A refined path loss prediction method with dual-mode network correction

Abstract

The application relates to a fine path loss prediction method of a bimodal network correction, and belongs to the technical field of wireless communication.The method is used for solving the problem that an existing method cannot balance the complexity and accuracy of path loss prediction, the method gives a statistical prediction value by using a path loss empirical model, completes common characterization of path loss, then a bimodal network is used to learn specific propagation characteristics of a communication environment from two modes of vision and features, a correction value of the path loss empirical model is obtained, and specific description of path loss at different positions is completed.The application is applied to the field of wireless communication, combines artificial intelligence and statistical methods, and can realize fine modeling of path loss.

Description

Technical Field

[0001] This invention relates to a refined path loss prediction method for dual-mode network correction, belonging to the field of wireless communication technology. Background Technology

[0002] Future mobile communication systems will not only continue to face the challenges of internet services but also need to meet the dynamic requirements of various application scenarios. To more rationally optimize network coverage and transmitter deployment for wireless communication systems, a fundamental task is to accurately characterize and model the propagation characteristics of wireless signals. Path loss is one of the most important characteristics of wireless propagation channels, effectively reflecting the power density distribution in the electromagnetic propagation environment and providing a basis for coverage and channel capacity analysis. Therefore, establishing accurate and flexible path loss models for typical propagation environments is an indispensable part of wireless communication system design.

[0003] From a modeling mechanism perspective, traditional path loss models can be divided into computationally-based deterministic models and statistically-based empirical path loss models. Deterministic models include ray tracing methods and finite-difference time-domain (FDTD) models. These methods calculate path loss during propagation based on electromagnetic computation theory. Empirical path loss models, on the other hand, are path loss models derived from measurement data through mathematical statistics, aiming to obtain a general formula for electromagnetic propagation in specific geographical environments. Empirical path loss models rely on parameter statistics, have low computational complexity, and are universally applicable to similar electromagnetic environments. Currently, there are works using machine learning for path loss prediction; however, they often require high-quality, representative features. These features often need to be extracted manually, requiring significant financial and time resources. Furthermore, these methods merely fit data, ignoring the expert knowledge specific to the communications field.

[0004] However, the aforementioned path loss prediction methods all have some shortcomings. Deterministic models have high computational complexity and long modeling time, and require high-quality detailed information about the communication environment; empirical path loss models can only provide statistical results and cannot accurately predict path loss values ​​at specific locations. Summary of the Invention

[0005] The technical problem solved by this invention is to overcome the shortcomings of existing technologies and propose a refined path loss prediction method based on dual-modal network correction. To address the issue that existing methods cannot balance the complexity and accuracy of path loss prediction, this method utilizes an empirical path loss model to provide statistical prediction values, thus characterizing the commonalities of path loss. Then, a dual-modal network is used to learn the specific propagation characteristics of the communication environment from both visual and feature-based modalities, obtaining the correction values ​​of the empirical path loss model to complete a specific description of path loss at different locations. This invention is applied in the field of wireless communication, integrating artificial intelligence and statistical analysis methods to achieve refined modeling of path loss.

[0006] The technical solution of this invention is:

[0007] A refined path loss prediction method with dual-modal network correction, the method comprising the following steps:

[0008] The first step involves extracting a bimodal communication dataset based on the coordinates of the transmitter, receiver, and buildings in the environment. This bimodal dataset includes a propagation feature set and an environmental image. The propagation feature set is obtained through a two-stage feature extraction algorithm and includes the receiver's horizontal coordinates, vertical coordinates, height, azimuth, propagation distance, number of spaced buildings, average height of spaced buildings, and communication propagation status. The initial values ​​for the number of spaced buildings and the average height of spaced buildings are 0, while the initial value for the communication propagation status is 1. An initial value of 1 indicates line-of-sight (LoS) propagation.

[0009] The second step is to calculate the coarse path loss PL between the transmitter and receiver using an empirical path loss model. e The input to the path loss empirical model is the transmitter height, receiver height, propagation distance, and communication propagation status obtained in the propagation feature set in the first step.

[0010] The third step involves using a trained bimodal network to predict the path loss correction value PL between the transmitter and receiver. γ The input to the bimodal network is the environmental image obtained in the first step, as well as the receiver's horizontal coordinates, vertical coordinates, azimuth angle, number of spaced buildings, average height of spaced buildings, and communication propagation status.

