CN111479231B - An indoor fingerprint localization method for millimeter-wave massive MIMO system - Google Patents

Abstract

The invention discloses an indoor fingerprint positioning method for a millimeter wave massive MIMO system. This method is based on the characteristics of millimeter-wave massive MIMO systems, combined with machine learning technology, to achieve a coarse-to-fine positioning process. In the coarse-grained positioning scheme, AOA information is used as fingerprint features, combined with deep learning methods, to adaptively extract feature quantities, and build a fingerprint library to achieve coarse-grained positioning. In the fine-grained positioning scheme, in order to overcome the high complexity brought by multiple antennas, the averaged multipath CSI amplitude is selected as the fingerprint feature for positioning, and a dynamic weighted K-nearest neighbor algorithm based on spatial mapping is proposed to achieve more good positioning accuracy. Simulation shows that the present invention can greatly improve the positioning accuracy. Compared with the existing fingerprint-based positioning scheme, the present invention adopts a single base station, which has lower cost, adopts a hierarchical positioning model, reduces the complexity of positioning operation, and can also achieve higher positioning accuracy in larger positioning scenarios.

Description

An indoor fingerprint localization method for millimeter-wave massive MIMO system

technical field

The invention belongs to the technical field of communication, and in particular relates to an indoor fingerprint positioning method for a millimeter-wave massive MIMO system.

Background technique

With the rapid development of wireless technology and the Internet, the Internet of Things (IoT) has gradually become an indispensable part of people's daily life. Location-based services are one of the most attractive applications related to IoT. The rapid development of IoT technology has promoted the development of many applications of high-precision positioning services. For example, future self-driving cars must achieve centimeter-level positioning accuracy to ensure accurate and high-speed driving. For some indoor scenarios, such as large-scale smart factories, the picking or inventorying of goods by intelligent robots requires precise location information of goods to speed up the picking process. In addition, luggage tracking in airports, personnel positioning in underground mining, smart home management and physical health monitoring all require high-precision positioning.

The main positioning technology in outdoor areas is based on the Global Navigation Satellite System (GNSS) positioning system. However, in indoor areas, due to its complex radio propagation environment, GNSS cannot provide satisfactory positioning performance. Therefore, in recent years, the indoor positioning technology based on radio signals has developed rapidly. Of these signals, the most commonly used are WiFi and Bluetooth in current smart devices. As one of the potential solutions for indoor positioning by radio signals, fingerprint positioning technology has attracted extensive attention from academia and industry. The fingerprint positioning method consists of two stages: offline training stage and online positioning stage. In the offline stage, professionals sample the location of the location area, and collect wireless signal characteristics at each sampling location, and store them in the location-fingerprint database. In the online positioning stage, the user sends the wireless signal fingerprint of his location to the positioning server, the server matches the query fingerprint with the database, and returns the position corresponding to the most similar fingerprint as the user's position estimate and returns it to the user.

As far as fingerprint positioning is concerned, received signal strength (RSS) is the most commonly used because of its simplicity and ease of measurement. However, RSS is sensitive to environmental changes due to shadow fading and multipath effects. In addition, RSS can only reflect rough channel information. Unlike RSS, channel state information (CSI) can also reflect channel characteristics such as shadow fading and multipath effects experienced by the received signal. By extracting CSI, fine-grained physical layer information of the system can be obtained, which is why CSI has been widely used in indoor fingerprint positioning in recent years. Additionally, millimeter wave (mmWave) technology is considered one of the key technologies for 5G networks and can be leveraged to determine the real-time location of vehicles. The high bandwidth and high path loss characteristics of millimeter waves can better observe the frequency selective fading caused by multipath in the frequency domain, which corresponds to the time domain and improves the resolution of the receiver for multipath signals. This characteristic helps It is used to improve the accuracy of positioning technology based on time, frequency difference, and RSS. In addition, millimeter waves can provide a larger effective signal bandwidth in high frequency bands, which improves the theoretical circle of positioning accuracy.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide an indoor fingerprint positioning method for a millimeter-wave massive MIMO system. The method adopts the fingerprint positioning technology, uses the CSI amplitude information and the AOA information obtained by CSI estimation as fingerprint features, and divides coarse-grained and fine-grained positioning in two stages to achieve coarse-to-fine positioning.

