels, and training the network until convergence and storing the trained parameters; and the position prediction module is used for collecting the current position fingerprint information of the user, removing noise by using the sparsity of the fingerprint, inputting the information into the trained positioning neural network, and outputting the result which is the estimated value of the three-dimensional coordinate of the position of the user.
The embodiment of the three-dimensional positioning device under the deep learning assisted large-scale antenna array and the embodiment of the three-dimensional positioning method under the deep learning assisted large-scale antenna array belong to the same inventive concept, and specific implementation details refer to the above method embodiments and are not repeated herein. In addition, based on the same inventive concept, the embodiment of the invention also discloses a three-dimensional positioning device under the deep learning assisted large-scale antenna array, which comprises a memory, a processor and a computer program which is stored on the memory and can run on the processor, wherein when the computer program is loaded to the processor, the three-dimensional positioning method under the deep learning assisted large-scale antenna array is realized.
Claims (8)
1. The three-dimensional positioning method under the deep learning assisted large-scale antenna array is characterized in that: the method comprises the following steps:
establishing a position fingerprint database, wherein the database stores position fingerprint information of a positioning reference point and corresponding three-dimensional coordinate information; the position fingerprint information is an angle delay domain channel energy matrix of a statistical channel, subscripts of elements of the position fingerprint information respectively represent angle and time delay, and values of the elements represent the magnitude of channel energy on the corresponding angle and time delay;
constructing a positioning neural network, wherein the positioning neural network is based on a three-dimensional convolution kernel and is formed by sequentially cascading a basic feature extraction module, a plurality of high-order feature extraction modules and a regression module; the input is a three-dimensional angle delay domain channel energy tensor processed by an angle delay domain channel energy matrix, the three dimensions of the tensor respectively represent a vertical angle, a horizontal angle and time delay, and the tensor is output as a three-dimensional coordinate;
training a positioning neural network, specifically, using position fingerprint information of a positioning reference point stored in a fingerprint database as input, using coordinate information of the stored positioning reference point as a label, and training the network until convergence and storing a trained parameter;
predicting the position information of the user, specifically, collecting the current position fingerprint information of the user, removing noise by using the sparsity of the fingerprint, inputting the information into a trained positioning neural network, and outputting the result, namely, the estimated value of the three-dimensional coordinate of the position of the user.
2. The deep learning assisted three-dimensional positioning method under large-scale antenna array according to claim 1, characterized in that: the position fingerprint information acquisition process comprises the following steps: considering the uplink of a user to a base station, acquiring channel state information by channel estimation at the base station side by transmitting a pilot signal at a user side; acquiring a channel response matrix of a spatial frequency domain from the channel state information, and converting the channel response matrix of the spatial frequency domain into a channel response matrix of an angle time delay domain through inverse discrete Fourier transform; and multiplying each element of the angle time delay domain channel response matrix by the conjugate of the element and obtaining the expectation of the element to obtain the energy matrix of the angle time delay domain.
3. The deep learning assisted three-dimensional positioning method under large-scale antenna array according to claim 1, characterized in that: the position fingerprint database is established at the base station side, positioning reference points of a positioning area are traversed, pilot signals and current position information are transmitted at each positioning reference point, the base station side processes the received pilot signals into position fingerprint information, and the position fingerprint information and the received corresponding position information are recorded in the database.