[0011] The fourth step is to convert the coarse path loss value PL obtained in the second step into a value that is not directly related to the previous step. e The path loss correction value PL obtained in step 3 γ The sums are then used to obtain the fine-grained path loss prediction value PL. eh To achieve refined path loss prediction for dual-modal network correction;

[0012] PL eh =PLγ +PL e

[0013] In the first step, there is one transmitter, N receivers, and M buildings; the specific method for obtaining the propagation feature set through a two-stage feature extraction algorithm is as follows:

[0014] Step 1.1: Calculate the two-dimensional distance between the transmitter and the i-th receiver based on their coordinates. And the azimuth angle θ between the transmitter and the i-th receiver i Let i = 1, ..., N, and the azimuth angle be the angle between the line connecting the transmitter and the i-th receiver and the horizontal plane. The position of the transmitter is considered as the origin o of a two-dimensional coordinate system, and the height of the transmitter is h. Tx The three-dimensional coordinates of the i-th receiver are represented as follows: P i Let be the two-dimensional coordinates of the i-th receiver. It is the horizontal coordinate of the i-th receiver. h is the vertical coordinate of the i-th receiver with the transmitter as the origin. i It is the height of the i-th receiver;

[0015] Step 1.2: Determine whether the line connecting the transmitter and the i-th receiver intersects the j-th building in a two-dimensional plane, where j = 1, ..., M. When making this determination, the line connecting the transmitter and the i-th receiver refers to line L1 connecting the coordinates of the transmitter and the i-th receiver. Whether it intersects the j-th building in a two-dimensional plane refers to whether line L1 intersects the boundary line L2 of the j-th building in a two-dimensional plane. The intersection determination is performed using a counter-clockwise direction test method based on computational geometry (CCW).

[0016] For the j-th building, viewed from a two-dimensional plane, it can be considered as a polygon composed of multiple boundary lines (line segments). Therefore, the vector map information B of the j-th building... j As described in the following formula:

[0017] B j =(H j K j , σ j )

[0018] Among them, H j K is the height of the j-th building. j σ is the number of boundary lines of the j-th building. For example, a cuboid building has 4 boundary lines. j Let be the set of boundary lines of the j-th building, as shown in the following formula:

[0019]

[0020] Let k be the boundary line of the j-th building. The two endpoints of the k-th boundary line are represented by the following formula:

[0021]

[0022] The CCW test method is used to determine the connection between the transmitter and the i-th receiver in a two-dimensional plane. Is it related to the k-th boundary line of the j-th building? The formula for calculating the intersection is:

[0023]

[0024] Where |·| represents determinant operation, CCW(a j，k b j,k p i The endpoint set (a) is determined. j，k b j，k p i Is it arranged in a counter-clockwise direction? When CCW(a) j，k b j,k p i When the calculation result of ) is positive, the endpoint set (a) j，k b j,k p i Arranged in counter-clockwise order, when CCW(a) j，k b j，k p i When the result of the calculation of ) is negative, the endpoint set (a j，k b j,k p i Arranged in clockwise order, when CCW(a) j，k b j,k p i When the result of the calculation is zero, it means that P i On the k-th boundary line of the j-th building or on the extension of the k-th boundary line of the j-th building;

[0025] When the endpoint set (a j，k b j,k p i ) and endpoint set (a j，k b j,k ,o) are arranged in the reverse rotation order, that is

[0026] CCW(a j，k b j,k p i )·CCW(a j，k b j,k,o)<0

[0027] And the endpoint set (o, p) i a j，k ) and endpoint set (o, p i b j，k Arranged in the reverse rotation order, i.e.

[0028] CCW(o, p) i a j，k )·CCW(o, p i b j,k ) < 0

[0029] If both are true, then and An intersection occurs; proceed to step 1.3.

[0030] At this time, endpoint a j，k and b j,k lie in At both ends, o and p i lie in The line connecting the transmitting end and the i-th receiving end intersects with the j-th building in a two-dimensional plane, and the coordinates of the two-dimensional intersection point are recorded as (x0, y0).

[0031] If any one of the conditions is not met, then and If no intersection occurs, update the boundary line k of the j-th building. If an intersection occurs, proceed to step 1.3. If the line connecting the transmitter and the i-th receiver does not intersect the boundary line of the j-th building in the two-dimensional plane, then there is no three-dimensional intersection, and proceed to step 1.4.

