Base station self-calibration method combined with direct non-line-of-sight error mitigation

Abstract

The invention provides a base station self-calibration method in combination with direct non-line-of-sight error mitigation, which directly mitigates non-line-of-sight errors through a parallel deep regression network, and iteratively updates the position of a base station in combination with global optimization and local optimization strategies. The problems that an existing base station self-calibration method is limited by non-line-of-sight shielding and a multipath effect, the convergence speed is low and the self-calibration precision is low are effectively solved, and a more efficient and accurate solution is provided for ultra-wideband base station position self-calibration in a complex environment.

Description

Technical Field

[0001] The invention relates to the field of radio signal processing, and in particular to a base station self-calibration method combined with direct non-line-of-sight error mitigation. Background Art

[0002] Ultra-wideband (UWB) positioning technology has become the most commonly used indoor positioning technology due to its high precision and anti-multipath interference capabilities. UWB technology uses high-frequency pulse signals for positioning and can provide centimeter-level accuracy in complex indoor environments. It is particularly suitable for application scenarios with high precision requirements, such as smart manufacturing, robot navigation, and IoT positioning. Its main advantage is that it can effectively solve the performance degradation problem of traditional positioning technology in non-line-of-sight (NLOS) environments, and has strong anti-interference and low power consumption.

[0003] Base station self-calibration is a key part of UWB positioning system. It is usually necessary to accurately determine the location of the base station to ensure the accuracy of the positioning system. However, due to environmental factors, signal multipath effects and NLOS errors, base station calibration is a tedious and challenging task. Traditional base station calibration methods usually rely on manual measurements or assume known environmental information, lacking automation and adaptability. Common base station self-calibration technologies currently include methods based on least squares optimization, multidimensional scaling and Kalman filtering. For example, Dalian Haoru Technology Co., Ltd. applied for a "UWB base station self-calibration method based on multidimensional scaling" (publication number: CN 117528413), which constructs an inner product matrix through distance measurement between base stations, reduces the dimension to calculate the coordinates of the target base station, simplifies the calibration process, and improves deployment efficiency and configuration flexibility; Changzhou Institute of Technology applied for a "UWB positioning system base station self-calibration method based on Kalman filtering" (publication number: CN 107708204), which processes the distance and height data between base stations through Kalman filtering, recursively solves the coordinates of each base station, achieves high-precision self-calibration, simplifies system layout, and improves mobility. However, these methods are easily affected by errors in dynamic environments, especially the interference of the NLOS effect, which makes it difficult for traditional base station self-calibration methods to converge, or converge to incorrect values;

[0004] Therefore, it is urgent to improve the existing technology. Summary of the invention

[0005] In view of the problem that the UWB indoor positioning system needs to manually measure and calibrate the positions of the deployed base stations, traditional methods usually rely on manual measurement and precise calculation, which is time-consuming and easily affected by human errors. At the same time, the existing base station self-calibration methods are limited by NLOS occlusion and multipath effects, making it difficult to perform effective calibration or converge to the correct position, and there is a challenge of low self-calibration efficiency. In response to these problems, the present invention proposes a base station self-calibration method combined with direct non-line-of-sight error mitigation. This method directly identifies and corrects NLOS errors through parallel deep regression prediction, provides a smaller iterative initial value, and thus accelerates the base station self-calibration calculation. By combining global and local optimization strategies, efficient and accurate base station position automatic calibration is achieved, effectively improving the accuracy and convergence speed of self-calibration.

