A positioning method based on 5G base station

Abstract

The present invention discloses a positioning method based on 5G base stations, and belongs to the technical field of positioning of 5G base stations. The present invention constructs a signal feature matrix and a signal gap coefficient matrix according to the signal characteristics of each 5G base station. The signal feature matrix reflects the signal characteristics of each 5G base station, and the signal gap coefficient matrix reflects the gap between the signal feature sequences reached in sequence. The difference in signal characteristics at adjacent moments enhances the feature distinction between different 5G base station positions. A multi-fusion positioning network is used to process the signal feature matrix and the signal gap coefficient matrix to obtain an initial position, so that the multi-fusion positioning network combines static characteristics and dynamic change information to accurately estimate the terminal position, while reducing the influence of instantaneous noise and outliers, improving the stability of the positioning result, and then using the terminal positions at multiple historical moments to correct the initial position, further improving the positioning accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of 5G base station positioning, and in particular to a 5G base station positioning method. Background Art

[0002] With the rapid development and widespread deployment of 5G technology, high-precision positioning services have become a research field that has attracted much attention. Traditional positioning methods such as GPS often fail to provide accurate positioning that meets the needs of modern applications in indoor environments or densely populated urban areas. The emergence of 5G networks provides new possibilities for solving this problem, especially in scenarios that require high-precision, low-latency positioning, such as smart transportation, augmented reality, and industrial automation. In 5G networks, terminals can usually receive signals from multiple base stations. How to coordinate and fuse this information to improve positioning accuracy is an important issue. Existing positioning technologies based on 5G base stations mainly include: 1. Determine the position by measuring the angle at which the signal reaches the terminal, and 2. Estimate the distance based on the received signal strength to determine the position. However, in reality, there are many obstacles that cause signal reflection and scattering, resulting in multipath effects and non-line-of-sight propagation, which affect positioning accuracy. Summary of the invention

[0003] In view of the above-mentioned deficiencies in the prior art, the present invention provides a positioning method based on 5G base stations to solve the problem of low positioning accuracy in the prior art.

[0004] In order to achieve the above-mentioned purpose of the invention, the technical solution adopted by the present invention is: a positioning method based on a 5G base station, comprising the following steps:

[0005] S1. Construct a signal feature sequence based on the signal features of each 5G base station;

[0006] S2. Sort the signal feature sequences according to the signal arrival time of each 5G base station, and construct a signal gap coefficient matrix and a signal feature matrix;

[0007] S3, using a multi-fusion positioning network to process the signal feature matrix and the signal gap coefficient matrix to obtain an initial position;

[0008] S4. According to the terminal positions at multiple historical moments, the initial position is corrected to obtain the terminal position at the current moment.

[0009] The beneficial effects of the present invention are as follows: the present invention constructs a signal feature sequence according to the signal characteristics of each 5G base station, sorts each signal feature sequence based on the arrival time of the signal, constructs a signal feature matrix, and then constructs a signal gap coefficient matrix. The signal feature matrix reflects the signal characteristics of each 5G base station, and the signal gap coefficient matrix reflects the gap between each signal feature sequence reached in sequence. The difference in signal characteristics at adjacent moments enhances the feature distinction between different 5G base station locations. A multi-fusion positioning network is used to process the signal feature matrix and the signal gap coefficient matrix to obtain an initial position, so that the multi-fusion positioning network combines static characteristics and dynamic change information to accurately estimate the terminal position, while reducing the influence of instantaneous noise and outliers, improving the stability of the positioning results, and then using the terminal positions at multiple historical moments to correct the initial position to further improve the positioning accuracy.

[0010] Furthermore, the elements in the signal characteristic sequence in S1 include: received signal power, signal arrival time, arrival angle and channel state information.

[0011] Furthermore, the S2 comprises the following sub-steps:

[0012] S21. Sort each signal feature sequence according to the signal arrival time of each 5G base station;

[0013] S22, using the sorted signal feature sequence as a row vector to form a signal feature matrix;

[0014] S23, subtract the two signal feature sequences at adjacent moments to obtain the signal feature difference sequence, D=X t -X t-1 , where D is the signal feature gap sequence, X t is the signal characteristic sequence at the tth arrival time, X t-1 is the signal feature sequence of the t-1th arrival time, t is the number of the arrival time;

[0015] S24, normalizing the elements in each signal feature gap sequence to obtain a signal gap coefficient sequence;

[0016] S25. Take each signal gap coefficient sequence as a row vector to form a signal gap coefficient matrix.