To achieve the above object, the present invention adopts the following technical solutions to realize:

An indoor fingerprint positioning method for a millimeter-wave massive MIMO system, a receiving end obtains a time-domain separable CSI multipath signal of the millimeter-wave massive MIMO system, and performs data collection and preprocessing on the CSI multipath signal; In the positioning scheme, the AOA information estimated by the multi-path CSI signal is used as the fingerprint feature, combined with the deep learning method, the feature quantity is adaptively extracted, and the fingerprint library is built to realize the coarse-grained positioning; in order to overcome the high complexity brought by multiple antennas , in the fine-grained positioning scheme, the averaged multi-path CSI information is used as the fingerprint feature for positioning; at the same time, considering that the traditional WKNN algorithm largely relies on the pre-selected fixed K value, the dynamic weighted K-nearest neighbor algorithm is used. , to achieve better positioning accuracy.

A further improvement of the present invention is that the method is specifically as follows:

First, in the coarse-grained positioning stage, the classical MUSIC algorithm is used to estimate the multi-path AOA information from the collected multi-path CSI fingerprint data. Fingerprint library, and then through the softmax classifier, the position label with the largest output probability is used as the final estimated position of the coarse-grained positioning stage;

Then, in the fine-grained localization stage, on the basis of the estimated coarse-grained localization labels, fine-grained labels are generated based on the estimated coarse-grained localization labels. The value information is used as the fingerprint feature in the fine-grained localization stage, and a dynamic weighted K-nearest neighbor algorithm based on spatial mapping is proposed for fine-grained localization;

Finally, in order to improve the positioning accuracy, in the fine-grained stage, the following fact is considered: the same CSI difference may correspond to different geometric distances, that is, for CSI differences of different magnitudes, the signal distances it represents may correspond to different geometric distances; therefore, the relationship between the signal distance space and the geometric distance space is established by training the ELM to achieve spatial mapping and prevent the influence of distance mismatch which leads to the degradation of localization accuracy.

A further improvement of the present invention is that the data collection and preprocessing in the coarse-grained positioning stage specifically include the following steps:

Step 1: Divide the cell into N ₁ blocks evenly, and use its geometric center as the classification position label;

Step 2: In the coarse-grained positioning stage, since the selected coarse-grained positioning labels are sparsely distributed and small in number, the multi _- path CSI information collected in the divided N1 blocks is estimated by the classical MUSIC estimation algorithm to obtain all coarse-grained positioning labels. Multi-path AOA information of the location tag point, and use it as a fingerprint feature;

Step 3: Add the sample features to the label, and the sample in the coarse-grained positioning stage is represented as:

为第N_ray条散射路径的AOA，T为矩阵转置。where φ ₀ is the AOA of the direct path, φ ₁ is the AOA of the first scattering path,

is the AOA of the Nth _ray scattering path, and T is the matrix transpose.

A further improvement of the present invention is that the task of the coarse-grained positioning stage is to train the parameters of the deep convolutional neural network and the regression classifier network according to the obtained labeled training data, and the training goal is to make the training label and the network output The mean square error is the smallest;

The offline training process of the coarse-grained positioning link is as follows:

For the deep learning network, a deep convolutional neural network is used, and the excitation function of each layer node adopts the ReLU function. After the training data is input into the network, the output of each layer is obtained according to the excitation function, which is used as the input of the next layer, and is propagated forward layer by layer. , and finally get the network output; construct the penalty function according to the principle of minimum mean square error and use the stochastic gradient descent algorithm to update and iterate to obtain the final training parameters, and the weights W, b after training are stored as part of the fingerprint library;

Then the neural network training output data is used as the input of the softmax classifier, and then it is divided into C classes, the probability of the input data belonging to each class is used as the output of the classifier, the penalty function is constructed according to the principle of minimum mean square error and stochastic gradient descent is used. The algorithm is updated and iterated to obtain the final training parameters, and W, b, θ are combined to form a fingerprint library, and θ is the classifier parameter.

A further improvement of the present invention is that the output of the classifier is as follows:

是一个C×1的矩阵，每一项表示在

给定的情况下，

属于每一类的概率，

为神经网络的训练输出，即回归分类器的输入，θ为分类器的参数。in,

is a C × 1 matrix, each item is represented in

Given the circumstances,

the probability of belonging to each class,

is the training output of the neural network, that is, the input of the regression classifier, and θ is the parameter of the classifier.

A further improvement of the present invention is that the online stage positioning process of the coarse-grained positioning link is as follows:

经过机器学习网络正向传播与回归分类器分类，得到该未知数据属于每一待定位置的概率；Step 1: After receiving the CSI information from the unknown location user, after the AOA information is estimated by the MUSIC algorithm, that is,

After the forward propagation of the machine learning network and the classification of the regression classifier, the probability that the unknown data belongs to each undetermined position is obtained;

Step 2: Using a probabilistic method, the position label with the highest output probability is used as the final position estimate for coarse-grained positioning.