[0032] Step 1.3: Determine whether the line connecting the transmitter and the i-th receiver intersects the j-th building in the three-dimensional plane, and calculate the height h of the three-dimensional intersection point. I As described in the following formula:

[0033]

[0034] If h I The height H of the j-th building is less than j If the line connecting the transmitter and the i-th receiver intersects the j-th building in the three-dimensional plane, it indicates that the transmitter's signal is blocked by the j-th building. Therefore, the number of interval buildings is updated (increased by 1); the average height of the interval buildings is updated (increased by H). j Update communication propagation status L i =0 indicates non-line-of-sight (NLoS) propagation; finally, exit the intersection judgment of the j-th building and proceed to step 1.4.

[0035] Step 1.4: Update building j. Repeat steps 1.2 and 1.3 above until all buildings j = 1, ..., M are traversed. Finally, divide the updated average height of the interval buildings by the number of updated interval buildings to obtain the final average height of the interval buildings.

[0036] Step 1.5: Repeat steps 1.1 to 1.4 above, traversing all receivers i = 1, ..., N, and concatenating all acquired features to obtain the propagation feature set of all receivers. The propagation feature set of any receiver is represented as follows:

[0037]

[0038] In the first step, the environmental image represents the terrain information surrounding the i-th receiver. The area captured by this environmental image is the region extending outwards from the location of the receiver, with a set distance of 75 meters. This region is converted into a rectangular matrix C for processing by the dual-modal network, as described below:

[0039]

[0040] Where, the element c in the matrix gw The value of represents the height of the terrain or building at that point, and G×W represents the resolution of the environmental image;

[0041] In the second step, the path loss empirical model used is the path loss model in the 3GPP TR38.901UMa standardized channel model. This model predicts path loss based on the Alpha-Beta-Gamma (ABG) model. The specific method for predicting the coarse value of path loss is as follows:

[0042] Step 2.1: Using the communication propagation state in the propagation feature set in the first step, the communication bimodal dataset is first divided into a LosS subset and an NLoS subset. When the communication propagation state is 1, the communication bimodal dataset is assigned to the LosS subset. When the communication propagation state is 0, the communication bimodal dataset is assigned to the NLoS subset. The coarse path loss values ​​of the LosS subset and the NLoS subset are predicted respectively.

[0043] Step 2.2, for the data in the Loss subset, calculate the coarse path loss value based on the propagation distance d, the transmitter height, and the receiver height using the following formula:

[0044]

[0045] in, This is a coarse value for path loss at line-of-sight (LoS), in dB; d2d is the two-dimensional distance between the transmitter and receiver; f is the communication frequency used by the transmitter and receiver; d BP This is the breakpoint distance. Near this distance, the path loss may vary significantly due to the effects of electromagnetic propagation, and it is calculated using the following formula:

[0046]

[0047] Where c represents the speed of light;

[0048] Step 2.3: For the data in the NLoS subset, based on the characteristic propagation distance d, transmitter height, and receiver height extracted in the propagation feature set in the first step, calculate the coarse path loss value using the following formula:

[0049]

[0050] in, This is a coarse value of path loss in NLoS state;

[0051] In the third step, the dual-modal network consists of a feature extractor, a visual extractor, and a fusion processor. The feature extractor is an artificial neural network composed of multiple fully connected layers, and its input is the propagation feature set. The visual extractor is a convolutional neural network, and its input is the environmental image. The fusion processor is also an artificial neural network composed of multiple fully connected layers. The fusion processor concatenates the high-dimensional features extracted by the feature extractor and the visual extractor, and fits the predicted path loss correction value through the neural network. The overall process is as follows:

[0052] PL r =ρ(v(I, w) v , ε v )⊙δ(F,w δ , ε δ ), w ρ , ε ρ )

[0053] Among them, PL r ρ represents the path loss correction value predicted by the bimodal network, and ρ represents the fusion processor. ρ , ε ρ ) represents the network weights and hyperparameters of the fusion processor; (w v , ε v ) represents the network weights and hyperparameters of the visual processor; (w δ , ε δ ) represents the network weights and hyperparameters of the feature extractor;

[0054] The parameter optimization process of the bimodal network is based on the backpropagation principle, and the root mean square error (RMSE) is chosen as the loss function:

[0055]

[0056] Where m represents the total number of samples in the dataset. This represents the refined path loss prediction value for the i-th receiver. PL is obtained by adding the coarse value of the predicted path loss to the corrected value of the path loss. i Let ||·||2 represent the predicted actual path loss of the i-th receiver, and let ||·||2 represent the L2 norm.