[0017] Furthermore, the multi-fusion positioning network in S3 includes: a first scale feature extraction unit, a second scale feature extraction unit, a third scale feature extraction unit, a fourth scale feature extraction unit, a first feature fusion unit, a second feature fusion unit, a first fully connected layer, a second fully connected layer, an adder A1 and an output layer;

[0018] The input end of the first scale feature extraction unit is used to input a signal feature matrix;

[0019] The input end of the second scale feature extraction unit is used for inputting a signal feature matrix;

[0020] The input end of the third scale feature extraction unit is used for inputting a signal gap coefficient matrix;

[0021] The input end of the fourth scale feature extraction unit is used for inputting a signal gap coefficient matrix;

[0022] The input end of the first feature fusion unit is connected to the output end of the first scale feature extraction unit and the output end of the third scale feature extraction unit respectively, and the output end thereof is connected to the input end of the first fully connected layer;

[0023] The input end of the second feature fusion unit is connected to the output end of the second scale feature extraction unit and the output end of the fourth scale feature extraction unit respectively, and the output end thereof is connected to the input end of the second fully connected layer;

[0024] The first input terminal of the adder A1 is connected to the output terminal of the first fully connected layer, the second input terminal thereof is connected to the output terminal of the second fully connected layer, and the output terminal thereof is connected to the input terminal of the output layer;

[0025] The output end of the output layer serves as the output end of the multi-fusion positioning network.

[0026] The beneficial effect of the above further scheme is: the present invention processes the same signal feature matrix through the first scale feature extraction unit and the second scale feature extraction unit respectively to extract features of different scales, processes the same signal gap coefficient matrix through the third scale feature extraction unit and the fourth scale feature extraction unit to extract features of different scales, fuses the output features of the first scale feature extraction unit and the third scale feature extraction unit at the first feature fusion unit, and fuses the output features of the second scale feature extraction unit and the fourth scale feature extraction unit at the second feature fusion unit, so that in one feature fusion unit, features of the same scale are fused, and then processed through a fully connected layer, each neuron in the fully connected layer is connected to all neurons in the previous layer, and by learning the weights of these connections, the network can automatically determine the importance of different features and optimize the combination of features. The output layer integrates the output of adder A1 to obtain the initial position.

[0027] Further, the first scale feature extraction unit, the second scale feature extraction unit, the third scale feature extraction unit and the fourth scale feature extraction unit each include: a first convolution block, an SPP layer, a second convolution block, a Sigmoid layer, a multiplier M1 and an adder A2;

[0028] The input end of the first convolution block is used as the input end of the first scale feature extraction unit, the second scale feature extraction unit, the third scale feature extraction unit or the fourth scale feature extraction unit; the input end of the SPP layer is connected to the output end of the first convolution block, and its output end is respectively connected to the input end of the second convolution block and the first input end of the adder A2; the output end of the second convolution block is respectively connected to the input end of the Sigmoid layer and the first input end of the multiplier M1; the second input end of the multiplier M1 is connected to the output end of the Sigmoid layer, and its output end is connected to the second input end of the adder A2; the output end of the adder A2 is used as the output end of the first scale feature extraction unit, the second scale feature extraction unit, the third scale feature extraction unit or the fourth scale feature extraction unit;

[0029] The size of the convolution kernel in the first convolution block in the first scale feature extraction unit and the third scale feature extraction unit is ;

[0030] The size of the convolution kernel in the first convolution block in the second scale feature extraction unit and the fourth scale feature extraction unit is .

[0031] The beneficial effects of the above further scheme are as follows: the present invention adopts the first scale feature extraction unit and the third scale feature extraction unit to extract features of the same scale, and the second scale feature extraction unit and the fourth scale feature extraction unit to extract features of the same scale. The SPP layer is adopted to enhance the adaptability of the network to inputs of different scales, and the robustness of feature extraction is improved. The Sigmoid layer and the multiplier M1 enable the network to adaptively focus on important features, ignore irrelevant information, and improve the efficiency and accuracy of feature extraction. The adder A2 realizes the residual connection, adds the output of the SPP layer to the features after attention processing, alleviates the gradient vanishing problem of the deep network, and promotes the training of deeper networks.

