Interference source positioning method based on neural network and ray tracing model

Abstract

The invention discloses an interference source positioning method based on a neural network and a ray tracing model, and belongs to the field of wireless communication technology and source positioning. According to the method, the neural network is combined with traditional geometric calculation, and the trained neural network can fit an originally complicated mathematical calculation or iterative optimization process to quickly position the interference source. Complex electromagnetic equation calculation is replaced by using a geometric calculation and neural network mixed mode, so that the calculation complexity of interference source positioning processing is reduced, and the prediction efficiency is improved; when an interference source position is predicted, space coordinates can be positioned, path prediction can be performed on main interference signal directions, an advancing path of an interference signal can be drawn, and path information can be acquired. The method is a technology for fusion development of a wireless communication technology and a deep learning / neural network technology, and provides a thought for improvement of a source positioning processing method.

Description

Technical Field

[0001] The present invention relates to the fields of wireless communication technology and source positioning, and in particular to an interference source positioning method based on a neural network and a ray tracing model. Background Art

[0002] With the rapid development of wireless communication technology, interference source positioning has become one of the key technologies to ensure normal communication and information security.

[0003] Traditional interference source location technology mainly measures the arrival time, direction, strength and other parameters of the wireless signal at the receiving end, and determines the location of the interference source through complex mathematical calculations using the triangulation positioning principle or other positioning algorithms. Its adaptability to complex environments and the robustness and accuracy of the calculation results need to be improved.

[0004] Existing improved methods, such as "An iterative implementation-based approach for joint source localization and association under multipath propagation environments", proposed an iterative implementation algorithm based on the minimum mean square error framework, combined with semi-unitary and sparsity constraints and subspace technology, to improve the positioning accuracy and estimation robustness of unknown sources in complex multipath propagation environments. However, this algorithm involves a very complex iterative optimization process with high computational cost, which affects the overall algorithm's operating efficiency, scalability and generalization, and is limited in large-scale problems and real-time applications. Similarly, in order to ensure high positioning accuracy for source positioning in existing literature, it is necessary to solve complex electromagnetic propagation equations or design complex iterative optimization processes. In view of this, it is necessary to propose a technology that focuses on the combination of neural networks and traditional positioning algorithms to balance positioning accuracy and algorithm operation efficiency. Summary of the invention

[0005] The present invention provides a method for locating interference sources based on a neural network and a ray tracing model. The method combines a neural network with traditional geometric calculations. The trained neural network can fit the originally very complex mathematical calculations or iterative optimization processes, and quickly locate the interference source. While ensuring the positioning accuracy, the computational complexity of the interference source positioning processing is reduced to improve the operating efficiency of the interference source positioning processing.

[0006] The technical solution adopted by the present invention is:

[0007] A method for locating interference sources based on a neural network and a ray tracing model, the method comprising the following steps:

[0008] Step 1, using 3D reconstruction to obtain structural information of the scene, the structural information includes information of a number of triangular facets represented by the spatial coordinates of three points;

[0009] Step 2: Acquire a wireless signal data set in the scene, the data set includes a number of data points, each data point includes the spatial coordinates Tx of the transmitter, the spatial coordinates Rx of the receiver, and the direct reflection path information between Tx and Rx; wherein the direct reflection path information includes: the number of paths, the reflection order of each path, the departure angle, the arrival angle and the path delay;

[0010] Step 3, constructing a training set for training the ray tracing model;

[0011] The structural information of the scene obtained in step 1 is used as the input data of the ray tracing model, and the structural information of the scene is stored as a list, each element in the list is the information of the triangular face element, and the one-hot encoding is performed according to the index of the triangular face element in the list to obtain the encoding vector of each triangular face element; and the Tx, Rx and the direct reflection path information between Tx and Rx of the data point corresponding to each triangular face element are used as the true value of the ray tracing model;

[0012] Step 4, training the ray tracing model based on the training set;

[0013] Each time training, a data point in the training set is randomly selected as the true value;

[0014] Taking the current data point Tx as the starting point, according to the departure angle of each path in the direct reflection path information of the current data point, a ray is emitted in the corresponding direction; and the ray tracing model is used to perform ray stepping iterative processing on each emitted ray, wherein each iterative processing includes:

[0015] The intersection point of the current ray and the scene geometry is calculated by using the ray-triangle intersection method, and the triangle surface element corresponding to the intersection point is obtained to obtain the input of the neural network structure of the ray tracing model; wherein the input of the neural network structure includes: the unique hot encoding f of the triangle surface element i , the surface normal vector n i , and the distance vector b composed of the distance between the transmitter, the receiver, the intersection point and the triangle element i , where the subscript i is the index of the triangle face;

[0016] The neural network structure is a stacked fully connected layer structure, including at least three fully connected layers, where the input of the last fully connected layer also includes the direction information of the incident ray The output of the last fully connected layer is the corresponding Euler-Rodriguez parameter; it is then converted into the corresponding rotation matrix A based on the direction information of the incident ray Calculate the direction information of the outgoing ray And the corresponding outgoing ray equation is constructed as: Among them, t represents the time variable of the ray step distance, the subscript of each parameter in the equation is the ray index, and the superscript is the iteration number index. Indicates the outgoing ray corresponding to the current iteration processing, represents the starting point of the outgoing ray, which corresponds to the intersection point calculated by the ray-triangle intersection method;

[0017] Determine whether the reflection order of the current outgoing ray reaches the same as the reflection order corresponding to the true value of the current data point. If so, determine that the propagation of the current outgoing ray is terminated and no ray stepping is performed; otherwise, use the current outgoing ray as the new incident ray and perform the next round of ray stepping iteration processing;

[0018] After all rays have terminated their propagation, a set of predicted paths is obtained, and the training loss is calculated by comparing the predicted path with the real path corresponding to the direct reflection path information in the true value of the current data point, so as to update the network parameters of the neural network structure in the form of back-propagation until the preset training end condition is met, thereby obtaining a trained ray tracing model;

[0019] Step 5: For the trained ray tracing model, the receiving end position, the incident direction and delay information of the interference signal received by the receiving end are used as input information, and the predicted path set obtained based on the ray tracing model is used as the interference signal path set;

[0020] The interference source is then located based on the last-order ray of each path in the interference signal path set, and the position of the interference source is determined based on the point closest to multiple rays.

[0021] Furthermore, the distance vector Among them, x tx With x rx They represent the spatial coordinates of the sending end and the receiving end respectively, and sdf(·) represents the distance from the spatial point to the plane.

[0022] Further, using The ray-triangle intersection algorithm computes the intersection point of the current ray with the scene geometry.

[0023] Furthermore, the activation function of each fully connected layer of the neural network structure is set to the ReLU activation function.

[0024] Furthermore, the expression of the rotation matrix A is:

[0025]

[0026] Among them, (θ, nx, ny, nz) is the Euler-Rodriguez parameter output by the last fully connected layer of the neural network structure, θ is the rotation angle of the rotation axis, (nx, ny, nz) is the direction of the rotation axis, and I is the unit matrix.

[0027] Furthermore, in step 4, the training loss is specifically set to:

[0028] The predicted path is matched with the actual path one by one according to the departure angle and reflection order in the direct reflection path information of the data point; at the receiving end, the training loss is obtained by the weighted sum of the cosine error of the arrival direction and the mean square error of the path delay.

[0029] Furthermore, in step 5, the least square method is used to solve the point closest to the multiple rays.

[0030] The technical solution provided by the present invention brings at least the following beneficial effects:

[0031] (1) Compared with the interference source positioning technology in the field of wireless communication technology, the present invention uses a hybrid method of geometric calculation and neural network to replace complex electromagnetic equation calculations, thereby reducing the overall calculation complexity when the method of the present invention is actually used for positioning after training is completed, and improving the prediction efficiency.

[0032] (2) According to the ray tracing model used in training of the present invention, when predicting the location of the interference source, in addition to being able to locate the spatial coordinates, it is also possible to predict the path of the main interference signal direction, draw the travel path of the interference signal, and obtain path information.