4. The deep learning assisted three-dimensional positioning method under large-scale antenna array according to claim 1, characterized in that: the localization neural network uses three-dimensional convolution-normalization-activation (3D CNA) units as a basis to build three main classes of modules: a basic feature extraction module, a high-order feature extraction module and a regression module;
the basic feature extraction module is used for extracting a basic feature map from a three-dimensional angle time delay domain channel energy tensor; the method comprises two parallel branches, wherein each branch comprises two 3D CNA layers and a maximum pooling layer, physical information of a vertical angle-time delay dimension and a horizontal angle-time delay dimension is extracted as a feature map tensor by setting an asymmetric three-dimensional convolution kernel and a pooling step length, the branch outputs are cascaded along a channel dimension at the tail end, and a feature map based on the extraction of one layer of symmetrical 3D CNA layer and one layer of symmetrical maximum pooling layer is used as an output;
the high-order feature extraction module is used for further extracting a higher-order feature map from the feature map output obtained by the last module; the method comprises a plurality of parallel branches formed by 3D CNA layers, wherein each branch is provided with symmetrical convolution kernels with different sizes, different feature information is extracted, and the different feature information is integrated at an output end to be used as a higher-order feature map;
the regression module is used for mapping the output of the high-order feature extraction module into a three-dimensional coordinate; the method comprises a global pooling layer and a full-connection layer, wherein the global pooling layer is used for averaging each input feature map, so that multidimensional tensor input is converted into a multidimensional vector, and the full-connection layer is converted into a three-dimensional coordinate vector.
5. The deep learning assisted three-dimensional positioning method under large-scale antenna array according to claim 1, characterized in that: in the training of the positioning neural network, the position fingerprint information is small in noise component under high signal-to-noise ratio or ideally noiseless.
6. The deep learning assisted three-dimensional positioning method under large-scale antenna array according to claim 1, characterized in that: procedure for removing noise when predicting location information of a user: setting a corresponding threshold value according to the specific distribution mode of noise pollution and the signal-to-noise ratio in the channel environment by using the sparsity of the angle time delay domain channel energy matrix, and setting elements smaller than the threshold value in the acquired angle time delay domain channel energy matrix of the user mobile terminal as 0.
7. Three-dimensional positioner under supplementary large-scale antenna array of deep learning, its characterized in that: the method comprises the following steps:
the position fingerprint database is used for storing the position fingerprint information of the positioning reference point and the corresponding three-dimensional coordinate information; the position fingerprint information is an angle delay domain channel energy matrix of a statistical channel, subscripts of elements of the position fingerprint information respectively represent angle and time delay, and values of the elements represent the magnitude of channel energy on the corresponding angle and time delay;
the positioning neural network is based on a three-dimensional convolution kernel and is formed by sequentially cascading a basic feature extraction module, a plurality of high-order feature extraction modules and a regression module; the input is a three-dimensional angle delay domain channel energy tensor processed by an angle delay domain channel energy matrix, the three dimensions of the tensor respectively represent a vertical angle, a horizontal angle and time delay, and the tensor is output as a three-dimensional coordinate;
the network training module is used for inputting the position fingerprint information of the positioning reference points stored in the fingerprint database, using the coordinate information of the stored positioning reference points as labels, and training the network until convergence and storing the trained parameters;
and the position prediction module is used for collecting the current position fingerprint information of the user, removing noise by using the sparsity of the fingerprint, inputting the information into the trained positioning neural network, and outputting the result which is the estimated value of the three-dimensional coordinate of the position of the user.