[0032] Furthermore, the expression of the first feature fusion unit is: , where H 1 is the output of the first feature fusion unit, Y 1 is the output of the first scale feature extraction unit, Y 3 is the output of the third scale feature extraction unit, For Hadamard;

[0033] The expression of the second feature fusion unit is: , where H 2 is the output of the second feature fusion unit, Y 2 is the output of the second scale feature extraction unit, Y 4 is the output of the fourth scale feature extraction unit.

[0034] The beneficial effect of the above further scheme is: the present invention multiplies the features extracted from the signal feature matrix by elements with the features extracted from the signal gap coefficient matrix, and combines the static and dynamic features to more accurately describe the changing characteristics of the signal in space and time.

[0035] Furthermore, the expression of the output layer is: , , among which, long 0 is the longitude of the initial position, lat 0 is the latitude of the initial position, x i is the i-th eigenvalue output by adder A1, ω 1,i is the longitude weight of the i-th eigenvalue, b 1,i is the longitude bias of the i-th eigenvalue, ω 2,i is the latitude weight of the i-th eigenvalue, b 2,i is the latitude bias of the i-th eigenvalue, N is the number of eigenvalues ​​output by adder A1, and i is a positive integer.

[0036] Further, the S4 comprises the following steps:

[0037] S41. According to the terminal positions at multiple historical moments, a quadratic polynomial is used to fit multiple longitudes to obtain a terminal movement longitude trajectory, and a quadratic polynomial is used to fit multiple dimensions to obtain a terminal movement dimension trajectory;

[0038] S42, using the terminal's moving longitude trajectory to predict the longitude at the next moment, and using the terminal's moving latitude trajectory to predict the latitude at the next moment, to obtain a predicted position;

[0039] S43, calculating the deviation coefficient between the initial position and the predicted position;

[0040] S44. Perform weighted processing on the initial position and the predicted position according to the deviation coefficient between the initial position and the predicted position to obtain the terminal position at the current moment.

[0041] Furthermore, the formula for calculating the bias coefficient between the initial position and the predicted position in S43 is: , where γ is the deviation coefficient, arctan is the inverse tangent function, and long 0 is the longitude of the initial position, lat 0 is the latitude of the initial position, is the longitude of the predicted location, is the latitude of the predicted location.

[0042] Furthermore, the weighted processing formula in S44 is: , , where x is the longitude of the terminal's location at the current moment, and y is the latitude of the terminal's location at the current moment.

[0043] The beneficial effects of the above further scheme are as follows: the present invention uses quadratic polynomials to fit historical location data and establish mobile trajectory models of longitude and latitude respectively. This method can capture the nonlinear characteristics of terminal motion. Then, the position at the next moment is predicted based on these trajectory models, which provides a powerful priori information for position estimation. By calculating the deviation coefficient between the initial position and the predicted position, the influence of historical trajectory information and current measurement data can be dynamically balanced. Finally, a weighted processing method is used to fuse the initial position and the predicted position, which takes into account the historical movement trend and retains the accuracy of the current measurement. This method can not only smooth the positioning results and reduce the impact of instantaneous errors, but also improve the continuity and stability of positioning. Especially in the case of unstable signals or noisy measurements, this method can significantly improve positioning accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a flow chart of a positioning method based on 5G base station;

[0045] Figure 2 It is a structural diagram of a multi-fusion positioning network;

[0046] Figure 3 It is a schematic diagram of the structures of a first scale feature extraction unit, a second scale feature extraction unit, a third scale feature extraction unit and a fourth scale feature extraction unit. DETAILED DESCRIPTION

[0047] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0048] like Figure 1 As shown, a positioning method based on a 5G base station includes the following steps:

[0049] S1. Construct a signal feature sequence based on the signal features of each 5G base station;

[0050] S2. Sort the signal feature sequences according to the signal arrival time of each 5G base station, and construct a signal gap coefficient matrix and a signal feature matrix;

[0051] S3, using a multi-fusion positioning network to process the signal feature matrix and the signal gap coefficient matrix to obtain an initial position;

[0052] S4. According to the terminal positions at multiple historical moments, the initial position is corrected to obtain the terminal position at the current moment.