[0057] Beneficial effects

[0058] (1) The present invention provides a refined path loss prediction method based on dual-modal network correction. An empirical path loss model characterizes the commonalities of path loss, and then the network is used to correct the deviation between the empirical path loss value and the actual path loss value, resulting in a refined path loss prediction value. This method integrates the learning ability of artificial intelligence with the expert knowledge of wireless communication.

[0059] (2) In the method of the present invention, a two-stage feature extraction algorithm is designed, which can automatically extract high-quality and representative features, avoiding the time and financial resources spent on manual feature extraction;

[0060] (3) In the method of the present invention, a dual-modal network is designed, which can summarize the functional relationship between the visual and feature modes and path loss, and learn the correction value of empirical path loss.

[0061] (4) Traditional path loss empirical models often use statistical expressions to predict path loss. These models are based on a series of assumptions, such as that the samples are independent of each other and follow a Gaussian distribution, but lack specific propagation details. Therefore, these methods can only give statistical values ​​of path loss in a region, such as the mean and variance of loss in a certain scenario. The dual-modal network proposed in this invention actually learns electromagnetic propagation characteristics from inputs that represent detailed environments based on the known empirical prediction values, thereby correcting the deviation of the empirical model prediction values ​​and obtaining refined path loss prediction values. This method realizes an effective combination of artificial intelligence methods and prior knowledge of experts in the field of communication.

[0062] (5) In machine learning, the prediction accuracy of the model often depends on a large number of high-quality data samples. The feature construction of these data samples is often time-consuming and laborious. In previous channel modeling work, it was often necessary to manually label features, which is a time-consuming and laborious process. The two-stage feature extraction method proposed in this invention can automatically obtain the propagation feature set required by the dual-modal network based on geometric information.

[0063] (6) The dual-modal network designed in this invention can mine electromagnetic propagation characteristics from data in two modes: features and images. Features mainly describe communication propagation information from a mathematical and statistical perspective, while images directly and intuitively express the detailed propagation environment near the receiving end in a visual form. The dual-modal network learns from the data in both modes to achieve accurate correction of the empirical model. Attached Figure Description

[0064] Figure 1 This is a flowchart of a refined path loss prediction method for dual-modal network correction according to the present invention.

[0065] Figure 2 The case study is the urban communication environment on the campus of Beijing Institute of Technology;

[0066] Figure 3 The proposed environmental image;

[0067] Figure 4 The prediction performance of the path loss method modified by the bimodal network is shown when the prediction parameter is path loss. Detailed Implementation

[0068] The refined path loss prediction method for dual-modal network correction according to the present invention will be described in detail below with reference to the accompanying drawings and embodiments:

[0069] Example

[0070] like Figure 1 As shown, a refined path loss prediction method based on dual-modal network correction is proposed. The steps of this method include:

[0071] The first step is to set the coordinates of the transmitter as (0m, 0m, 41m) and the coordinates of the receiver as follows: There is only one building in the communication environment, and the vector map information of the building is B. 1 ={H 1 K 1 , σ 1 The height H of the building 1 The length is 35m, and the number of building boundary lines is K. 1 There are 4 lines, and the building's boundary lines are... First boundary line The coordinates are ((30, 15), (45, 15)), the second boundary line The coordinates are ((45, 15), (45, 0)), the third boundary line The coordinates are ((45, 0), (30, 0)), the fourth boundary line The coordinates are ((30, 0), (30, 15)).

[0072] Step 1.1: Calculate the two-dimensional distance d between the transmitter and receiver based on their coordinates. 2D The distance is 30 meters, and the azimuth angle θ between the transmitter and the i-th receiver is 53°.

[0073] Step 1.2: Determine whether the line connecting the transmitter and receiver intersects the building in a two-dimensional plane. This determination is made using a counter-clockwise direction test method based on computational geometry (CCW); determine the endpoint set (a 1，k b 1，k p) and endpoint set (a 1，k b 1，k Whether (o, p, a) are arranged in the reverse rotation order, and the endpoint set (o, p, a) 1，k ) and endpoint set (o, p i b 1 ，k Whether they are arranged in the reverse rotation order.