A further improvement of the present invention is that the data collection and preprocessing in the fine-grained positioning stage specifically includes the following steps:

Step 1: On the basis of the coarse-grained positioning results, take it as the core, and expand outward at equal intervals to generate fine-grained labels, which are called extended sliding windows. According to this method, N ₂ fine-grained location labels are generated in the sliding window. ;

Step 2: For the fine-grained positioning stage, in order to further reduce the data dimension and reduce the computational complexity, the CSI data of each path is sampled, and then the amplitude channel matrices on different antennas are averaged, and based on the known CSI information and The corresponding relationship of user positions, grouping and numbering the amplitude data;

Step 3: Add the sample features to the label, and the samples in the fine-grained positioning stage are represented as:

N₂为细粒度定位阶段的位置标签数。in,

N ₂ is the number of location labels in the fine-grained localization stage.

A further improvement of the present invention is that the task of the fine-grained positioning stage is based on the coarse-grained positioning, with the estimated position as the center extending outward into a sliding window full of refined virtual labels; considering that there is a phase offset in the CSI, At the same time, in order to overcome the high complexity caused by multiple antennas, the averaged multipath CSI amplitude information is used as the location label, and an adaptive dynamic weighted K-nearest neighbor algorithm based on spatial mapping is used. The algorithm steps are as follows:

Step 1: Calculate the signal distance between all sample points in the fingerprint library:

Step 2: Delete the singular distance, set the threshold T=α×D ₁ , retain the signal distance not greater than the threshold T in the above formula, and denote it as D ₁ ,...,D _S in order from small to large, s=1,2 ,...,S, S represents the number of reserved adjacent reference points;

Step 3: Calculate the average distance difference of the reserved points from each other:

Among them, Δd _j,s represents the distance difference between D _j and D _s ;

Step 4: Estimate the final position, the dynamic weights are expressed as:

则最终的位置估计的具体形式如下：Wherein, ΔD=D _K -D ₁ ; in particular, when D _K =D ₁ , ω _j =1; definition:

The specific form of the final location estimation is as follows:

Among them, (x _j , y _j ) represents the coordinates of the jth reserved position label.

A further improvement of the present invention is that, in the proposed step (1) of the dynamic weighted K-nearest neighbor algorithm based on spatial mapping, when calculating the signal distance, considering the mismatch between the signal distance and the physical distance at different CSI amplitude levels In fact, using an extreme learning machine-based spatial mapping method, the problem is described as follows:

Among them, S _D and S _C are the signal space distance and the geometric space distance, respectively.

The present invention at least has the following beneficial technical effects:

A good positioning method not only pursues positioning accuracy, but also pursues the goal of meeting application requirements and adapting to environmental characteristics. Therefore, the present invention designs an indoor fingerprint for a millimeter-wave massive MIMO system by using a fingerprint location technology based on wireless signals for scenarios with a large positioning range, such as large indoor parking lots, large smart factories and other indoor scenarios. The positioning method realizes the positioning process from coarse to fine, and can achieve high positioning accuracy. The core of the present invention is that the receiving end obtains the CSI multipath signals separable in the time domain of the millimeter wave massive MIMO system, and performs data collection and preprocessing on the multipath signals. In the coarse-grained positioning scheme, the AOA information estimated by the multi-path CSI signal is used as the fingerprint feature, combined with the deep learning method, the feature quantity is adaptively extracted, and the fingerprint database is built to realize the coarse-grained positioning. In order to overcome the high complexity brought by multiple antennas, in the fine-grained localization scheme, the averaged multipath CSI information is used as the fingerprint feature for localization. At the same time, considering that the traditional WKNN algorithm relies heavily on the pre-selected fixed K value, a dynamic weighted K-nearest neighbor algorithm is designed to achieve better positioning accuracy. In general, the present invention has the following advantages:

1. The present invention proposes an indoor fingerprint positioning method based on a hierarchical model, which is used in a millimeter-wave massive MIMO system, realizes a coarse-to-fine positioning process, reduces system complexity, and adopts a single base station, reducing location cost.

2. The present invention uses multi-path AOA information as coarse-grained fingerprints, and uses convolutional neural network and softmax regression classifier to adaptively extract fingerprint features to construct a fingerprint database. Because the indoor environment is rich in multi-path, and the direct path may not exist in most cases, the accuracy of the traditional AOA estimation algorithm largely depends on the direct path, so the estimation accuracy in indoor scenes is greatly reduced, and the small error of AOA will cause Serious positioning errors, especially in scenes with a large positioning range. The invention uses the multi-path AOA information as the fingerprint feature, and utilizes the uniqueness of the multi-path AOA in different positions to realize fingerprint positioning, which not only overcomes the problem of high-precision AOA estimation in the multi-path environment, but also improves the positioning accuracy. In addition, the convolutional neural network shows excellent performance in video image classification. The present invention utilizes the convolutional neural network to achieve feature extraction and build a fingerprint library. Compared with other networks, such as BP network, the extracted features are more abundant, which is helpful for improving positioning performance.