[0053] The elements in the signal characteristic sequence in S1 include: received signal power, signal arrival time, arrival angle and channel state information.

[0054] The present invention comprehensively utilizes multi-dimensional features such as received signal power, signal arrival time, arrival angle and channel state information to construct a comprehensive signal feature sequence to fully capture the rich characteristics of 5G signals.

[0055] The S2 comprises the following sub-steps:

[0056] S21. Sort each signal feature sequence according to the signal arrival time of each 5G base station;

[0057] S22, using the sorted signal feature sequence as a row vector to form a signal feature matrix;

[0058] In this embodiment, the order is sorted in ascending order according to the arrival time. Therefore, the first row in the signal feature matrix has the earliest arrival time, and the last row has the latest arrival time.

[0059] S23, subtract the two signal feature sequences at adjacent moments to obtain the signal feature difference sequence, D=X t -X t-1 , where D is the signal feature gap sequence, X t is the signal characteristic sequence at the tth arrival time, X t-1 is the signal feature sequence of the t-1th arrival time, t is the number of the arrival time;

[0060] S24, normalizing the elements in each signal feature gap sequence to obtain a signal gap coefficient sequence;

[0061] S25. Take each signal gap coefficient sequence as a row vector to form a signal gap coefficient matrix.

[0062] like Figure 2 As shown, the multi-fusion positioning network in S3 includes: a first scale feature extraction unit, a second scale feature extraction unit, a third scale feature extraction unit, a fourth scale feature extraction unit, a first feature fusion unit, a second feature fusion unit, a first fully connected layer, a second fully connected layer, an adder A1 and an output layer;

[0063] The input end of the first scale feature extraction unit is used to input a signal feature matrix;

[0064] The input end of the second scale feature extraction unit is used for inputting a signal feature matrix;

[0065] The input end of the third scale feature extraction unit is used for inputting a signal gap coefficient matrix;

[0066] The input end of the fourth scale feature extraction unit is used for inputting a signal gap coefficient matrix;

[0067] The input end of the first feature fusion unit is connected to the output end of the first scale feature extraction unit and the output end of the third scale feature extraction unit respectively, and the output end thereof is connected to the input end of the first fully connected layer;

[0068] The input end of the second feature fusion unit is connected to the output end of the second scale feature extraction unit and the output end of the fourth scale feature extraction unit respectively, and the output end thereof is connected to the input end of the second fully connected layer;

[0069] The first input terminal of the adder A1 is connected to the output terminal of the first fully connected layer, the second input terminal thereof is connected to the output terminal of the second fully connected layer, and the output terminal thereof is connected to the input terminal of the output layer;

[0070] The output end of the output layer serves as the output end of the multi-fusion positioning network.

[0071] The present invention processes the same signal feature matrix through a first scale feature extraction unit and a second scale feature extraction unit to extract features of different scales, processes the same signal gap coefficient matrix through a third scale feature extraction unit and a fourth scale feature extraction unit to extract features of different scales, fuses the output features of the first scale feature extraction unit and the third scale feature extraction unit at a first feature fusion unit, and fuses the output features of the second scale feature extraction unit and the fourth scale feature extraction unit at a second feature fusion unit, so that features of the same scale are fused in one feature fusion unit, and then processed through a fully connected layer, each neuron in the fully connected layer is connected to all neurons in an upper layer, and by learning the weights of these connections, the network can automatically determine the importance of different features, optimize the combination mode of features, and the output of the output layer synthesizes the output of the adder A1 to obtain the initial position.