8. Three-dimensional positioning apparatus under a deep learning assisted large scale antenna array, comprising a memory, a processor and a computer program stored on the memory and executable on the processor, characterized in that the computer program, when loaded into the processor, implements the three-dimensional positioning method under a deep learning assisted large scale antenna array according to any of claims 1-6.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010105977.5A CN111313943A (en) | 2020-02-20 | 2020-02-20 | Three-dimensional positioning method and device under deep learning assisted large-scale antenna array |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010105977.5A CN111313943A (en) | 2020-02-20 | 2020-02-20 | Three-dimensional positioning method and device under deep learning assisted large-scale antenna array |
Publications (1)
Publication Number | Publication Date |
---|---|
CN111313943A true CN111313943A (en) | 2020-06-19 |
Family
ID=71147303
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202010105977.5A Pending CN111313943A (en) | 2020-02-20 | 2020-02-20 | Three-dimensional positioning method and device under deep learning assisted large-scale antenna array |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN111313943A (en) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111931121A (en) * | 2020-09-21 | 2020-11-13 | 南昌大学 | Distributed positioning and speed measuring method based on large-scale antenna array |
CN112180326A (en) * | 2020-09-21 | 2021-01-05 | 南昌大学 | Layered distributed positioning and speed measuring method based on large-scale antenna array |
CN112995892A (en) * | 2021-02-08 | 2021-06-18 | 东南大学 | Large-scale MIMO fingerprint positioning method based on complex neural network |
Citations (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CA2442950A1 (en) * | 2003-09-26 | 2005-03-26 | Chahe Nerguizian | Method and system for indoor geolocation using an impulse response fingerprinting technique |
US7333774B2 (en) * | 2003-10-07 | 2008-02-19 | The Board Of Trustees Of The Leland Stanford Junior University | Method of optimizing wireless communication links using stored channel characteristics of different locations |
US20160277899A1 (en) * | 2015-03-18 | 2016-09-22 | Sony Corporation | Fingerprint matching in massive mimo systems |
CN107592611A (en) * | 2017-09-11 | 2018-01-16 | 东南大学 | The extensive mimo system wireless location method in broadband and system |
CN109411890A (en) * | 2018-08-13 | 2019-03-01 | 成都旸谷信息技术有限公司 | A kind of single station high-precision three-dimensional localization method based on multiple antennas |
CN109922427A (en) * | 2019-03-06 | 2019-06-21 | 东南大学 | Utilize the intelligent radio positioning system and method for large scale array antenna |
CN110381436A (en) * | 2019-06-25 | 2019-10-25 | 东南大学 | A kind of quick fingerprint positioning method based on the mono- station system of extensive MIMO |
CN110418278A (en) * | 2019-07-25 | 2019-11-05 | 李印 | A kind of 3 D positioning system based on Evolutionary Neural Network |
US20190369204A1 (en) * | 2018-05-31 | 2019-12-05 | Electronics And Telecommunications Research Institute | Position estimation method for estimating position of interference signal source and position estimation system for performing the method |
-
2020
- 2020-02-20 CN CN202010105977.5A patent/CN111313943A/en active Pending
Patent Citations (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CA2442950A1 (en) * | 2003-09-26 | 2005-03-26 | Chahe Nerguizian | Method and system for indoor geolocation using an impulse response fingerprinting technique |
US7333774B2 (en) * | 2003-10-07 | 2008-02-19 | The Board Of Trustees Of The Leland Stanford Junior University | Method of optimizing wireless communication links using stored channel characteristics of different locations |
US20160277899A1 (en) * | 2015-03-18 | 2016-09-22 | Sony Corporation | Fingerprint matching in massive mimo systems |
CN107592611A (en) * | 2017-09-11 | 2018-01-16 | 东南大学 | The extensive mimo system wireless location method in broadband and system |
US20190369204A1 (en) * | 2018-05-31 | 2019-12-05 | Electronics And Telecommunications Research Institute | Position estimation method for estimating position of interference signal source and position estimation system for performing the method |
CN109411890A (en) * | 2018-08-13 | 2019-03-01 | 成都旸谷信息技术有限公司 | A kind of single station high-precision three-dimensional localization method based on multiple antennas |
CN109922427A (en) * | 2019-03-06 | 2019-06-21 | 东南大学 | Utilize the intelligent radio positioning system and method for large scale array antenna |