[0006] The technical solution adopted by the present invention to solve its technical problem is:

[0007] A base station self-calibration method combined with direct non-line-of-sight error mitigation comprises the following steps:

[0008] S1: First, the distance value d between each base station is obtained by bilateral two-way flight time ranging method. ij , for any two base stations i and j, their coordinates are p i =(x i ,y i ,z i ) and p j =(x j ,y j ,z j ), the ranging value satisfies:

[0009]

[0010] S2: Perform direct non-line-of-sight error mitigation on the ranging values ​​between all base stations to obtain the ranging matrix D after error mitigation m ={dm ij}, providing a smaller initial value for optimization iteration input, accelerating the calculation of the base station self-calibration position;

[0011] S3: Construct global error function:

[0012]

[0013] S4: Local optimization, all base stations are divided into several subsets, and the location is solved using local graph optimization in each subset, that is, three base stations form a triangle, and the solution is solved using the intersection of two spheres; four base stations form a tetrahedron, and the solution is solved using the intersection of three spheres;

[0014] S5: Local position solution, select N ≥ 3 base stations, and initialize the position using the following rules:

[0015] The coordinates of fixed base station 1 are (0,0,0), so:

[0016]

[0017] Next, determine the location of base station 2, assuming it is located on the x-axis, that is, (d 12 ,0,0);

[0018] The position of base station 3 is determined again, and it is limited to the xy-plane according to the two sphere intersection method:

[0019]

[0020] If N=3, there are three base stations, forming a triangle, and the solution stops here.

[0021] S6: Determine the location of base station 4, based on the distance values ​​d between base station 4 and base stations 1, 2, and 3 respectively. 14 , d 24 , d 34 , solve its position by intersecting three spheres:

[0022] First, establish the equation for the intersection of the three spheres:

[0023]

[0024] After expanding each equation and performing elimination, we get:

[0025]

[0026] Simplifying these two linear equations is:

[0027]

[0028] The linear equation is iteratively solved by the Levenberg-Marquardt method to obtain the optimal coordinates (x4, y4, z4) of the base station 4;

[0029] S7: Global optimization, using the global error E as the objective function, adjusting the relative positions of each local geometric figure so that their overlapping areas satisfy the consistency constraint, and using the genetic algorithm to perform nonlinear optimization on all base station coordinates;

[0030] S8: Stop when the global error E converges to the set threshold ε or reaches the maximum number of iterations;

[0031] S9: Output the relative coordinates p1, p2, ..., p of N base stations N ;

[0032] Furthermore, the bilateral two-way time-of-flight ranging method comprises the following steps:

[0033] S11: First, a communication polling frame is sent from base station i to base station j, wherein base station i and base station j respectively record their respective hardware timestamps: t0 and t1;

[0034] S12: Delay at fixed time D round1 Afterwards, base station j replies with a response frame, and the corresponding timestamps are t2 and t3;

[0035] S13: Delay D for another fixed time round2 Afterwards, base station i sends the last frame to base station j, recording the timestamps as t4 and t5;

[0036] S14: These timestamps are then used to calculate the bilateral two-way flight time:

[0037]

[0038] Calculate the distance between base station i and base station j based on the bilateral two-way flight time:

[0039] d ij =Timediff×c (9)

[0040] Where c is the propagation speed of UWB radio waves in the current environment;

[0041] Further, the direct non-line-of-sight error mitigation is achieved by parallel depth regression prediction, and the parallel depth regression prediction specifically includes the following steps:

[0042] S21: first, the ultra-wideband base station receives the signal sent from other base stations, and extracts the channel impulse response data and seven artificial features from the signal;

[0043] S22: input the channel impulse response data into the first branch to extract deep features, and input the seven artificial features into the second branch to extract further features;

[0044] S23: The channel impulse response data is first processed by a multi-layer deep convolutional feature extraction network in the first branch. Secondly, the processed feature map is further extracted with a squeezed attention mechanism. Finally, the obtained key features are flattened and input into a long short-term memory network layer to extract temporal features, and the finally processed feature map is input into a feature fusion layer.