[0072] like Figure 3 As shown, the first scale feature extraction unit, the second scale feature extraction unit, the third scale feature extraction unit and the fourth scale feature extraction unit all include: a first convolution block, an SPP layer, a second convolution block, a Sigmoid layer, a multiplier M1 and an adder A2;

[0073] The input end of the first convolution block is used as the input end of the first scale feature extraction unit, the second scale feature extraction unit, the third scale feature extraction unit or the fourth scale feature extraction unit; the input end of the SPP layer is connected to the output end of the first convolution block, and its output end is respectively connected to the input end of the second convolution block and the first input end of the adder A2; the output end of the second convolution block is respectively connected to the input end of the Sigmoid layer and the first input end of the multiplier M1; the second input end of the multiplier M1 is connected to the output end of the Sigmoid layer, and its output end is connected to the second input end of the adder A2; the output end of the adder A2 is used as the output end of the first scale feature extraction unit, the second scale feature extraction unit, the third scale feature extraction unit or the fourth scale feature extraction unit;

[0074] The size of the convolution kernel in the first convolution block in the first scale feature extraction unit and the third scale feature extraction unit is ;

[0075] The size of the convolution kernel in the first convolution block in the second scale feature extraction unit and the fourth scale feature extraction unit is .

[0076] The SPP layer is the Spatial Pyramid Pooling layer, and the convolution block includes: convolution layer, ReLU layer and BN layer.

[0077] In this embodiment, the multi-fusion positioning network is trained by the gradient descent method.

[0078] The present invention adopts a first scale feature extraction unit and a third scale feature extraction unit to extract features of the same scale, and a second scale feature extraction unit and a fourth scale feature extraction unit to extract features of the same scale. The SPP layer is adopted to enhance the adaptability of the network to inputs of different scales, and the robustness of feature extraction is improved. The Sigmoid layer and the multiplier M1 enable the network to adaptively focus on important features, ignore irrelevant information, and improve the efficiency and accuracy of feature extraction. The adder A2 realizes the residual connection, adds the output of the SPP layer to the features after attention processing, alleviates the gradient vanishing problem of the deep network, and promotes the training of deeper networks.

[0079] The expression of the first feature fusion unit is: , where H 1 is the output of the first feature fusion unit, Y 1 is the output of the first scale feature extraction unit, Y 3 is the output of the third scale feature extraction unit, For Hadamard;

[0080] The expression of the second feature fusion unit is: , where H 2 is the output of the second feature fusion unit, Y 2 is the output of the second scale feature extraction unit, Y 4 is the output of the fourth scale feature extraction unit.

[0081] The present invention multiplies the features extracted from the signal feature matrix by elements and the features extracted from the signal gap coefficient matrix by elements, and combines the static and dynamic features to more accurately describe the changing characteristics of the signal in space and time.

[0082] The expression of the output layer is: , , among which, long 0 is the longitude of the initial position, lat 0 is the latitude of the initial position, x i is the i-th eigenvalue output by adder A1, ω 1,i is the longitude weight of the i-th eigenvalue, b 1,i is the longitude bias of the i-th eigenvalue, ω 2,i is the latitude weight of the i-th eigenvalue, b 2,i is the latitude bias of the i-th eigenvalue, N is the number of eigenvalues ​​output by adder A1, and i is a positive integer.

[0083] In the present invention, the output layer includes two types of weights and biases to predict longitude and latitude respectively.

[0084] The S4 comprises the following steps:

[0085] S41. According to the terminal positions at multiple historical moments, a quadratic polynomial is used to fit multiple longitudes to obtain a terminal movement longitude trajectory, and a quadratic polynomial is used to fit multiple dimensions to obtain a terminal movement dimension trajectory;

[0086] In this embodiment, the quadratic polynomial is: 1 = a 1 t 2 +b 1 t+c 1 , f 2 = a 2 t 2 +b 2 t+c 2 , f 1 is the terminal moving longitude trajectory curve, f 2 is the terminal moving dimension trajectory, t is time, a 1 、b 1 、a 2 、b 2 For four multipliers, c 1 、c2 for two increments;

[0087] S42, using the terminal's moving longitude trajectory to predict the longitude at the next moment, and using the terminal's moving latitude trajectory to predict the latitude at the next moment, to obtain a predicted position;

[0088] S43, calculating the deviation coefficient between the initial position and the predicted position;

[0089] S44. Perform weighted processing on the initial position and the predicted position according to the deviation coefficient between the initial position and the predicted position to obtain the terminal position at the current moment.

[0090] The formula for calculating the bias coefficient between the initial position and the predicted position in S43 is: , where γ is the deviation coefficient, arctan is the inverse tangent function, and long 0 is the longitude of the initial position, lat 0 is the latitude of the initial position, is the longitude of the predicted location, is the latitude of the predicted location.