3. In the fine-grained positioning stage of the present invention, in order to reduce the computational complexity, only the averaged amplitude of the channel state information is selected as the fingerprint feature, and the averaged multipath CSI information is used as the fingerprint feature for positioning. At the same time, considering that the traditional WKNN algorithm relies heavily on the fixed K value selected, a dynamic weighted K-nearest neighbor algorithm is proposed to achieve better localization accuracy. The traditional WKNN algorithm relies on a pre-selected fixed K value. The dynamic weighted K nearest neighbor algorithm proposed by the present invention enables each target to be assigned an appropriate K value during positioning, which greatly improves the positioning performance. In addition, the present invention also considers the problem of mismatch between the signal distance and the corresponding real physical distance in the algorithm, studies the mapping relationship between the two, and uses a machine learning method (ELM) to realize the mapping between the signal space and the physical space, So as to achieve high-precision positioning. As far as I know, this is the first time that a machine learning algorithm has been used to implement spatial mapping and achieve good localization performance.

4. The present invention is aimed at indoor scenes, especially for scenes with a large positioning range, such as large indoor factories, intelligent indoor parking lots, etc., the positioning advantages are more obvious, and the positioning accuracy is effectively improved. Simulations show that the method can achieve decimeter-level positioning accuracy in an indoor environment of 120×70×3m ³ .

Description of drawings

Figure 1 is a schematic diagram of a 3D indoor positioning scene;

Fig. 2 is a schematic diagram of a flow chart of a hierarchical positioning system;

3 is a schematic diagram of a data training process of a convolutional neural network;

4 is a schematic diagram of virtual labels in a fine-grained positioning scheme;

5 is a schematic diagram of the positioning effect of randomly selecting 120 test points;

Fig. 6 is the performance comparison between the present invention and the positioning scheme;

Figure 7 shows the influence of the number of refined virtual labels on the average positioning error;

Figure 8 shows the effect of the threshold coefficient α on the average positioning error.

Detailed ways

The technical solutions of the present invention will be further described in detail below through the accompanying drawings and embodiments.

The positioning environment is shown in Figure 1, which is a 3D indoor positioning environment, including a base station (Access point) with a known position, equipped with multiple antennas, and a device to be positioned (MP) with an unknown position as a receiver. Since in mmWave channels, the direct path (LOS) dominates, and the non-direct path (NLOS) is severely attenuated after one or two reflections. Therefore, only single-hop and double-hop reflections are considered in the considered localization environment. Since the receiver (MP) is located at a certain height above the ground, the ground reflections are considered in addition to the reflections from the four walls in a single hop.

See Figure 2 for the flow chart of the positioning system. For scenes with a large positioning range, the positioning process is divided into two stages: coarse-grained positioning (left in the figure) and fine-grained positioning (right in the figure).

First, in the coarse-grained positioning stage, the classical MUSIC algorithm is used to estimate the multi-path AOA information from the collected multi-path CSI fingerprint data, which is used as fingerprint features. Build the fingerprint library. Then through the softmax classifier, the position label with the largest output probability is used as the final estimated position in the coarse-grained localization stage.

Then, in the fine-grained localization stage, on the basis of the estimated coarse-grained localization labels, fine-grained labels are generated based on them. Therefore, the averaged CSI amplitude information is selected as the fingerprint feature in the fine-grained positioning stage, and a dynamic weighted K-nearest neighbor algorithm based on spatial mapping is proposed for fine-grained positioning.

1) The specific process of the coarse-grained positioning stage is as follows:

(1) The data collection and preprocessing in the coarse-grained positioning stage specifically includes the following steps:

Step 1: The cell is evenly divided into N ₁ blocks, and its geometric center is used as the classification position label. See Figure 4.

Step 2: In the coarse-grained positioning stage, since the selected coarse-grained positioning labels are sparsely distributed and the number is small, the multi _- path CSI information collected in the divided N1 blocks is estimated by the classical MUSIC algorithm to obtain all coarse-grained positioning labels. Multi-path AOA information of the location tag point, and use it as a fingerprint feature;

Step 3: Add sample features to labels. The samples in the coarse-grained positioning stage can be expressed as:

为第N_ray条散射路径的AOA，T为矩阵转置。where φ ₀ is the AOA of the direct path, φ ₁ is the AOA of the first scattering path,

is the AOA of the Nth _ray scattering path, and T is the matrix transpose.