[0045] S24: The seven artificial features are input into the second branch, and are first processed by three fully connected layers to capture the nonlinear relationship in the input data, and then activated by the sigmoid function to normalize the feature values ​​to the range [0, 1], and finally input into the feature fusion layer;

[0046] S25: fusing the deep features of the channel impulse response data extracted by the first branch and the second branch and the further features of the seven artificial features through a feature fusion layer;

[0047] S26: The fused features are reduced in dimension through two fully connected layers and finally input into the regression layer to obtain the predicted value Δd of the ranging error. ij , and mitigate the ranging error:

[0048] dm ij =d ij -Δd ij (10)

[0049] Furthermore, the following ranging constraint relationship exists between the base stations:

[0050] The distance d between each base station and base station 1 1i A spherical constraint is provided:

[0051]

[0052] The distance between two base stations ij An elliptical surface constraint is provided;

[0053] The elliptical surface constraint is defined as follows: Assume that the coordinates of the base station k to be solved are p k =(x k ,y k ,z k ), the distance between base station k and base station i is d ik , the distance between base station k and base station j is d jk , we get the following two spherical constraints:

[0054]

[0055]

[0056] These two constraints indicate that base station k must be located on a sphere centered on base stations i and j, with radii d and ik and d jk ;

[0057] After splitting and merging, the equation of the elliptical surface is obtained:

[0058]

[0059] Furthermore, the process of solving the linear equation by Levenberg-Marquardt iteration is as follows:

[0060] S61: Setting the damping factor λ>0 and the stop condition;

[0061] S62: Construct an error function and define the error function of each pair of base stations (i, j):

[0062] e ij =||p i -p j ||-d ij (15)

[0063] Combine all error terms into a single error vector:

[0064] e=[e 12 e 13 ... e N-1,N ] (16)

[0065] S63: Define the objective function as the sum of the squares of the error vector:

[0066] L=e T e (17)

[0067] S64: Construct the Jacobian matrix J, each item of which is the partial derivative of the error function with respect to the kth coordinate:

[0068]

[0069] For e ij , its partial derivative form is:

[0070]

[0071] S65: The Levenberg-Marquardt method is updated as follows:

[0072] Δp=-(J T J+λI) -1 J T e (20)

[0073] Where I is the identity matrix;

[0074] Update the coordinates to:

[0075] p←p+Δp(21)

[0076] S66: Iterate to find the optimal value. First, calculate the new objective function value L. If the error decreases, reduce λ and accept the update; otherwise, increase λ and reject the update.

[0077] Repeat steps S64 and S65 until convergence, and obtain the optimal coordinates of base station 4;

[0078] Furthermore, in the parallel deep regression prediction, feature extraction of the first branch and the second branch is performed simultaneously, and when the first branch has no channel impulse response data input, if the second branch has seven artificial features input, its feature extraction will still be performed, and the output of the feature fusion layer is only the feature map after the seven artificial features are extracted; when the second branch does not have seven artificial features input, if the first branch has channel impulse response data input, it will still be performed, and the output of the feature fusion layer is only the feature map after the channel impulse response data is extracted;

[0079] In the multi-layer deep convolutional feature extraction network, each convolutional layer is sequentially connected to a batch normalization layer and a maximum pooling layer;

[0080] In the second branch, a batch normalization layer, a ReLU activation function, and a dropout layer are sequentially added between the first fully connected layer and the second fully connected layer;

[0081] The seven artificial features include: a distance measurement value between base stations, a first path received signal strength difference, a signal received strength indication, a first path distance difference, a kurtosis factor, an average excess delay, and a root mean square delay spread;

[0082] The beneficial effects of the present invention are mainly manifested in:

[0083] (1) By combining global optimization and local optimization, the coordination problem of global error and local error in the multi-base station self-calibration process is solved. Global optimization improves the overall positioning accuracy, while local optimization specifically corrects the error between each pair of base stations, thereby further refining the accurate position of each base station while ensuring system consistency;

[0084] (2) The parallel deep regression network simultaneously processes the ranging errors between multiple base stations, directly identifies and mitigates the NLOS error, provides a smaller iterative initial error, and solves the problems of high computational complexity and low optimization efficiency of traditional methods. This parallel processing significantly accelerates the base station location solution process and improves the iterative efficiency of self-calibration;