[0033] (3) The present invention does not use a pure neural network to predict the entire wireless ray propagation process, but combines traditional geometric calculations with neural networks to enhance the interpretability of the entire ray tracing process. In addition, since strong geometric prior constraints are given in traditional calculations, only a simple network structure is needed to make a relatively accurate prediction of the ray effect. This is a technology that integrates the development of wireless communication technology and deep learning / neural network technology, and provides ideas for improving source positioning processing methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0035] Figure 1 A schematic diagram of a scene used in a specific embodiment of the present invention;

[0036] Figure 2 It is an overall flow chart of the interference source positioning method based on neural network and ray tracing model of the present invention.

[0037] Figure 3 A ray tracing flow chart of a ray tracing channel model based on a neural network according to the present invention;

[0038] Figure 4 It is a schematic diagram of the process of ray stepping and ray receiving in steps S6-S9 of the present invention;

[0039] Figure 5 A neural network structure diagram used in the "ray-surface interaction" in step S8 of the present invention;

[0040] Figure 6 This is a graph showing the prediction effect and average distance error result of the transmitter (interference source) of the test data set in a specific embodiment of the present invention. DETAILED DESCRIPTION

[0041] To make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be described in detail and completely in conjunction with the drawings in the implementation of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings can be arranged and designed using different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the present application for protection, but merely represents selected embodiments of the present invention.

[0042] An embodiment of the present invention provides an interference source positioning method based on a neural network and a ray tracing model. The method focuses on the combination of a neural network and a traditional positioning algorithm, and balances the positioning accuracy and algorithm operation efficiency by using a neural network to replace complex mathematical equations.

[0043] The data set and specific configuration used in this embodiment are as follows:

[0044] like Figure 1 As shown in the scenario, three indoor scenarios Office1, Office2 and Meeting Room provided by the CloudRT official website were used for simulation testing, and the scenario information was represented by triangular face elements. For each scenario, by randomly sampling the transmitter position Tx and the receiver position Rx, and using CloudRT online simulation, 200 Tx-Rx point pairs and their direct and reflected path information were obtained, including the number of paths, the arrival angle, departure angle and path delay of each path. 90% of them were used as the training set for training, and the other 10% were used as the test set for predicting the transmitter Tx.

[0045] According to the above parameter settings, Figure 2 The overall process of an interference source positioning method based on a neural network and a ray tracing model provided by an embodiment of the present invention, specifically, the simulation experiment of this embodiment includes the following steps:

[0046] S1. Use existing 3D reconstruction technology to obtain scene structure information. In this embodiment, the scene structure information is directly provided by CloudRT, using triangular face elements as basic elements, and each triangular face element is represented by the spatial coordinates of three points.

[0047] S2. Measure or simulate the wireless signal data set in the scenario. This example uses the CloudRT platform for simulation and obtains a data set containing 200 data points as the true value. Each data point consists of: the spatial coordinates of the transmitter Tx, the spatial coordinates of the receiver Rx, and the direct reflection path information between Tx and Rx (including the number of paths, the reflection order of each path, the departure angle, the arrival angle, and the path delay).

[0048] S3, data reading and preprocessing. The input data required by the present invention are: the scene geometry information in step S1 (composed of multiple triangular facet information) and the wireless signal data set in step S2. After reading the data, the scene geometry information is stored as a list, and each element in the list is a triangular facet represented by the spatial coordinates of three points. Unique hot encoding is performed according to the index of the facet in the list to obtain a coding vector uniquely corresponding to each facet.

[0049] The wireless signal data set is stored separately according to the information described in S2 to be used as the true value during neural network training, and is divided into a training set and a test set in a ratio of 9:1.

[0050] S4. Train the ray tracing model based on the neural network. Use the training set divided in step S3 to train and optimize the ray tracing model. Each time training, randomly select a data point in the training set as the true value. Figure 3 The specific process of the neural network-based ray tracing channel model is demonstrated, which can be described as three steps: ray emission, ray stepping, and ray reception.

[0051] S5, ray emission. Taking the spatial coordinate Tx of the transmitting end as the starting point, rays are emitted in the corresponding direction according to the departure angle of each path in the data.

[0052] S6, ray stepping. The ray stepping iteration is performed for each emitted ray. Figure 4The specific process of ray stepping and ray reception is demonstrated. The first two modules are an iterative process of ray stepping, which mainly consists of two stages: "ray-triangle intersection" and "ray-surface interaction".

