A two-stage high-precision positioning method in NLOS environment

Abstract

The present invention discloses a two-stage high-precision positioning method in an NLOS environment. The method comprises the following steps: step 1, in an NLOS scenario with a length of L and a width of W, base stations are evenly distributed on a square array with a distance of D; step 2, CIR data acquisition; step 3, TOA prediction; step 4, TOA positioning; step 5, setting the sample size to be represented by N sp ,(x i ,y i ) and #imgabs0# distribution represent the actual position and predicted position coordinates. The method of the present invention obtains CIR data through the channel information between the base station and the user and establishes a mapping relationship between CIR information and TOA through a TOA prediction network, thereby accurately predicting the time TOA when the base station reaches the user. A nonlinear mapping relationship is established between TOA and the user position through a deep learning method, thereby ensuring that the user position can be accurately predicted. The proposed deep learning method establishes a mapping function between CIR, TOA and user position by learning a large amount of channel information, which greatly improves the positioning accuracy in the NLOS environment and saves equipment costs.

Description

Technical Field

[0001] The present invention belongs to the fields of wireless positioning, 5G, and deep learning technology, and in particular relates to a two-stage high-precision positioning method in an NLOS environment. Background Art

[0002] First, the application scenarios, traditional positioning solutions, and deep learning technology (DL) related background knowledge of the present invention are introduced. According to 3GPP TS 22.261 and TR 22.804, in 5G networks, both commercial positioning services and industrial Internet positioning services have new positioning requirements. In the 5G positioning service area, the horizontal accuracy required for outdoor scenes (villages and towns) at a maximum speed of 250km / h is 1m; the horizontal positioning accuracy of flexible and modular assembly areas in smart factories is required to be below 0.5m. With the increase in positioning accuracy requirements, traditional positioning solutions such as TOA, TDOA, and AOA are mostly designed for direct (Line of Sight, LOS) channel environments, and have poor performance in non-direct environments (Non Line of Sight, NLOS), and the accuracy that can be achieved by existing positioning solutions is limited. Therefore, it is necessary to study sub-meter high-precision positioning solutions in NLOS scenarios.

[0003] Traditional positioning solutions measure the physical quantities of time of arrival (TOA), time difference of arrival (TDOA) and angle of arrival (AOA) for positioning. TOA positioning also has the following disadvantages: 1. TOA positioning requires strict time synchronization between the base station and the user, which has high requirements for equipment, so the measurement error is large and the accuracy is not high; 2. In some application scenarios, the NLOS situation is very severe, and it is difficult to obtain accurate TOA data. In traditional positioning algorithms, TOA calculations are completely dependent on time, so the time synchronization requirements of the system are very high. Therefore, TOA positioning is rarely used in actual scenarios.

[0004] At present, with the improvement of machine computing power and the expansion of data scale, deep learning technology has developed rapidly. The nonlinear processing ability of deep learning has successfully solved a series of problems that were difficult to handle before, and has shown strong performance in the fields of speech recognition and computer vision. In the field of wireless communications, the combination of deep learning technology and communication networks can bring new solutions to traditional communications and achieve good results in channel prediction, signal processing, and network optimization. Therefore, in wireless positioning, by introducing deep learning algorithms, complex nonlinear features in multiple dimensions of time domain, frequency domain, and spatial domain can be effectively extracted, thereby solving the time synchronization error and NLOS problems faced by traditional positioning methods and improving positioning accuracy. Summary of the invention

[0005] The object of the present invention is to provide a two-stage high-precision positioning method in a NLOS environment that can overcome the above-mentioned technical problems. The present invention comprises the following steps:

[0006] Step 1: In a NLOS scenario with a length of L and a width of W, the base stations are evenly distributed on a square matrix with a distance of D. The number of base stations is Ng, and the received signal from each base station to the user is shown in the following formula:

[0007]

[0008] Among them, x n (t) is the base station transmission signal, a n is the signal gain, τ n is the path delay, b(t) is Gaussian white noise, and N is the number of multipaths:

[0009] Step 1.1, N g The CIR information from base stations to users is predicted by the designed deep learning network. g TOA from base station to user,