[0085] (3) By combining the fast convergence of the Levenberg-Marquardt algorithm with the feature extraction capability of the deep network, the problems of slow convergence and inaccurate error processing in base station location optimization are effectively solved, and the self-calibration accuracy and optimization efficiency are significantly improved. At the same time, the accuracy of regression prediction is further enhanced through the squeezed attention mechanism and parallel processing steps in the deep regression network. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] Figure 1 It is an operation flow chart of the present invention in combination with the direct non-line-of-sight error mitigation base station self-calibration method;

[0087] Figure 2is a schematic diagram of ultra-wideband bilateral two-way time-of-flight ranging;

[0088] Figure 3 is an operational flow chart for direct non-line-of-sight error mitigation using parallel depth regression prediction;

[0089] Figure 4 It is the operation flow chart of the Levenberg-Marquardt iteration to solve the linear equation of base station location. DETAILED DESCRIPTION

[0090] In order to facilitate the understanding of the present invention, the present invention will be described more comprehensively and meticulously below in conjunction with the accompanying drawings and preferred embodiments of the present invention, but the protection scope of the present invention is not limited to the following specific embodiments.

[0091] like Figure 1 As shown, a base station self-calibration method combined with direct non-line-of-sight error mitigation in this embodiment includes the following steps:

[0092] S1: First, the distance value d between each base station is obtained by bilateral two-way flight time ranging method. ij , for any two base stations i and j, their coordinates are p i =(x i ,y i ,z i ) and p j =(x j ,y j ,z j ), the ranging value satisfies:

[0093]

[0094] S2: Perform direct non-line-of-sight error mitigation on the ranging values ​​between all base stations to obtain the ranging matrix D after error mitigation m ={dm ij}, providing a smaller initial value for optimization iteration input, accelerating the calculation of the base station self-calibration position;

[0095] S3: Construct global error function:

[0096]

[0097] S4: Local optimization, all base stations are divided into several subsets, and the location is solved using local graph optimization in each subset, that is, three base stations form a triangle, and the solution is solved using the intersection of two spheres; four base stations form a tetrahedron, and the solution is solved using the intersection of three spheres;

[0098] S5: Local position solution, select N ≥ 3 base stations, and initialize the position using the following rules:

[0099] The coordinates of fixed base station 1 are (0,0,0), so:

[0100]

[0101] Next, determine the location of base station 2, assuming it is located on the x-axis, that is, (d 12 ,0,0);

[0102] The position of base station 3 is determined again, and it is limited to the xy-plane according to the two sphere intersection method:

[0103]

[0104] If N=3, there are three base stations, forming a triangle, and the solution stops here.

[0105] S6: Determine the location of base station 4, based on the distance values ​​d between base station 4 and base stations 1, 2, and 3 respectively. 14 , d 24 , d 34 , solve its position by intersecting three spheres:

[0106] First, establish the equation for the intersection of the three spheres:

[0107]

[0108] After expanding each equation and performing elimination, we get:

[0109]

[0110] Simplifying these two linear equations is:

[0111]

[0112] The linear equation is iteratively solved by the Levenberg-Marquardt method to obtain the optimal coordinates (x4, y4, z4) of the base station 4;

[0113] S7: Global optimization, using the global error E as the objective function, adjusting the relative positions of each local geometric figure so that their overlapping areas satisfy the consistency constraint, and using the genetic algorithm to perform nonlinear optimization on all base station coordinates;

[0114] S8: Stop when the global error E converges to the set threshold ε or reaches the maximum number of iterations;

[0115] S9: Output the relative coordinates p1, p2, ..., p of N base stations N ;

[0116] like Figure 2 As shown, the bilateral two-way time-of-flight ranging method is specifically implemented by the following steps:

[0117] S11: First, a communication polling frame is sent from base station i to base station j, wherein base station i and base station j respectively record their respective hardware timestamps: t0 and t1;

[0118] S12: Delay at fixed time D round1 Afterwards, base station j replies with a response frame, and the corresponding timestamps are t2 and t3;

[0119] S13: Delay D for another fixed time round2 Afterwards, base station i sends the last frame to base station j, recording the timestamps as t4 and t5;

[0120] S14: These timestamps are then used to calculate the bilateral two-way flight time:

[0121]