(2) The offline training process of the coarse-grained positioning link is as follows:

For the deep learning network, a deep convolutional neural network is used, and the excitation function of each layer node adopts the ReLU function. After the training data is input into the network, the output of each layer is obtained according to the excitation function, which is used as the input of the next layer, and is propagated forward layer by layer. , and finally get the network output; construct the penalty function according to the principle of minimum mean square error and use the stochastic gradient descent algorithm to update and iterate to obtain the final training parameters, and the weights W, b after training are stored as part of the fingerprint library;

Then the neural network training output data is used as the input of the softmax classifier, and then it is divided into C classes, the probability of the input data belonging to each class is used as the output of the classifier, the penalty function is constructed according to the principle of minimum mean square error and stochastic gradient descent is used. The algorithm is updated and iterated to obtain the final training parameters, and W, b, θ are combined to form a fingerprint library, and θ is the classifier parameter.

The output of the classifier is as follows:

是一个C×1的矩阵，每一项表示在

给定的情况下，

属于每一类的概率，

为神经网络的训练输出，即回归分类器的输入，θ为分类器的参数。in,

is a C × 1 matrix, each item is represented in

Given the circumstances,

the probability of belonging to each class,

is the training output of the neural network, that is, the input of the regression classifier, and θ is the parameter of the classifier.

See Figure 3 for the data training process of the coarse-grained localization stage using a deep convolutional neural network (DCNN). First, get the AOA data as input, and estimate the AOA value for each location through MUSIC. Then, an AOA tensor is built for each position, which is simple and convenient for DCNN to process in all layers. In the first convolutional layer and downsampling layer, 18 convolution kernels of size 2×1 are used for convolution operation to obtain the same number of feature maps of the same size as the input. To ensure the invariance of feature maps, down-sampling is performed with a size of 2 × 1 to obtain the same number of feature maps. For all the layers in the designed DCNN, for simplicity, the convolutional and pooling layers use the same size of convolution kernels for term convolution and downsampling. The difference is the stride of the second and fourth pooling layers. In addition, it is worth noting that the Dropout layer is introduced into the designed DCNN, that is, during the training process, the weights of the DCNN will be temporarily discarded with a certain probability and will not be updated temporarily, which can effectively prevent overfitting.

(3) The online stage positioning process of the coarse-grained positioning link is as follows:

经过机器学习网络正向传播与回归分类器分类，得到该未知数据属于每一待定位置的概率。Step 1: After receiving the CSI information from the unknown location user, after the AOA information is estimated by the MUSIC algorithm, that is,

After the forward propagation of the machine learning network and the classification of the regression classifier, the probability that the unknown data belongs to each undetermined position is obtained.

Step 2: Using a simple probabilistic approach, use the location label with the highest output probability as the final location estimate for coarse-grained localization.

2) The task of the fine-grained localization stage is to extend outwards into a sliding window full of refined virtual labels with the estimated position as the center on the basis of coarse-grained localization. Considering the existence of phase offset in CSI, and to overcome the high complexity caused by multiple antennas, an adaptive dynamic weighted K-nearest neighbor algorithm based on spatial mapping is proposed by using the averaged multipath CSI amplitude information as the location label. Figure 4 shows the virtual label construction process in the fine-grained positioning scheme. Among them, the left of the figure is a schematic diagram of the distribution of coarse-grained location labels, and the five-pointed star is marked as a coarse-grained location label; the right of the figure is a schematic diagram of the distribution of fine-grained location labels. bit target position.

(1) The data collection and preprocessing in the fine-grained positioning stage specifically includes the following steps:

Step 1: On the basis of the coarse-grained positioning results, take it as the core, and expand the fine-grained labels outward at equal intervals, as shown in Figure 4, which is called the extended sliding window. According to this method, N ₂ fine-grained position labels are generated in the sliding window.

Step 2: For the fine-grained positioning stage, in order to further reduce the data dimension and reduce the computational complexity, the CSI data of each path is sampled, and then the amplitude channel matrices on different antennas are averaged, and based on the known CSI information and The corresponding relationship between user positions, grouping and numbering the amplitude data.

Step 3: Add sample features to labels. The samples in the fine-grained positioning stage can be expressed as:

N₂为细粒度定位阶段的位置标签数。以上便是两个阶段的输入数据。in,

N ₂ is the number of location labels in the fine-grained localization stage. The above is the input data of the two stages.