[0074] Step 1.4 repeats the judgment steps of Step 1.2 for all boundary lines of the buildings, and finds that the boundary lines of the buildings and the lines connecting the transmitter and receiver do not intersect in the two-dimensional plane. The average height of the spaced buildings is 0, the number of spaced buildings is 0, and the communication propagation state is LoS.

[0075] Step 1.5: Concatenate all the acquired features to obtain the propagation feature set of all receivers.

[0076] In the first step, an environmental image is captured. The captured area is a region extending outwards from the location of the receiving end, with a set extension distance of 75 meters. This region is then converted into an environmental image with a resolution of 128*128.

[0077] The second step is to calculate the coarse path loss PL between the transmitter and receiver using an empirical path loss model. e The input to the path loss empirical model is the receiver height, propagation distance, and communication propagation state obtained in the propagation feature set in the first step. Since the communication propagation state is LoS, the path loss value PL is calculated using the path loss formula under line-of-sight conditions. e

[0078] The third step involves using a trained bimodal network to predict the path loss correction value PL between the transmitter and receiver. γ The input to the bimodal network consists of the environmental image obtained in the first step, as well as the receiver's horizontal coordinates, vertical coordinates, azimuth angle, number of spaced buildings, average height of spaced buildings, and communication propagation status. The path loss correction value PL is obtained through network prediction. γ .

[0079] The fourth step is to convert the coarse path loss value PL obtained in the second step into a value that is not directly related to the previous step. e The path loss correction value PL obtained in step 3 γ The sums are then used to obtain the fine-grained path loss prediction value PL. eh This enables refined path loss prediction for dual-modal network correction.

[0080] Path loss parameters were generated using WirelessInsite software simulation, with the simulation area being the Zhongguancun campus of Beijing Institute of Technology. The simulation area was 1000m × 800m. The transmitter was fixed at the center, 40m above the ground. All receivers used omnidirectional antennas for transmission and reception, distributed along the road. Each receiver was 1.5 meters high, resulting in a total of 8231 different receiver locations. Path loss parameters were calculated at each location, and the data was divided into training, test, and validation sets in a 6:2:2 ratio. Figure 2 As shown.

[0081] Step 1: Based on the two-stage feature extraction algorithm described in Step 1, extract the propagation feature set for each receiving end. The specific structure is as follows:

[0082] (1) Horizontal coordinates of the receiving end The horizontal coordinate of the i-th receiver with the transmitter as the origin.

[0083] (2) Vertical coordinates of the receiving end The vertical coordinate of the i-th receiver with the transmitter as the origin.

[0084] (3) Azimuth (θ) i ): The angle between the horizontal transmitting end and the i-th receiving end.

[0085] (4) Two-dimensional distance The two-dimensional distance from the transmitter to the i-th receiver.

[0086] (5) Number of interval buildings (n) i ): The number of buildings that the line connecting the transmitter to the i-th receiver passes through.

[0087] (6) Average height of buildings at intervals (meters): Represents the average height of buildings that pass through the line connecting the transmitter to the z-th receiver.

[0088] (7) Communication propagation status (L) i ): The propagation condition between the transmitter and the i-th receiver. When L is 1, it is line-of-sight propagation; when L is 0, it is non-line-of-sight propagation.

[0089] Step two: Construct an environmental image based on the simulation environment and the coordinates of the receiving end, such as... Figure 3 As shown.

[0090] Step 3: Using the propagation feature set provided in Step 1, divide the dataset according to the propagation status of LoS and NLoS, and predict the coarse value of path loss using the empirical model of LoS path loss and the empirical model of NLoS path loss respectively.

[0091] Step 4: Predict path loss correction values ​​using a bimodal network. After hyperparameter optimization, the feature extractor in the bimodal network has three neural network layers with 6, 16, and 128 neurons, and the activation function is ReLU. Specifically, the bimodal network structure, parameters, training, testing, and validation are all performed using the Python-based PyTorch architecture. During network training, the samples in the dataset are used as features extracted by the feature extractor with dimensions [1, 128]. The visual extractor uses a ResNet50 network structure. After extraction via a convolutional neural network, the image feature dimensions are [1, 128]. The fusion processor has five neural network layers with 256, 64, 16, 4, and 1 neurons, and the activation function is ReLU. The predicted path loss value after correction by the bimodal network is shown below. Figure 4 As shown.