[0122] Calculate the distance between base station i and base station j based on the bilateral two-way flight time:

[0123] d ij =Timediff×c (9)

[0124] Where c is the propagation speed of UWB radio waves in the current environment;

[0125] like Figure 3 As shown, the direct non-line-of-sight error mitigation is achieved by parallel depth regression prediction, and the parallel depth regression prediction specifically includes the following steps:

[0126] S21: first, the ultra-wideband base station receives the signal sent from other base stations, and extracts the channel impulse response data and seven artificial features from the signal;

[0127] S22: input the channel impulse response data into the first branch to extract deep features, and input the seven artificial features into the second branch to extract further features;

[0128] S23: The channel impulse response data is first processed by a multi-layer deep convolutional feature extraction network in the first branch. Secondly, the processed feature map is further extracted with a squeezed attention mechanism. Finally, the obtained key features are flattened and input into a long short-term memory network layer to extract temporal features, and the finally processed feature map is input into a feature fusion layer.

[0129] S24: The seven artificial features are input into the second branch, and are first processed by three fully connected layers to capture the nonlinear relationship in the input data, and then activated by the sigmoid function to normalize the feature values ​​to the range [0, 1], and finally input into the feature fusion layer;

[0130] S25: fusing the deep features of the channel impulse response data extracted by the first branch and the second branch and the further features of the seven artificial features through a feature fusion layer;

[0131] S26: The fused features are reduced in dimension through two fully connected layers and finally input into the regression layer to obtain the predicted value Δd of the ranging error. ij , and mitigate the ranging error:

[0132] dm ij =d ij -Δd ij (10)

[0133] The following ranging constraints exist between the base stations: the distance d between each base station and base station 1 1i A spherical constraint is provided:

[0134]

[0135] The distance between two base stations ij An elliptical surface constraint is provided;

[0136] The elliptical surface constraint is defined as follows: Assume that the coordinates of the base station k to be solved are p k =(x k ,y k ,z k ), the distance between base station k and base station i is d ik , the distance between base station k and base station j is d jk , we get the following two spherical constraints:

[0137]

[0138]

[0139] These two constraints indicate that base station k must be located on a sphere centered on base stations i and j, with radii d and ik and d jk ;

[0140] After splitting and merging, the equation of the elliptical surface is obtained:

[0141]

[0142] like Figure 4 As shown, the process of the Levenberg-Marquardt iterative solution of linear equations is as follows:

[0143] S61: Setting the damping factor λ>0 and the stop condition;

[0144] S62: Construct an error function and define the error function of each pair of base stations (i, j):

[0145] e ij =||p i -p j ||-d ij (15)

[0146] Combine all error terms into a single error vector:

[0147] e=[e 12 e 13 ...e N-1,N ] (16)

[0148] S63: Define the objective function as the sum of the squares of the error vector:

[0149] L=e T e (17)

[0150] S64: Construct the Jacobian matrix J, each item of which is the partial derivative of the error function with respect to the kth coordinate:

[0151]

[0152] For e ij , its partial derivative form is:

[0153]

[0154] S65: The Levenberg-Marquardt method is updated as follows:

[0155] Δp=-(J T J+λI) -1 J T e (20)

[0156] Where I is the identity matrix;

[0157] Update the coordinates to:

[0158] p←p+Δp(21)

[0159] S66: Iterate to find the optimal value. First, calculate the new objective function value L. If the error decreases, reduce λ and accept the update; otherwise, increase λ and reject the update.