[0053] S7, ray-triangle intersection. Calculate the first interaction between the current ray and the scene geometry through geometric calculation. The scene geometry information is represented by triangle face elements, using The ray-triangle intersection algorithm can quickly calculate the intersection point between the current ray and the triangle surface element.

[0054] Specifically, for the kth ray emitted, assuming that it has been reflected r times, use the symbol To express, in order to construct the intersection equation of the ray and the triangle, the ray It can be expressed as: of the form, where is the starting point of the ray, is the direction vector of the ray, and t is the time variable of the ray step distance. For a certain triangle face in the environment, the three vertex coordinates of the triangle are used to represent it, p0, p1, and p2. When the ray intersects the triangle, the intersection point satisfies the ray equation and is located inside the triangle, so we can get formula (1):

[0055]

[0056] By solving equation (1), we can get the solution of the intersection point. We take the minimum t that satisfies t>0, which is the time delay of the current ray step. With f i That is, the intersection point and the corresponding triangle element when the ray interacts with the spatial triangle element for the first time.

[0057] S8, ray-surface interaction. A three-layer fully connected neural network is used to replace the calculation of complex electromagnetic equations. Figure 5 The specific neural network structure used is shown. The specific function of the network is to fit the interaction (such as reflection) that occurs when the ray intersects the scene surface. Each layer of the network uses the Pytorch framework, including a linear layer and a ReLU activation function. The detailed network model design is as follows:

[0058] The input of the first layer of the network is the surface f that currently interacts with the ray i The one-hot encoding f i , the surface normal vector n i And the three-dimensional vector b composed of three signed distances i . b i The specific content of is given by formula (2)

[0059]

[0060] where x tx With x rx They represent the spatial coordinates of the sender and the receiver respectively, and sdf(x,f) represents the distance from the spatial point x to the plane f. i The three quantities in the figure represent the transmitter, receiver, intersection point and surface f. i The point-to-plane distance relationship.

[0061] In terms of network structure, the first fully connected layer converts the input into a 256-dimensional feature vector; the second layer inputs a 256-dimensional feature vector and also outputs a 256-dimensional feature vector; in the third layer, the input is a 256-dimensional feature vector and a 3-dimensional direction vector. (direction information of the incident ray), the output is a four-dimensional vector ρ = (θ, nx, ny, nz), which is the Euler-Rodriguez parameter, where θ is the rotation angle of the rotation axis, (nx, ny, nz) represents the direction of the rotation axis, that is, nx, ny, nz are the values ​​of the three directional dimensions of the rotation axis.

[0062] The parameters are converted into a rotation matrix A using formula (3).

[0063]

[0064] Where I is the identity matrix.

[0065] Then through Calculate the direction information of the outgoing ray At this time, the outgoing ray equation can be constructed as: Completes one ray step.

[0066] S9, ray reception and termination detection. For the outgoing ray, it is necessary to detect whether to stop ray stepping. Judge according to the reflection order of each path provided by the training data. When the reflection order of the outgoing ray is the same as the actual reflection order of the training data, it is considered that the outgoing ray will terminate during this propagation and no ray stepping will be performed. For the ray that does not meet the termination condition, it is regarded as the new incident ray, and the ray stepping process of steps S6, S7, and S8 is re-executed to perform the next iteration until the obtained outgoing ray meets the termination condition.

[0067] S10, loss calculation, back propagation. After all rays are terminated, a set of predicted paths is obtained. By calculating the loss of the predicted path and the true path, back propagation is performed to train the neural network used in step S8.

[0068] First, the predicted path is matched with the real path one by one according to the departure angle and reflection order given by the data. At the receiving end, the cosine error of the arrival direction and the mean square error of the path delay together form the loss function, where the real arrival direction is calculated by the arrival angle given by the data, and the predicted arrival direction is the direction corresponding to the last ray in the predicted path; the real path delay is directly given by the data, and the predicted path delay is obtained by accumulating the time t calculated by each ray on the path in step S7.