(2) Aiming at the problem that the traditional WKNN algorithm largely depends on the selected fixed K value, a dynamic weighted K-nearest neighbor algorithm based on spatial mapping is proposed in the fine-grained localization stage to achieve better localization accuracy. The algorithm steps are as follows:

Step 1: Calculate the signal distance between all sample points in the fingerprint library:

Step 2: Remove singular distances. Set the threshold value T=α×D ₁ , keep the signal distance not greater than the threshold value T in the above formula, and denote it as D ₁ ,...,D _S in order from small to large, s=1,2,...,S, S represents The number of adjacent reference points to keep.

Step 3: Calculate the average distance difference of the reserved points from each other:

Among them, Δd _j,s represents the distance difference between D _j and D _s .

Step 4: Estimate the location. The dynamic weight is expressed as:

则最终的位置估计的具体形式如下：Wherein, ΔD=D _K −D ₁ . In particular, when D _K =D ₁ , ω _j =1. definition:

The specific form of the final location estimation is as follows:

Among them, (x _j , y _j ) represents the coordinates of the jth reserved position label.

In step (1) of the proposed dynamic weighted K-nearest neighbor algorithm based on spatial mapping, when calculating the signal distance, considering the fact that the signal distance and the physical distance do not match at different CSI amplitude levels, a new algorithm based on A spatial mapping method for Extreme Learning Machine (ELM). Describe the problem as follows:

Among them, S _D and S _C are the signal space distance and the geometric space distance, respectively. The purpose is to try to find the mapping relationship between the two. Therefore, an adaptive spatial mapping method based on ELM is designed. To the best of our knowledge, this is the first time that machine learning is used for spatial mapping of signal-space distance and geometric-space distance, and the localization accuracy is effectively and significantly improved.

3) In order to verify the performance of the hierarchical fingerprint positioning method proposed by the present invention, the following simulations were carried out:

To better evaluate the performance of the proposed method, experiments are carried out in a 28 GHz mmWave channel with an indoor propagation environment with dimensions of 70 × 120 × ⁴ m. The location of the base station (AP) is set to (47.5m, 57.5m, 4.0m). A single base station with 256 antennas is used at the transmitter and 32 antennas are configured at the receiver. The distance between adjacent antennas and the radius of each uniform circular array are half the wavelength and twice the wavelength of the mmWave carrier frequency, respectively. In addition, for the millimeter-wave indoor propagation environment, the close-in free space reference distance path loss model is adopted, namely:

分别为路径损耗因子和阴影衰落，FSPL(f,d₀)＝20log₁₀(4πf/c)为自由空间路径损耗。where dl is used to denote the length of the _lth path, γ and

are the path loss factor and shadow fading, respectively, and FSPL(f,d ₀ )=20log ₁₀ (4πf/c) is the free-space path loss.

The performance of the experiment is mainly evaluated by three aspects: Root Mean Square Error Distance (RMSED), median error and Cumulative Distribution Function (CDF), where RMSED is the average of the actual position and the estimated position. Square error:

First, 400 test points were selected for testing, and 120 test points were randomly selected to draw the positioning effect map. See Figure 5, where the six-pointed star marks the actual position and the circle marks the estimated position. It can be clearly seen from the figure that the estimated positions of all test points are very close to the actual positions, or even coincide. Furthermore, it shows that the proposed hierarchical localization system has good localization performance.

Then, to further verify the effectiveness of the proposed scheme, the localization performance of different localization methods is compared. As shown in Table 1, the second column is the average error distance, the third column is the median error, and the last column is the maximum error distance. It can be seen from the table that the proposed positioning method can finally achieve a positioning error of 0.4738m. The median error is 0.2619m, and the maximum error distance is also significantly smaller than other comparison methods.

Table 1 Comparison of results of different positioning methods

The performance comparison between the present invention and other positioning schemes is shown in FIG. 6 . It can be seen from the figure that the proposed positioning system is significantly better than other algorithms, the positioning error of about 77% of the test points is below 0.5m, and the positioning error of 88% of the test points is below 1m. The test points with positioning errors within 1m of other algorithms are only 55%, 60% and 64% respectively.

See Figure 7 for the CDF comparison of the proposed algorithm under different number of refined labels (n=7, 9, 11, 13). It can be seen from the figure that with the increase of the number of refined labels in the sliding window, the CDF value gradually increases, indicating that the positioning effect is getting better and better. Especially when n=13, the test points with error within 1m account for about 97%. Furthermore, as n increases, the maximum error distance decreases by about 0.5m each time. This is because the fixed interval between adjacent refinement labels is set to 0.5m. When n changes from 7 to 9, the side length of the sliding window increases by 1m, but in fact the sliding window only expands by 0.5m in four directions. So the maximum error is reduced by about 0.5m. The above analysis shows that the error level can be controlled within a certain range by reasonably setting the number of refined labels.