[0160] Repeat steps S64 and S65 until convergence, and obtain the optimal coordinates of base station 4;

[0161] The feature extraction of the first branch and the second branch in the parallel deep regression prediction is performed simultaneously. When the first branch has no channel impulse response data input, if the second branch has seven artificial features input, its feature extraction will still be performed, and the output of the feature fusion layer is only the feature map after the seven artificial features are extracted; when the second branch does not have seven artificial features input, if the first branch has channel impulse response data input, it will still be performed, and the output of the feature fusion layer is only the feature map after the channel impulse response data is extracted;

[0162] In the multi-layer deep convolutional feature extraction network, each convolutional layer is sequentially connected to a batch normalization layer and a maximum pooling layer;

[0163] In the second branch, a batch normalization layer, a ReLU activation function, and a dropout layer are sequentially added between the first fully connected layer and the second fully connected layer;

[0164] The seven artificial features include: a distance measurement value between base stations, a first path received signal strength difference, a signal received strength indication, a first path distance difference, a kurtosis factor, an average excess delay, and a root mean square delay spread;

[0165] With the aid of the teachings presented in the foregoing description and the associated drawings, many modifications and other embodiments of the present invention will occur to those skilled in the art to which the present invention pertains. Therefore, it is to be understood that the present invention is not limited to the specific embodiments disclosed, and modifications and other embodiments are considered to be included within the scope of the appended claims. Although specific terms are used herein, they are used in a generic and descriptive sense only and not for limitation.

Claims

1. A base station self-calibration method combined with direct non-line-of-sight error mitigation, characterized by: The following steps are involved: S1: First, the distance value d between each base station is obtained by bilateral two-way flight time ranging method. ij , for any two base stations i and j, their coordinates are p i =(x i ,y i ,z i ) and p j =(x j ,y j ,z j ), the ranging value satisfies: S2: Perform direct non-line-of-sight error mitigation on the ranging values ​​between all base stations to obtain the ranging matrix D after error mitigation m ={dm ij }, providing a smaller initial value for optimization iteration input, accelerating the calculation of the base station self-calibration position; S3: Construct global error function: S4: Local optimization, all base stations are divided into several subsets, and the location is solved using local graph optimization in each subset, that is, three base stations form a triangle, and the solution is solved using the intersection of two spheres; four base stations form a tetrahedron, and the solution is solved using the intersection of three spheres; S5: Local position solution, select N ≥ 3 base stations, and initialize the position using the following rules: The coordinates of fixed base station 1 are (0, 0, 0), so: Next, determine the location of base station 2, assuming it is located on the x-axis, that is, (d 12 ,0,0); The position of base station 3 is determined again, and it is limited to the xy-plane according to the two sphere intersection method: If N=3, three base stations form a triangle, and the solution stops here S6: Determine the position of base station 4 according to the distance values ​​d between base station 4 and base stations 1, 2, and 3 respectively. 14 , d 24 , d 34 , solve its position by intersecting the three spheres: First, establish the equation for the intersection of the three spheres: After expanding each equation and performing elimination, we get: Simplifying these two linear equations is: The linear equation is iteratively solved by the Levenberg-Marquardt method to obtain the optimal coordinates (x4, y4, z4) of the base station 4; S7: Global optimization, using the global error E as the objective function, adjusting the relative positions of each local geometric figure so that their overlapping areas satisfy the consistency constraint, and using the genetic algorithm to perform nonlinear optimization on all base station coordinates; S8: Stop when the global error E converges to the set threshold ε or reaches the maximum number of iterations; S9: Output the relative coordinates p1, p2, ..., p of N base stations N .

2. The base station self-calibration method combined with direct non-line-of-sight error mitigation according to claim 1, characterized in that: The bilateral two-way time-of-flight ranging method comprises the following steps: S11: First, a communication polling frame is sent from base station i to base station j, wherein base station i and base station j respectively record their respective hardware timestamps: t0 and t1; S12: Delay at fixed time D round1 Afterwards, base station j replies with a response frame, and the corresponding timestamps are t2 and t3; S13: Delay D for another fixed time round2 Afterwards, base station i sends the last frame to base station j, recording the timestamps as t4 and t5; S14: These timestamps are then used to calculate the bilateral two-way flight time: Calculate the distance between base station i and base station j based on the bilateral two-way flight time: d ij =Timediff×c (9) where c is the propagation speed of UWB radio waves in the current environment.