[0069] In this embodiment, the loss calculation function used is shown in formula (4):

[0070]

[0071] Among them, w t With w d are preset weights for delay and arrival direction, which are artificially set hyperparameters to balance the size gap between the two items. They represent the delay of the real path and the delay of the predicted path respectively, d represents the real arrival direction, Represents the predicted arrival direction. The first term of this loss function aims to minimize the delay error between the path and the true path, and the second term aims to minimize the cosine loss between the predicted receiving ray direction and the actual receiving ray direction.

[0072] S11, interference signal path prediction. The ray tracing model based on the neural network is trained through steps S4-S10 to obtain a converged model. For the converged model, the input information is modified to: the position of the receiving end, the incident direction and delay information of the main interference signal received by the receiving end.

[0073] According to the input information, the ray is emitted at the receiving end and the ray is traced using the above model. The difference is that in step S9, the algorithm determines whether the ray should be terminated by comparing the current step time of each path with the real delay information. When the step time exceeds the real time, it means that the ray has been terminated and no more iterations are performed.

[0074] After all rays are terminated, the resulting path set is the interference signal path predicted by the model.

[0075] S12. Predict the location of the interference source based on the predicted path. According to the path information predicted in step S11, assuming that there are n paths in total, take the last-order ray of each path for positioning. The positioning method is: use the least squares idea to solve the point closest to the n rays. The specific algorithm principle is described as follows:

[0076] First, let’s analyze the simple problem of finding the closest point between two rays. The equations of the two rays can be written as:

[0077] r1=o1+t1·d1; r2=o2+t2·d2 (5)

[0078] Among them, o1 and o2 represent the incident directions of the two rays, d1 and d2 represent the exit directions of the two rays, and t1 and t2 represent the time variables of the step distances of the two rays.

[0079] Assuming that (x1, y1, z1) and (x2, y2, z2) are points on two rays, the constraints of the xyz axes can be obtained according to the ray equations:

[0080]

[0081] At this time, in order to solve the closest point, we can assume that if the two rays intersect, then (x1, y1, z1) and (x2, y2, z2) represent the same point; otherwise, (x1, y1, z1) and (x2, y2, z2) represent the points closest to the two rays obtained by the least squares method. Assuming this point is (x, y, z), the above constraints can be rewritten in matrix form as follows:

[0082]

[0083] make

[0084]

[0085] Through matrix operations, we can solve m=(M T M) -1 M T b, where the first three dimensions of m are the points closest to the two straight lines.

[0086] By expanding the above solution idea, we can get the solution for the closest point of any number greater than two rays. Assuming we want to solve the closest point of n rays, M is a 3n×(3+n)-dimensional matrix and b is a 3n×1-dimensional vector. We can perform the same matrix operation and get the vector m.

[0087] S13, the first three dimensions of the output vector m are the spatial coordinates of the point closest to the n straight lines. That is, the predicted spatial position coordinates of the interference source (x tx ,y tx ,z tx ).

[0088] Figure 6Visual examples of transmitter predictions in three scenarios and the average distance error on the test set are given. In the graph, the red dots represent the receiver, the blue dots represent the predicted transmitter, the green dots represent the actual transmitter, and the yellow rays are the paths predicted by the model. The data in the table show that in the three scenarios, the average distance error between the predicted transmitter location and the actual transmitter location is less than 0.1m.

[0089] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

[0090] The above are only some embodiments of the present invention. For those skilled in the art, several modifications and improvements can be made without departing from the creative concept of the present invention, which all belong to the protection scope of the present invention.