[0091] The weighted processing formula in S44 is: , , where x is the longitude of the terminal's location at the current moment, and y is the latitude of the terminal's location at the current moment.

[0092] The present invention uses quadratic polynomials to fit historical location data and establish mobile trajectory models for longitude and latitude respectively. This method can capture the nonlinear characteristics of terminal movement. Then, based on these trajectory models, the position at the next moment is predicted, which provides a powerful priori information for position estimation. By calculating the deviation coefficient between the initial position and the predicted position, the influence of historical trajectory information and current measurement data can be dynamically balanced. Finally, a weighted processing method is used to fuse the initial position and the predicted position, which not only takes into account the historical movement trend, but also retains the accuracy of the current measurement. This method can not only smooth the positioning results and reduce the impact of instantaneous errors, but also improve the continuity and stability of positioning. Especially when the signal is unstable or there is noise in the measurement, this method can significantly improve the positioning accuracy.

[0093] The present invention constructs a signal feature sequence according to the signal characteristics of each 5G base station, sorts each signal feature sequence based on the arrival time of the signal, constructs a signal feature matrix, and then constructs a signal gap coefficient matrix. The signal feature matrix reflects the signal characteristics of each 5G base station, and the signal gap coefficient matrix reflects the gap between each signal feature sequence reached in sequence. The difference in signal characteristics at adjacent moments enhances the feature distinction between different 5G base station locations. A multi-fusion positioning network is used to process the signal feature matrix and the signal gap coefficient matrix to obtain an initial position, so that the multi-fusion positioning network combines static characteristics and dynamic change information to accurately estimate the terminal position, while reducing the influence of instantaneous noise and outliers, improving the stability of the positioning result, and then using the terminal positions at multiple historical moments to correct the initial position to further improve the positioning accuracy.

[0094] In reality, due to the presence of many obstacles, signal reflection and scattering will occur, resulting in multipath effects and non-line-of-sight propagation, which will affect the positioning accuracy. The present invention simultaneously constructs a signal feature matrix based on the signal characteristics of multiple 5G base stations, and constructs a signal difference coefficient matrix based on the signal arrival time, which characterizes the signal differences of different 5G base stations, and uses a multi-fusion positioning network to process the signal differences and signal characteristics, predict the initial positioning, effectively reduce the accumulation of single-point measurement errors, and provide more stable and accurate positioning results.

[0095] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A positioning method based on a 5G base station, characterized in that: The following steps are involved: S1. Construct a signal feature sequence based on the signal features of each 5G base station; S2. Sort the signal feature sequences according to the signal arrival time of each 5G base station, and construct a signal gap coefficient matrix and a signal feature matrix; S3, using a multi-fusion positioning network to process the signal feature matrix and the signal gap coefficient matrix to obtain an initial position; S4. According to the terminal positions at multiple historical moments, the initial position is corrected to obtain the terminal position at the current moment; The S2 comprises the following sub-steps: S21. Sort each signal feature sequence according to the signal arrival time of each 5G base station; S22, using the sorted signal feature sequence as a row vector to form a signal feature matrix; S23, subtract the two signal feature sequences at adjacent moments to obtain the signal feature difference sequence, D=X t -X t-1 , where D is the signal feature gap sequence, X t is the signal characteristic sequence at the tth arrival time, X t-1 is the signal feature sequence of the t-1th arrival time, t is the number of the arrival time; S24, normalizing the elements in each signal feature gap sequence to obtain a signal gap coefficient sequence; S25, taking each signal gap coefficient sequence as a row vector to form a signal gap coefficient matrix; The multi-fusion positioning network in S3 includes: a first scale feature extraction unit, a second scale feature extraction unit, a third scale feature extraction unit, a fourth scale feature extraction unit, a first feature fusion unit, a second feature fusion unit, a first fully connected layer, a second fully connected layer, an adder A1 and an output layer; The input end of the first scale feature extraction unit is used to input a signal feature matrix; The input end of the second scale feature extraction unit is used for inputting a signal feature matrix; The input end of the third scale feature extraction unit is used for inputting a signal gap coefficient matrix; The input end of the fourth scale feature extraction unit is used for inputting a signal gap coefficient matrix; The input end of the first feature fusion unit is connected to the output end of the first scale feature extraction unit and the output end of the third scale feature extraction unit respectively, and the output end thereof is connected to the input end of the first fully connected layer; The input end of the second feature fusion unit is connected to the output end of the second scale feature extraction unit and the output end of the fourth scale feature extraction unit respectively, and the output end thereof is connected to the input end of the second fully connected layer; The first input end of the adder A1 is connected to the output end of the first fully connected layer, the second input end thereof is connected to the output end of the second fully connected layer, and the output end thereof is connected to the input end of the output layer; The output end of the output layer serves as the output end of the multi-fusion positioning network; The S4 comprises the following steps: S41. According to the terminal positions at multiple historical moments, a quadratic polynomial is used to fit multiple longitudes to obtain a terminal movement longitude trajectory, and a quadratic polynomial is used to fit multiple dimensions to obtain a terminal movement dimension trajectory; S42, predicting the longitude at the next moment using the terminal's moving longitude trajectory, and predicting the latitude at the next moment using the terminal's moving latitude trajectory, to obtain a predicted position; S43, calculating the deviation coefficient between the initial position and the predicted position; S44, performing weighted processing on the initial position and the predicted position according to the deviation coefficient between the initial position and the predicted position to obtain the terminal position at the current moment; The formula for calculating the bias coefficient between the initial position and the predicted position in S43 is: , where γ is the deviation coefficient, arctan is the inverse tangent function, long0 is the longitude of the initial position, and lat0 is the latitude of the initial position. is the longitude of the predicted location, is the latitude of the predicted location; The weighted processing formula in S44 is: , , where x is the longitude of the terminal's location at the current moment, and y is the latitude of the terminal's location at the current moment.