[0010] Step 1.2, map TOA and user location through a deep learning network, and predict the user location based on the mapping relationship learned by the network;

[0011] Step 2, CIR data acquisition;

[0012] In a NLOS environment with a length of L and a width of W, 18 base stations are evenly placed, and the CIR information is shown in the following formula:

[0013]

[0014] For the CIR information between each base station and the user, 256 sampling points are selected as the input information of the TOA prediction network;

[0015] Step 3, TOA prediction: The TOA information is predicted by extracting the CIR data information to model the channel characteristics. The number of base stations is set to 18, the CIR length is 256, and the imaginary part information is included. The data dimension input to the TOA prediction network is (18, 256, 2). The TOA prediction network consists of four blocks, which specifically include the following steps:

[0016] Step 3.1, the input data first passes through Block 1, which consists of a layer of convolutional network. After the input data passes through a layer of convolutional network with a convolution kernel size of 3, it is normalized and the activation function channel dimension is increased to 64;

[0017] In the designed TOA prediction network, batch normalization (BN) is used to process the data to stabilize the model training process and avoid gradient disappearance or gradient explosion; the specific process is as follows:

[0018] The mean is calculated as shown in the following formula:

[0019]

[0020] The variance is calculated as shown in the following formula:

[0021]

[0022] Normalization is shown in the following formula:

[0023]

[0024] The activation function is shown in the following formula:

[0025]

[0026] Among them, α is 0.01; the advantage of using LeakyReLU is that the gradient can be calculated for the part of the LeakyReLU activation function that is less than 0 during the back propagation process;

[0027] Step 3.2, after two Block2s, Block2 consists of a two-layer convolutional network, adding the original data to the data after passing through the two-layer convolutional network to retain the previous features and prevent the gradient from disappearing or exploding due to the deepening of the network;

[0028] The shortcut connection directly skips the middle nonlinear layer. The connection combines the shortcut connection output and the direct output by performing an identity mapping. The residual block consists of multiple cascaded convolutional layers and a cross-layer direct connection edge. After the ReLU activation function, the output is passed, and a certain layer of data is directly skipped over multiple layers and introduced into the input part of the subsequent data layer. Dropout is used before the output of Block2 so that a part of the neurons are temporarily discarded from the network according to the set probability during the training process to reduce the scale of the neural network and prevent overfitting.

[0029] Step 3.3, after 2 layers of Block3, Block3 is also composed of 2 layers of convolutional networks. The step size s of the first layer of convolution is set to 2, so that the size of the feature map is halved. The original data passes through a convolutional network with a convolution kernel size of 1×1 and a step size s of 2 to keep consistent with the feature map after two layers of convolution;

[0030] Step 3.4, after Block 4, Block 4 further reduces the size of the feature map through convolution and passes the flattened feature with a dimension of 2048 through a fully connected network to obtain the predicted TOA information;

[0031] The Flatten layer is used to expand the convolutional feature map into a one-dimensional feature vector;

[0032] The Dense layer is used to compress the features of the input data after convolution and pooling operations and plays a role in mapping the learned features to the label space of the sample in the AI ​​network;

[0033] During the model training process, the mean square error (MSE) is selected as the loss function, as shown in the following formula:

[0034]

[0035] Among them, y n and Represent the real TOA and predicted TOA information respectively;

[0036] Step 4: TOA positioning. After obtaining TOA, the user location is predicted using the positioning network. After obtaining the predicted TOA data, it is directly input into the positioning network to learn the relationship between the TOA information and the location coordinates between the base station and the user to reduce the interference caused by the NLOS environment.

[0037] Step 5: Set the sample size to be denoted as N sp ,(x i ,y i )and The distribution represents the actual position and the predicted position coordinates, then the average error can be expressed as:

[0038]

[0039] In 3GPP Rel-18, the horizontal accuracy achieved by 90% of users is selected as the evaluation index, and the error cumulative distribution function (CDF) is drawn. The error when the cumulative probability reaches 90% is selected as the performance evaluation index of the present invention.