Claims

1. A method for locating interference sources based on a neural network and a ray tracing model, characterized in that: The following steps are involved: Step 1, using 3D reconstruction to obtain structural information of the scene, the structural information includes information of a number of triangular facets represented by the spatial coordinates of three points; Step 2: Acquire a wireless signal data set in the scene, the data set includes a number of data points, each data point includes the spatial coordinates Tx of the transmitter, the spatial coordinates Rx of the receiver, and the direct reflection path information between Tx and Rx; wherein the direct reflection path information includes: the number of paths, the reflection order of each path, the departure angle, the arrival angle and the path delay; Step 3, constructing a training set for training the ray tracing model; The structural information of the scene obtained in step 1 is used as the input data of the ray tracing model, and the structural information of the scene is stored as a list, each element in the list is the information of the triangular face element, and the one-hot encoding is performed according to the index of the triangular face element in the list to obtain the encoding vector of each triangular face element; and the Tx, Rx and the direct reflection path information between Tx and Rx of the data point corresponding to each triangular face element are used as the true value of the ray tracing model; Step 4, training the ray tracing model based on the training set; Each time training, a data point in the training set is randomly selected as the true value; Taking the current data point Tx as the starting point, according to the departure angle of each path in the direct reflection path information of the current data point, a ray is emitted in the corresponding direction; and the ray tracing model is used to perform ray stepping iterative processing on each emitted ray, wherein each iterative processing includes: The intersection point of the current ray and the scene geometry is calculated by using the ray-triangle intersection method, and the triangle surface element corresponding to the intersection point is obtained to obtain the input of the neural network structure of the ray tracing model; wherein the input of the neural network structure includes: the unique hot encoding f of the triangle surface element i , the surface normal vector n i , and the distance vector b composed of the distances between the transmitter, receiver, intersection point and triangle element i , where the subscript i is the index of the triangle face; The neural network structure is a stacked fully connected layer structure, including at least three fully connected layers, where the input of the last fully connected layer also includes the direction information of the incident ray The output of the last fully connected layer is the corresponding Euler-Rodriguez parameter; it is then converted into the corresponding rotation matrix A based on the direction information of the incident ray Calculate the direction information of the outgoing ray And the corresponding outgoing ray equation is constructed as: Among them, t represents the time variable of the ray step distance, the subscript of each parameter in the equation is the ray index, and the superscript is the iteration number index. Indicates the outgoing ray corresponding to the current iteration processing, Indicates the starting point of the outgoing ray; Determine whether the reflection order of the current outgoing ray reaches the same as the reflection order corresponding to the true value of the current data point. If so, determine that the propagation of the current outgoing ray is terminated and no ray stepping is performed; otherwise, use the current outgoing ray as the new incident ray and perform the next round of ray stepping iteration processing; After all rays have terminated their propagation, a set of predicted paths is obtained, and the training loss is calculated by comparing the predicted path with the real path corresponding to the direct reflection path information in the true value of the current data point, so as to update the network parameters of the neural network structure in the form of back-propagation until the preset training end condition is met, thereby obtaining a trained ray tracing model; Step 5: For the trained ray tracing model, the receiving end position, the incident direction and delay information of the interference signal received by the receiving end are used as input information, and the predicted path set obtained based on the ray tracing model is used as the interference signal path set; The interference source is then located based on the last-order ray of each path in the interference signal path set, and the position of the interference source is determined based on the point closest to multiple rays.

2. The method according to claim 1, characterized in that The distance vector is specifically set as: Among them, x tx With x rx They represent the spatial coordinates of the sending end and the receiving end respectively, and sdf(·) represents the distance from the spatial point to the plane.

3. The method according to claim 1, characterized in that use The ray-triangle intersection algorithm computes the intersection point of the current ray with the scene geometry.

4. The method according to claim 1, characterized in that The activation function of each fully connected layer of the neural network structure is set to the ReLU activation function.

5. The method according to claim 4, characterized in that The neural network structure is set to three fully connected layers, and the number of output channels of the first and second fully connected layers are both set to 256 dimensions.

6. The method according to claim 1, characterized in that The expression of the rotation matrix A is: Among them, (θ, nx, ny, nz) is the Euler-Rodriguez parameter output by the last fully connected layer of the neural network structure, θ is the rotation angle of the rotation axis, (nx, ny, nz) is the direction of the rotation axis, and I is the unit matrix.

7. The method according to claim 1, characterized in that The training loss in step 4 is specifically set to: The predicted path is matched with the actual path one by one according to the departure angle and reflection order in the direct reflection path information of the data point; at the receiving end, the training loss is obtained by the weighted sum of the cosine error of the arrival direction and the mean square error of the path delay.

8. The method according to claim 1, characterized in that In step 5, the least square method is used to find the point closest to multiple rays.

Images