2. The 5G base station-based positioning method according to claim 1, characterized in that: The elements in the signal characteristic sequence in S1 include: received signal power, signal arrival time, arrival angle and channel state information.

3. The positioning method based on a 5G base station according to claim 1, characterized in that: The first scale feature extraction unit, the second scale feature extraction unit, the third scale feature extraction unit and the fourth scale feature extraction unit each include: a first convolution block, an SPP layer, a second convolution block, a Sigmoid layer, a multiplier M1 and an adder A2; The input end of the first convolution block is used as the input end of the first scale feature extraction unit, the second scale feature extraction unit, the third scale feature extraction unit or the fourth scale feature extraction unit; the input end of the SPP layer is connected to the output end of the first convolution block, and its output end is respectively connected to the input end of the second convolution block and the first input end of the adder A2; the output end of the second convolution block is respectively connected to the input end of the Sigmoid layer and the first input end of the multiplier M1; the second input end of the multiplier M1 is connected to the output end of the Sigmoid layer, and its output end is connected to the second input end of the adder A2; the output end of the adder A2 is used as the output end of the first scale feature extraction unit, the second scale feature extraction unit, the third scale feature extraction unit or the fourth scale feature extraction unit; The size of the convolution kernel in the first convolution block in the first scale feature extraction unit and the third scale feature extraction unit is 3*3; The size of the convolution kernel in the first convolution block in the second scale feature extraction unit and the fourth scale feature extraction unit is 5*5.

4. The positioning method based on 5G base station according to claim 1, characterized in that: The expression of the first feature fusion unit is: , where H1 is the output of the first feature fusion unit, Y1 is the output of the first scale feature extraction unit, and Y3 is the output of the third scale feature extraction unit. For Hadamard; The expression of the second feature fusion unit is: , where H2 is the output of the second feature fusion unit, Y2 is the output of the second scale feature extraction unit, and Y4 is the output of the fourth scale feature extraction unit.

5. The positioning method based on a 5G base station according to claim 1, characterized in that: The expression of the output layer is: , , where long0 is the longitude of the initial position, lat0 is the latitude of the initial position, and x i is the i-th eigenvalue output by adder A1, ω 1,i is the longitude weight of the i-th eigenvalue, b 1,i is the longitude bias of the i-th eigenvalue, ω 2,i is the latitude weight of the i-th eigenvalue, b 2,i is the latitude bias of the i-th eigenvalue, N is the number of eigenvalues ​​output by adder A1, and i is a positive integer.

Images