[0040] The method of the present invention has the following beneficial effects:

[0041] 1. The method of the present invention obtains CIR data through the channel information between the base station and the user and establishes a mapping relationship between CIR information and TOA through a TOA prediction network, thereby accurately predicting the time TOA when the base station reaches the user. Considering the TOA positioning application and the LOS scenario, a TOA positioning network in the NLOS scenario is designed. A nonlinear mapping relationship is established between TOA and the user position through a deep learning method, thereby ensuring that the user position can be accurately predicted;

[0042] 2. The deep learning method proposed by the method of the present invention establishes a mapping function between CIR, TOA and user position by learning a large amount of channel information, which greatly improves the positioning accuracy in the NLOS environment and saves equipment costs;

[0043] 3. The method of the present invention considers the positioning problem in the NLOS scenario and obtains CIR data through the channel information between the base station and the user, and uses the deep learning network to extract CIR data information to accurately model the mapping relationship between CIR and TOA to obtain TOA information between the base station and the user; then positioning is performed through the predicted TOA information, which can significantly improve the positioning accuracy compared with the TOA data positioning obtained by direct measurement;

[0044] 4. After the method of the present invention predicts the TOA data, a TOA positioning network is designed to establish a mapping relationship between the arrival time from the base station to the user and the user position for positioning. The TOA information is accurately predicted in the existing equipment and NLOS environment through a deep learning network, and the TOA data with small errors is used for positioning, thereby reducing the equipment cost and solving the problem of insufficient accuracy of traditional positioning solutions. The CIR data features from the base station to the user are accurately extracted through a deep learning network, thereby solving the problems of difficult time synchronization between devices and large TOA measurement errors caused by NLOS scenarios. The deep learning network is directly used to predict the user position, avoiding the positioning error problem caused by measurement errors in traditional TOA positioning.

[0045] 5. Compared with the traditional positioning algorithms of TOA, TDOA, and AOA, the high-precision positioning solution proposed by the method of the present invention has low requirements on equipment, which can not only effectively reduce the measurement cost, but also improve the positioning accuracy. In terms of application scenarios, the method of the present invention can be applied to a variety of actual scenarios and has strong scalability. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is a schematic diagram of the overall scheme design process of the method of the present invention;

[0047] Figure 2 is a schematic diagram of the spatial distribution of base stations of the method of the present invention;

[0048] Figure 3 It is a TOA prediction network schematic diagram of the method of the present invention;

[0049] Figure 4 is a schematic diagram of the residual structure of the method of the present invention;

[0050] Figure 5 It is a schematic diagram of the positioning network of the method described in the present invention. DETAILED DESCRIPTION

[0051] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0052] The present invention comprises the following steps:

[0053] like Figure 1 As shown, step 1, in the NLOS scenario with a length of L and a width of W, the base stations are evenly distributed on a square matrix with a distance of D, the number of base stations is Ng, and the received signal from each base station to the user is shown in the following formula:

[0054]

[0055] Among them, x n (t) is the base station transmission signal, a n is the signal gain, τ n is the path delay, b(t) is Gaussian white noise, and N is the number of multipaths:

[0056] Step 1.1, N g The CIR information from base stations to users is predicted by the designed deep learning network. g TOA from base station to user,

[0057] Step 1.2, map TOA and user location through a deep learning network, and predict the user location based on the mapping relationship learned by the network;

[0058] Step 2, CIR data acquisition;

[0059] like Figure 2 As shown in the figure, in a NLOS environment with a length of L and a width of W, 18 base stations are evenly placed, and the CIR information is shown in the following formula:

[0060]

[0061] For the CIR information between each base station and the user, 256 sampling points are selected as the input information of the TOA prediction network;

[0062] Step 3, TOA prediction; Figure 3 As shown, by extracting CIR data information to model channel characteristics to predict TOA information, the number of base stations is set to 18, the CIR length is 256, including the time imaginary part information, and the data dimension input to the TOA prediction network is (18, 256, 2). The TOA prediction network consists of four blocks, which specifically include the following steps:

[0063] Step 3.1, the input data first passes through Block 1, which consists of a layer of convolutional network. After the input data passes through a layer of convolutional network with a convolution kernel size of 3, it is normalized and the activation function channel dimension is increased to 64;

[0064] In the designed TOA prediction network, batch normalization (BN) is used to process the data to stabilize the model training process and avoid gradient disappearance or gradient explosion; the specific process is as follows:

[0065] The mean is calculated as shown in the following formula:

[0066]

[0067] The variance is calculated as shown in the following formula:

[0068]

[0069] Normalization is shown in the following formula:

[0070]

[0071] The activation function is shown in the following formula:

[0072]

[0073] Among them, α is 0.01; the advantage of using LeakyReLU is that the gradient can be calculated for the part of the LeakyReLU activation function that is less than 0 during the back propagation process;

[0074] Step 3.2, after two Block2s, Block2 consists of a two-layer convolutional network, adding the original data to the data after passing through the two-layer convolutional network to retain the previous features and prevent the gradient from disappearing or exploding due to the deepening of the network;

[0075] The design of Block2 is as follows Figure 4 As shown in the figure, the shortcut connection directly skips the middle nonlinear layer. The connection combines the shortcut connection output and the direct output by performing an identity mapping. The residual block consists of multiple cascaded convolutional layers and a cross-layer direct connection edge. After the ReLU activation function, the output is output, and a certain layer of data is directly skipped over multiple layers and introduced into the input part of the subsequent data layer. Dropout is used before the output of Block2 so that a part of the neurons are temporarily discarded from the network according to the set probability during the training process to reduce the scale of the neural network and prevent overfitting.

[0076] Step 3.3, after 2 layers of Block3, Block3 is also composed of 2 layers of convolutional networks. The step size s of the first layer of convolution is set to 2, so that the size of the feature map is halved. The original data passes through a convolutional network with a convolution kernel size of 1×1 and a step size s of 2 to keep consistent with the feature map after two layers of convolution;

[0077] Step 3.4, after Block 4, Block 4 further reduces the size of the feature map through convolution and passes the flattened feature with a dimension of 2048 through a fully connected network to obtain the predicted TOA information;

[0078] The Flatten layer is used to expand the convolutional feature map into a one-dimensional feature vector;

[0079] The Dense layer is used to compress the features of the input data after convolution and pooling operations and plays a role in mapping the learned features to the label space of the sample in the AI ​​network;

[0080] During the model training process, the mean square error (MSE) is selected as the loss function, as shown in the following formula:

[0081]

[0082] Among them, y n and Represent the real TOA and predicted TOA information respectively;

[0083] Step 4: TOA positioning. After obtaining TOA, use Figure 5The positioning network shown predicts the user location, and after obtaining the predicted TOA data, it is directly input into the positioning network to learn the relationship between the TOA information from the base station to the user and the location coordinates to reduce the interference caused by the NLOS environment;

[0084] Step 5: Set the sample size to be denoted as N sp ,(x i ,y i )and The distribution represents the actual position and the predicted position coordinates, then the average error can be expressed as:

[0085]

[0086] In 3GPP Rel-18, the horizontal accuracy achieved by 90% of users is selected as the evaluation index, and the error cumulative distribution function (CDF) is drawn. The error when the cumulative probability reaches 90% is selected as the performance evaluation index of the present invention.

[0087] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by any technician familiar with the technical field within the scope disclosed in the present invention should be covered by the protection scope of the claims of the present invention.