[0092] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A refined path loss prediction method based on dual-modal network correction, characterized in that... The steps of this method include: The first step is to extract a bimodal communication dataset based on the coordinate information of the transmitter, receiver, and buildings in the environment. The bimodal communication dataset includes a propagation feature set and an environmental image. The propagation feature set includes the receiver's horizontal coordinates, vertical coordinates, height, azimuth, propagation distance, number of spaced buildings, average height of spaced buildings, and communication propagation status. The initial values ​​for the number of spaced buildings and the average height of spaced buildings are both 0, while the initial value for the communication propagation status is 1, indicating line-of-sight propagation. The propagation feature set is obtained through a two-stage feature extraction algorithm, specifically: Step 1.1: Given one transmitter, N receivers, and M buildings, calculate the two-dimensional distance between the transmitter and the i-th receiver based on their coordinates. And the azimuth angle θ between the transmitter and the i-th receiver i Let i = 1, ..., N, and the azimuth angle be the angle between the line connecting the transmitter and the i-th receiver and the horizontal plane. The position of the transmitter is considered as the origin o of a two-dimensional coordinate system, and the height of the transmitter is h. Tx The three-dimensional coordinates of the i-th receiver are represented as follows: P i Let be the two-dimensional coordinates of the i-th receiver. It is the horizontal coordinate of the i-th receiver. h is the vertical coordinate of the i-th receiver with the transmitter as the origin. i It is the height of the i-th receiver; Step 1.2: Determine whether the line connecting the transmitter and the i-th receiver intersects the j-th building in a two-dimensional plane, i.e., whether the line L1 connecting the coordinates of the transmitter and the i-th receiver intersects the boundary line L2 of the j-th building in a two-dimensional plane, j = 1, ..., M; Vector map information B of building j j As described in the following formula: Among them, H j K is the height of the j-th building. j σ is the number of boundary lines of the j-th building. j Let be the set of boundary lines of the j-th building, as shown in the following formula: This is the k-th boundary line of the j-th building, and its two endpoints are represented by the following formula: Determine the connection between the transmitter and the i-th receiver in a two-dimensional plane. Is it related to the k-th boundary line of the j-th building? The formula for calculating the intersection is: Where |·| represents determinant operation, when When the calculation result is positive, the endpoint set Arranged in counter-clockwise order, the endpoint set is used when the calculation result is negative. Arranged in clockwise order, when the calculation result is zero, it represents P. i On the k-th boundary line of the j-th building or on the extension of that boundary line; When endpoint set and endpoint set Arranged in the reverse rotation order, that is And endpoint set and endpoint set Arranged in the reverse rotation order, that is If both are true, then and An intersection occurs; proceed to step 1.

3. At this time, endpoint a j，k and b j，k lie in At both ends, o and p i lie in The line connecting the transmitting end and the i-th receiving end intersects with the j-th building in a two-dimensional plane, and the coordinates of the two-dimensional intersection point are recorded as (x0, y0). If any one of the conditions is not met, then and If no intersection occurs, update the boundary line k of the j-th building. If an intersection occurs, proceed to step 1.

3. If the line connecting the transmitter and the i-th receiver does not intersect the boundary line of the j-th building in the two-dimensional plane, then there is no three-dimensional intersection, and proceed to step 1.

4. Step 1.3: Determine whether the line connecting the transmitter and the i-th receiver intersects the j-th building in the three-dimensional plane, and calculate the height h of the three-dimensional intersection point. I As described in the following formula: If h I The height H of the j-th building is less than j If the line connecting the transmitter and the i-th receiver intersects the j-th building in the three-dimensional plane, it indicates that the transmitter's signal is blocked by the j-th building. Therefore, the intersection check for the j-th building is terminated, and the number of spaced buildings and the average height of the spaced buildings are updated (i.e., the number of spaced buildings increases by 1, and the average height of the spaced buildings increases by H). j Communication propagation state L i =0, proceed to step 1.4; Step 1.4: Update building j. Repeat steps 1.2 and 1.3 above until all buildings j = 1, ..., M are traversed. Finally, divide the updated average height of the interval buildings by the number of updated interval buildings to obtain the final average height of the interval buildings. Step 1.5: Repeat steps 1.1 and 1.4 above, traverse all receivers i = 1, ..., N, and concatenate all the acquired features to obtain the propagation feature set of all receivers; The second step is to calculate the coarse path loss between the transmitter and receiver using the propagation feature set obtained in the first step and the path loss empirical model. The third step involves using the propagation feature set and environmental images obtained in the first step to predict the path loss correction value between the transmitter and receiver using a trained bimodal network. The fourth step is to add the coarse path loss value obtained in the second step to the path loss correction value obtained in the third step to obtain the fine path loss prediction value, thus completing the fine path loss prediction of the dual-modal network correction.