Table 2 shows the comparison of the positioning effect under different number of refined labels. It can be seen that with the increase of n, the average positioning error, the median error and the maximum value all gradually decrease. When n=13, the mean error value is reduced to 0.3101m. This is because, as the number of refined labels increases, the size of the sliding window increases, so the localization performance will get better and better.

Table 2 Comparison of positioning effects under different number of refined labels

Number of refinement labels Mean(m) Median(m) Max(m) n=7 0.7006 0.3653 3.9613 n=9 0.4738 0.2619 3.4169 n=11 0.3729 0.2506 2.9777 n=13 0.3101 0.2315 2.4903

In the proposed SMWKNN algorithm, before calculating the average signal distance, the candidate labels are screened by setting a threshold, and the label points corresponding to the singular signal distance are deleted. Therefore, when evaluating the performance, the influence of the threshold coefficient α on the positioning error was compared, and in order to select the universally optimal threshold coefficient, experiments were carried out with different number of refined labels, see Figure 8. Obviously, when α=2, the localization performance is optimal for all the number of refined labels set in the experiment, so the threshold coefficient is chosen to be 2.

Therefore, it can be seen from the above that the indoor fingerprint positioning method based on the hierarchical model proposed by the present invention can effectively reduce the positioning complexity, improve the positioning accuracy, and at the same time reduce the positioning cost.

The above content is a further detailed description of the present invention in conjunction with the specific preferred embodiments, and it cannot be considered that the specific embodiments of the present invention are limited to this. Below, some simple deductions or substitutions can also be made, all of which should be regarded as belonging to the invention and the scope of patent protection determined by the submitted claims.

Claims (4)

1. An indoor fingerprint positioning method for a millimeter wave large-scale MIMO system is characterized in that a receiving end acquires a CSI multipath signal of the millimeter wave large-scale MIMO system, which can be separated in time domain, and performs data collection and preprocessing on the CSI multipath signal; in the coarse-grained positioning scheme, AOA information estimated by a multipath CSI signal is used as fingerprint characteristics, a deep learning method is combined, characteristic quantities are extracted in a self-adaptive mode, and a fingerprint database is established to achieve coarse-grained positioning; in order to overcome the high complexity caused by multiple antennas, in a fine-grained positioning scheme, positioning is carried out by taking averaged multipath CSI (channel state information) as fingerprint characteristics; meanwhile, considering the problem that the traditional WKNN algorithm depends on a preselected fixed K value to a great extent, a dynamic weighted K nearest neighbor algorithm is adopted to realize better positioning accuracy; the method comprises the following specific steps:

firstly, in a coarse-grained positioning stage, estimating acquired multipath CSI fingerprint data by using a classical MUSIC algorithm to obtain multipath AOA information as fingerprint characteristics, extracting the fingerprint characteristics in a self-adaptive manner through a multilayer convolutional neural network to construct a fingerprint database, and then using a position label with the maximum output probability as a final estimation position in the coarse-grained positioning stage through a softmax classifier;

the data collection and pretreatment in the coarse-grained positioning stage specifically comprise the following steps:

step 1: uniformly dividing cells into N₁A block having its geometric center as a classification position label;

step 2: in the coarse-grained location stage, the selected coarse-grained location labels are sparsely distributed and are few in number, so that the N divided is₁Multipath CSI information respectively collected in the blocks is estimated through a classical MUSIC estimation algorithm to obtain multipath AOA information of all coarse-grained position label points, and the multipath AOA information is used as fingerprint characteristics;

and step 3: adding the sample characteristics into a label, wherein the sample in the coarse-grained positioning stage is represented as follows:

wherein phi is₀AOA, phi of direct path₁Scattering for the first stripeThe AOA of the path is determined,

is the Nth_rayThe AOA and T of the scattering paths are matrix transposes;

the task of the coarse-grained positioning stage is to train parameters of a deep convolutional neural network and a regression classifier network according to the acquired labeled training data, and the training aim is to minimize the mean square error output by a training label and the network;

the off-line stage training process of the coarse-grained location link is as follows:

for a deep learning network, a deep convolutional neural network is adopted, a ReLU function is adopted as an excitation function of each layer of nodes, after training data are input into the network, the output of each layer is obtained according to the excitation function and is used as the input of the next layer, and finally, the network output is obtained through layer-by-layer forward propagation; constructing a penalty function according to a minimum mean square error principle, updating and iterating by using a random gradient descent algorithm to obtain a final training parameter, and storing a trained weight W, b as a part of a fingerprint library;