Claims

1. A two-stage high-precision positioning method in a NLOS environment, characterized in that: The following steps are involved: Step 1: In a NLOS scenario with a length of L and a width of W, the base stations are evenly distributed on a square matrix with a distance of D. The number of base stations is Ng, and the received signal from each base station to the user is shown in the following formula: Among them, x n (t) is the base station transmission signal, a n is the signal gain, τ n is the path delay, b(t) is Gaussian white noise, and N is the number of multipaths; Step 2, CIR data acquisition; In a NLOS environment with a length of L and a width of W, 18 base stations are evenly placed, and the CIR information is shown in the following formula: For the CIR information between each base station and the user, 256 sampling points are selected as the input information of the TOA prediction network; Step 3, TOA prediction: The TOA information is predicted by extracting the CIR data information to model the channel characteristics, setting the number of base stations to 18, the CIR length to 256, including the time imaginary part information, and the data dimension input to the TOA prediction network is (18, 256, 2). The TOA prediction network consists of four blocks; Step 4: TOA positioning. After obtaining TOA, the user location is predicted using the positioning network. After obtaining the predicted TOA data, it is directly input into the positioning network to learn the relationship between the TOA information and the location coordinates between the base station and the user to reduce the interference caused by the NLOS environment. Step 5: Set the sample size to be denoted as N sp ,(x i ,y i )and The distribution represents the actual position and the predicted position coordinates, then the average error can be expressed as: In 3GPP Rel-18, the horizontal accuracy achieved by 90% of users is selected as the evaluation indicator, the error cumulative distribution function is drawn, and the error when the cumulative probability reaches 90% is selected as the performance evaluation indicator.

2. The two-stage high-precision positioning method in a NLOS environment according to claim 1, characterized in that: The step 1 comprises the following steps: Step 1.1, N g The CIR information from base stations to users is predicted by the designed deep learning network. g TOA from base station to user, Step 1.2, map TOA and user location through a deep learning network, and predict the user location based on the mapping relationship learned by the network.

3. The two-stage high-precision positioning method in a NLOS environment according to claim 1, characterized in that: The step 3 comprises the following steps: Step 3.1, the input data first passes through Block 1, which consists of a layer of convolutional network. After the input data passes through a layer of convolutional network with a convolution kernel size of 3, it is normalized and the activation function channel dimension is increased to 64; In the designed TOA prediction network, batch normalization is used to process the data to stabilize the model training process and avoid gradient vanishing or gradient exploding; Step 3.2, after two Block2s, Block2 consists of a two-layer convolutional network, adding the original data to the data after passing through the two-layer convolutional network to retain the previous features and prevent the gradient from disappearing or exploding due to the deepening of the network; Step 3.3, after 2 layers of Block3, Block3 is also composed of 2 layers of convolutional networks. The step size s of the first layer of convolution is set to 2, so that the size of the feature map is halved. The original data passes through a convolutional network with a convolution kernel size of 1×1 and a step size s of 2 to keep consistent with the feature map after two layers of convolution; Step 3.4, after Block 4, Block 4 further reduces the size of the feature map through convolution and passes the flattened features with a dimension of 2048 through a fully connected network to obtain the predicted TOA information.

4. The two-stage high-precision positioning method in a NLOS environment according to claim 3, characterized in that: In step 3.1: The mean is calculated as shown in the following formula: The variance is calculated as shown in the following formula: Normalization is shown in the following formula: The activation function is shown in the following formula: Among them, α is 0.01; the advantage of using LeakyReLU is that the gradient can be calculated for the part of the LeakyReLU activation function that is less than 0 during the back propagation process.

5. The two-stage high-precision positioning method in a NLOS environment according to claim 3, characterized in that: In step 3.2: The shortcut connection directly skips the middle nonlinear layer. The connection combines the shortcut connection output and the direct output by performing an identity mapping. The residual block consists of multiple cascaded convolutional layers and a cross-layer direct connection edge. After passing through the ReLU activation function, the output directly skips multiple layers of data from a certain layer and is introduced into the input part of the subsequent data layer. Dropout is used before the output of Block2 so that a part of the neurons are temporarily discarded from the network according to the set probability during the training process to reduce the scale of the neural network and prevent overfitting.

6. The two-stage high-precision positioning method in a NLOS environment according to claim 3, characterized in that: In step 3.4: The Flatten layer is used to expand the convolutional feature map into a one-dimensional feature vector; The Dense layer is used to compress the features of the input data after convolution and pooling operations and plays a role in mapping the learned features to the label space of the sample in the AI ​​network; During the model training process, the mean square error (MSE) is selected as the loss function, as shown in the following formula: Among them, y n and Represent the true TOA and predicted TOA information respectively.

Images