CN110381436A (en) * | 2019-06-25 | 2019-10-25 | 东南大学 | A kind of quick fingerprint positioning method based on the mono- station system of extensive MIMO |
CN110418278A (en) * | 2019-07-25 | 2019-11-05 | 李印 | A kind of 3 D positioning system based on Evolutionary Neural Network |
Non-Patent Citations (5)
Title |
---|
CHI WU ET AL: "Learning to Localize: A 3D CNN Approach to User Positioning in Massive MIMO-OFDM Systems", 《ARXIV:1910.12378V2 EESS.SP》 * |
JOAO VIEIRA ET AL: "Deep Convolutional Neural Networks for Massive MIMO Fingerprint-Based Positioning", 《2017 IEEE 28TH ANNUAL INTERNATIONAL SYMPOSIUM ON PERSONAL, INDOOR, AND MOBILE RADIO COMMUNICATIONS》 * |
SEONGAH JEONG ET AL: "Optimization of Massive Full-Dimensional MIMO for Positioning and Communication", 《IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS》 * |
XIAOYU SUN ET AL: "Fingerprint-Based Localization for Massive MIMO-OFDM System With Deep Convolutional Neural Networks", 《IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY》 * |
刘敏 等: "改进型神经网络在室内三维定位中的应用研究", 《计算机应用研究》 * |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111931121A (en) * | 2020-09-21 | 2020-11-13 | 南昌大学 | Distributed positioning and speed measuring method based on large-scale antenna array |
CN112180326A (en) * | 2020-09-21 | 2021-01-05 | 南昌大学 | Layered distributed positioning and speed measuring method based on large-scale antenna array |
CN112180326B (en) * | 2020-09-21 | 2023-11-21 | 南昌大学 | Hierarchical distributed positioning and speed measuring method based on large-scale antenna array |
CN112995892A (en) * | 2021-02-08 | 2021-06-18 | 东南大学 | Large-scale MIMO fingerprint positioning method based on complex neural network |
CN112995892B (en) * | 2021-02-08 | 2022-11-29 | 东南大学 | Large-scale MIMO fingerprint positioning method based on complex neural network |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Vieira et al. | Deep convolutional neural networks for massive MIMO fingerprint-based positioning | |
CN109655799B (en) | IAA-based covariance matrix vectorization non-uniform sparse array direction finding method | |
CN112712557B (en) | Super-resolution CIR indoor fingerprint positioning method based on convolutional neural network | |
CN112995892B (en) | Large-scale MIMO fingerprint positioning method based on complex neural network | |
CN111313943A (en) | Three-dimensional positioning method and device under deep learning assisted large-scale antenna array | |
CN111030952B (en) | Beam space channel estimation method and system of millimeter wave system | |
CN108387864B (en) | Method and device for calculating angle of arrival | |
CN108182474B (en) | Multi-target direct positioning method based on uncorrected array and neural network | |
CN112040400B (en) | Single-site indoor fingerprint positioning method based on MIMO-CSI, storage medium and equipment | |
CN109633538B (en) | Maximum likelihood time difference estimation method of non-uniform sampling system | |
CN108646213B (en) | Direct wave AOA (automatic optical inspection) judgment method in indoor multipath environment | |
CN110933628B (en) | Fingerprint indoor positioning method based on twin network | |
CN111965596A (en) | Low-complexity single-anchor node positioning method and device based on joint parameter estimation | |
WO2023000614A1 (en) | Wireless positioning parameter estimation method, apparatus and system, computer device, and storage medium | |
CN114386321A (en) | AOA and TOF joint estimation method and device for indoor positioning and storage medium | |
CN112637950A (en) | Fingerprint positioning method based on angle similarity | |
CN113595941A (en) | Deep learning compressed sensing large-scale MIMO channel estimation method and system | |
CN112929962A (en) | Positioning method, positioning device, computer equipment and storage medium | |
CN109709510A (en) | A kind of estimation method and system of coherent 2-d direction finding | |
WO2022166477A1 (en) | Positioning method and apparatus, base station, computer device, and storage medium | |
Wu et al. | Super-resolution TOA and AOA estimation for OFDM radar systems based on compressed sensing | |
CN108398659A (en) | A kind of Wave arrival direction estimating method that pencil of matrix is combined with rooting MUSIC | |
CN111610488A (en) | Method for estimating wave arrival angle of any array based on deep learning | |
CN114269014B (en) | Large-scale MIMO dynamic environment fingerprint positioning method based on domain adaptive network | |
CN115908547A (en) | Wireless positioning method based on deep learning |
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 | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20200619 |
|
RJ01 | Rejection of invention patent application after publication |