then, taking neural network training output data as input of a softmax classifier, then dividing the neural network training output data into C classes, taking the probability that the input data belongs to each class as output of the classifier, constructing a penalty function according to a minimum mean square error principle, updating and iterating by using a random gradient descent algorithm to obtain final training parameters, and forming W, b and theta together into a fingerprint library, wherein the theta is a classifier parameter;

the output of the classifier is as follows:

wherein,

is a C1 matrix, each term is represented in

In the given case of the situation where,

the probability of belonging to each of the classes,

the method comprises the following steps of (1) inputting training output of a neural network, namely input of a regression classifier, and theta is a parameter of the classifier;

the on-line stage positioning process of the coarse grain positioning link is as follows:

step 1: after receiving CSI information from users at unknown positions, the MUSIC algorithm estimates the received AOA information, namely

The probability that the unknown data belong to each to-be-determined position is obtained through forward propagation of a machine learning network and classification of a regression classifier;

step 2: using a probability method to take the position label with the maximum output probability as the final position estimation of coarse-grained positioning;

then, in a fine-grained positioning stage, on the basis of the coarse-grained positioning labels obtained through estimation, fine-grained labels are generated by taking the coarse-grained positioning labels as a core, and considering that the more the number of the fine-grained labels is, the finer the fine-grained positioning result is, the averaged CSI amplitude information is selected as the fingerprint characteristic of the fine-grained positioning stage, and a dynamic weighted K nearest neighbor algorithm based on space mapping is provided for fine-grained positioning;

finally, in order to improve the positioning accuracy, at the fine-grained stage, the following facts are considered: the same CSI difference may correspond to different geometric distances, that is, for CSI differences of different magnitudes, the signal distances represented by the CSI difference may correspond to different geometric distances; therefore, the relationship between the signal distance space and the geometric distance space is established by training the ELM to realize the spatial mapping and prevent the influence of the distance mismatch from causing the reduction of the positioning precision.

2. The indoor fingerprint positioning method for the millimeter wave massive MIMO system as claimed in claim 1, wherein the data collection and preprocessing at the fine-grained positioning stage comprises the following steps:

step 1: based on the coarse-grained positioning result, the coarse-grained positioning result is used as a core, the coarse-grained positioning result is expanded outwards at equal intervals to generate a fine-grained label called an expanded sliding window, and N is generated in the sliding window according to the method₂Fine-grained location tags;

step 2: for a fine-grained positioning stage, in order to further reduce data dimension and operation complexity, sampling CSI data of each path, then averaging amplitude channel matrixes on different antennas, and grouping and numbering the amplitude data according to the corresponding relation between known CSI information and user positions;

and step 3: adding the sample characteristics into a label, wherein the sample in the fine-grained positioning stage is represented as follows after the label is added:

wherein,

N₂the position label number of the fine-grained positioning stage.

3. The indoor fingerprint positioning method for the millimeter wave massive MIMO system as claimed in claim 2, wherein the task of the fine-grained positioning stage is to extend outward to form a sliding window full of the refined virtual tags with the estimated position as the center on the basis of the coarse-grained positioning; considering that phase shift exists in CSI, and in order to overcome high complexity brought by multiple antennas, averaged multipath CSI amplitude information is used as a position label, and a space mapping-based adaptive dynamic weighting K nearest neighbor algorithm comprises the following algorithm steps:

step 1: calculating the signal distance between all sample points of the fingerprint library:

step 2: deleting singular distance, and setting threshold T ═ alpha × D₁Keeping the signal distance which is not more than the threshold value T in the formula, and sequentially marking as D from small to large₁,...,D_SS-1, 2, …, S represents the number of retained neighboring reference points;

and step 3: calculate the mean distance difference of the retention points from each other:

wherein, Δ d_j,sRepresents D_jAnd D_sThe distance difference of (a);

and 4, step 4: estimating the final position, the dynamic weight value is expressed as:

wherein Δ D ═ D_K-D₁(ii) a In particular, when D_K＝D₁Time, omega_j1 is ═ 1; defining:

the specific form of the final position estimate is then as follows:

wherein (x)_j,y_j) The coordinates of the jth remaining location tag are represented.

4. The indoor fingerprint positioning method for mmwave massive MIMO system as claimed in claim 3, wherein in the proposed dynamic weighted K-nearest neighbor algorithm step (1) based on spatial mapping, when calculating the signal distance, considering the fact that the signal distance and the physical distance do not match at different CSI amplitude levels, the spatial mapping method based on extreme learning machine is adopted, and this problem is described as follows:

wherein S is_DAnd S_CRespectively signal space distance and geometric space distance.

Images