3. The base station self-calibration method combined with direct non-line-of-sight error mitigation according to claim 1, characterized in that: The direct non-line-of-sight error mitigation is achieved by parallel depth regression prediction, and the parallel depth regression prediction specifically includes the following steps: S21: first, the ultra-wideband base station receives the signal sent from other base stations, and extracts the channel impulse response data and seven artificial features from the signal; S22: input the channel impulse response data into the first branch to extract deep features, and input the seven artificial features into the second branch to extract further features; S23: The channel impulse response data is first processed by a multi-layer deep convolutional feature extraction network in the first branch. Secondly, the processed feature map is further extracted with a squeezed attention mechanism. Finally, the obtained key features are flattened and input into a long short-term memory network layer to extract temporal features, and the finally processed feature map is input into a feature fusion layer. S24: The seven artificial features are input into the second branch, and are first processed by three fully connected layers to capture the nonlinear relationship in the input data, and then activated by the sigmoid function to normalize the feature values ​​to the range [0, 1], and finally input into the feature fusion layer; S25: fusing the deep features of the channel impulse response data extracted by the first branch and the second branch and the further features of the seven artificial features through a feature fusion layer; S26: The fused features are reduced in dimension through two fully connected layers and finally input into the regression layer to obtain the predicted value Δd of the ranging error. ij , and mitigate the ranging error: dm ij =d ij -Δd ij (10)。 4. The base station self-calibration method combined with direct non-line-of-sight error mitigation according to claim 1, characterized in that: The following ranging constraints exist between the base stations: The distance d between each base station and base station 1 1i A spherical constraint is provided: The distance between two base stations ij An elliptical surface constraint is provided; The elliptical surface constraint is defined as follows: Assume that the coordinates of the base station k to be solved are p k =(x k ,y k ,z k ), the distance between base station k and base station i is d ik , the distance between base station k and base station j is d jk , we get the following two spherical constraints: These two constraints mean that base station k must be located on a sphere centered on base stations i and j, with radii d and ik and d jk ; After splitting and merging, the equation of the elliptical surface is obtained:

5. The base station self-calibration method combined with direct non-line-of-sight error mitigation according to claim 1, characterized in that: The process of the Levenberg-Marquardt iteration to solve the linear equation is as follows: S61: Setting the damping factor λ>0 and the stop condition; S62: Construct an error function and define the error function of each pair of base stations (i, j): e ij =||p i -p j ||-d ij (15) Combine all error terms into a single error vector: and=[and 12 And 13 ...And N-1,N ] (16) S63: Define the objective function as the sum of the squares of the error vector: L=e T and (17) S64: Construct the Jacobian matrix J, each item of which is the partial derivative of the error function with respect to the kth coordinate: For e ij , its partial derivative form is: S65: The Levenberg-Marquardt method is updated as follows: Δp=-(J T J+λI) -1 J T e (20) Where I is the identity matrix; Update the coordinates to: p←p+Δp(21) S66: Iterate to find the optimal value. First, calculate the new objective function value L. If the error decreases, reduce λ and accept the update; otherwise, increase λ and reject the update. Repeat steps S64 and S65 until convergence, and obtain the optimal coordinates of base station 4.

6. The base station self-calibration method combined with direct non-line-of-sight error mitigation according to claim 3, characterized in that: In the parallel deep regression prediction, feature extraction of the first branch and the second branch is performed simultaneously. When the first branch has no channel impulse response data input, if the second branch has seven artificial features input, Its feature extraction will still be performed, and the output of the feature fusion layer is only the feature map after the seven artificial features are extracted; when the second branch does not have seven artificial features input, the first branch will still be performed if there is channel impulse response data input, and the output of the feature fusion layer is only the feature map after the channel impulse response data is extracted; In the multi-layer deep convolutional feature extraction network, each convolutional layer is sequentially connected to a batch normalization layer and a maximum pooling layer; In the second branch, a batch normalization layer, a ReLU activation function, and a dropout layer are sequentially added between the first fully connected layer and the second fully connected layer; The seven artificial features include: a distance measurement value between base stations, a first path received signal strength difference, a signal received strength indicator, a first path distance difference, a kurtosis factor, an average excess delay and a root mean square delay spread.

Images