2. The refined path loss prediction method for dual-modal network correction according to claim 1, characterized in that: In step 1.2, when performing the intersection judgment, a counterclockwise direction test method based on computational geometry is adopted.

3. The refined path loss prediction method for dual-modal network correction according to claim 2, characterized in that: In the second step, the input to the path loss empirical model is the transmitter height, receiver height, propagation distance, and communication propagation status obtained in the propagation feature set in the first step.

4. The refined path loss prediction method for dual-modal network correction according to claim 3, characterized in that: In the third step, the input to the bimodal network is the environmental image obtained in the first step, as well as the receiver's horizontal coordinates, vertical coordinates, azimuth angle, number of spaced buildings, average height of spaced buildings, and communication propagation status.

5. The refined path loss prediction method for dual-modal network correction according to claim 4, characterized in that: In the first step, the environmental image represents the first... The terrain information surrounding the receiver is captured in the environmental image, which is a region extending outwards from the receiver's location. This region is then converted into a rectangular matrix for processing by the dual-modal network. : Among them, the elements in the matrix The value represents the height of the terrain or building at that point. G × W This indicates the resolution of the environmental image.

6. The refined path loss prediction method for dual-modal network correction according to claim 5, characterized in that: In the second step, the path loss empirical model used is the path loss model in the 3GPP TR 38.901 UMA standardized channel model. This model predicts path loss based on the Alpha-Beta-Gamma (ABG) model. The specific method for predicting the coarse value of path loss is as follows: Step 2.1: Using the communication propagation state in the propagation feature set in the first step, the communication bimodal dataset is first divided into a LosS subset and an NLoS subset. When the communication propagation state is 1, the communication bimodal dataset is assigned to the LosS subset. When the communication propagation state is 0, the communication bimodal dataset is assigned to the NLoS subset. The coarse path loss values ​​of the LosS subset and the NLoS subset are predicted respectively. Step 2.2, for the data in the Loss subset, based on the propagation distance The transmitter and receiver heights are used to calculate the coarse path loss using the following formula: in, This is a coarse value of path loss under line-of-sight conditions, in dB. It is the two-dimensional distance between the transmitter and receiver; It refers to the communication frequency used by the transmitting and receiving ends; It is the distance from the breakpoint: in, Represents the speed of light; For the data in the NLoS subset, the characteristic propagation distance is determined based on the propagation feature set extracted in the first step. The path loss is roughly calculated using the following formula: (The formula is missing from the original text.) in, This is a coarse value of path loss in NLoS state.

7. The refined path loss prediction method for dual-modal network correction according to claim 6, characterized in that: In the third step, the dual-modal network consists of a feature extractor, a visual extractor, and a fusion processor. The feature extractor is an artificial neural network composed of multiple fully connected layers, and its input is a propagated feature set. The visual extractor is a convolutional neural network, and its input is an environmental image. The fusion processor is an artificial neural network composed of multiple fully connected layers. The fusion processor concatenates the high-dimensional features extracted by the feature extractor and the visual extractor, and fits the predicted path loss correction value through the neural network. in, This represents the path loss correction value predicted by the bimodal network. Represents a fusion processor, Network weights and hyperparameters representing the fusion processor; Network weights and hyperparameters representing the visual processor; These represent the network weights and hyperparameters of the feature extractor.

8. The refined path loss prediction method for dual-modal network correction according to claim 7, characterized in that: The parameter optimization process of the bimodal network is based on backpropagation, and the root mean square error is chosen as the loss function: in, Represents the total number of samples in the dataset. Representing the Fine-grained path loss prediction for each receiver. It is obtained by adding the predicted coarse path loss value to the path loss correction value. Representing the The actual path loss prediction value for each receiver. This represents the L2 norm.

Images