Tunnel environment channel modeling method based on ray tracing and ensemble learning

Abstract

The invention provides a tunnel environment channel modeling method based on ray tracing and integration, and the method comprises the steps: building a channel characteristic parameter data set of various communication configurations in a tunnel environment through employing a ray tracing method, carrying out the feature extraction of channel environment parameters through employing an integrated learning method based on the combination of MLP and XGBoost, and carrying out the MLP pre-training, and the XGBoost performs more accurate prediction by using the characteristics, so that a high-precision channel fast model in a tunnel environment is realized. According to the invention, an intelligent channel model can be provided, and the model has high precision on the premise of high model speed. The method provides a basis for tunnel communication system design, helps to optimize network architecture and parameter setting, and ensures communication quality and reliability.

Description

Technical Field

[0001] The present invention belongs to the field of channel modeling and machine learning, and in particular relates to a tunnel environment channel modeling method based on ray tracing and integrated learning. Background Art

[0002] In rail transit, the vehicle-ground communication network is the basis for multiple applications such as information transmission between trains and ground control systems, passenger services, and safety monitoring. Wireless channel modeling aims to evaluate the channel environment before establishing the vehicle-ground communication system, help improve the reliability and performance of the communication system, and establish a good communication environment. The tunnel environment is an important environment in rail transit, but since the tunnel in rail transit is a closed area, data collection needs to be done during non-operating hours and requires advance application, so it is difficult to organize measurement activities. The ray tracing method does not require on-site testing, and only needs to know and configure the environmental parameters to simulate the channel. Therefore, the ray tracing method has become a more suitable method for studying the channel characteristics in tunnel scenarios. Compared with other channel modeling methods, the ray tracing method has higher accuracy and can accurately reflect the channel characteristics in tunnel scenarios.

[0003] There are two main problems with the ray tracing method: first, the ray tracing method is slow to simulate, and if different antenna positions and different frequency bands are taken into account, it will take a considerable amount of time. Second, when the size of the tunnel changes, the original model cannot be used for the new tunnel.

[0004] Machine learning is an effective way to solve the above problems. Many studies have shown that machine learning can capture the correlation between data and the ability to learn nonlinear mapping. Channel modeling methods based on machine learning usually use multilayer perceptrons (MLP) for supervised learning, but traditional MLP training takes a long time and the accuracy of predicted channel characteristic parameters is low, because slight changes in environmental parameters often lead to huge changes in the multipath channel environment, which greatly affects the final fitting accuracy of MLP. Simply increasing the scale and depth of MLP will lead to overfitting of model training, so a new machine learning model is needed to improve the accuracy of channel modeling. Summary of the invention

[0005] In order to solve the technical problems existing in the prior art, the present invention provides a tunnel environment channel modeling method based on ray tracing and ensemble learning. Based on the ray tracing method, a tunnel environment channel characteristic parameter data set is established, and the pre-training results are spliced ​​with the features through the MLP pre-training model, and further training is performed through XGBoost. The final trained model can be used as the input environment feature to obtain an end-to-end channel model of the channel characteristic parameters.

[0006] Technical Solution A tunnel environment channel modeling method based on ray tracing and ensemble learning comprises the following steps: Step 1: Establish multiple 3D models of tunnel environments of different scales, set different center frequencies and transmitter and receiver positions, perform ray tracing simulation, and generate a channel characteristic parameter data set; Specifically, the channel characteristic parameter data set includes: feature ,in represents the tunnel bottom width, represents the tunnel center height, Indicates the horizontal coordinate of the tunnel section where the receiver is located, Indicates the vertical coordinate of the tunnel section where the receiver is located, Indicates the horizontal coordinate of the tunnel section where the transmitter is located, Indicates the vertical coordinate of the tunnel section where the transmitter is located, Indicates the distance between the transmitter and the receiver. Indicates the center frequency of the signal; Label ,in represents the received power, represents the RMS delay spread, represents the Rice K factor, represents the RMS horizontal azimuth angle of arrival, represents the root mean square angle of arrival at the zenith; Step 2: Pre-train the channel model based on MLP to obtain the pre-training results After that, Concatenate with the input X of MLP as the new input of XGBoost , and then further train the channel model based on XGBoost to get the final output .

[0007] Step 3: Apply the model to tunnel environment channel modeling. It supports directly generating five channel characteristic parameters, including received power, root mean square delay spread, Ricean K factor, root mean square horizontal azimuth angle of arrival, and root mean square zenith angle of arrival, by inputting eight parameters, including tunnel bottom width, tunnel height, horizontal coordinate of the tunnel section where the transmitter is located, vertical coordinate of the tunnel section where the transmitter is located, horizontal coordinate of the tunnel section where the receiver is located, vertical coordinate of the tunnel section where the receiver is located, distance between the transmitter and the receiver, and signal center frequency.

[0008] Compared with the prior art, the present invention has the following advantages: 1) The modeling method of ensemble learning combined with ray tracing proposed in the present invention has high accuracy and extremely fast modeling speed.

[0009] 2) The MLP-XGBoost ensemble learning method proposed in this invention has higher accuracy than that of MLP or XGBoost alone. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 A processing flow chart of a tunnel environment channel modeling method based on ray tracing and ensemble learning in the present invention; Figure 2 A schematic diagram of the positions of the transmitter and the receiver on the xz section of the tunnel according to an embodiment of the present invention; Figure 3 Schematic diagram of the MLP-XGBoost model structure according to an embodiment of the present invention; Figure 4 Comparison of the results of various learning models on the validation set in the embodiments of the present invention. DETAILED DESCRIPTION

[0011] The technical solution provided by the present application will be further described below in conjunction with specific embodiments and accompanying drawings. The advantages and features of the present application will become more apparent with the following description.

[0012] The present invention is mainly used for rapid modeling of vehicle-ground wireless channels in arched tunnel scenarios of rail transit. For scenarios where it is necessary to establish wireless channel characteristic parameter models such as path loss, root mean square delay spread, Ricean K factor, root mean square horizontal azimuth angle of arrival, and root mean square zenith angle of arrival in an arched tunnel environment, data with the same accuracy as the ray tracing method can be obtained, while avoiding the time overhead required for establishing a 3D model and ray tracing simulation.

[0013] like Figure 1 As shown, the present invention relates to a tunnel environment channel modeling method based on ray tracing and ensemble learning, and the channel modeling method comprises the following steps: Step 1: Establish multiple 3D models of tunnel environments of different scales, set different center frequencies and transmitter and receiver positions, perform ray tracing simulation, and generate a channel characteristic parameter data set. The specific steps include: Step 1.1: Create multiple 3D models of arched tunnels with different bottom widths and heights and set the materials of the tunnel walls, unify the reflection, scattering and diffraction modes, and set different center frequencies and transmitter coordinates. The coordinates of the receiver .like Figure 2 The tunnel cross section is plane, the tunnel extension direction is the y-axis, and the transmitter is located on a semicircle 10 cm away from the tunnel wall ( Figure 2 The blue dashed line) is the random position of the receiver. The cross section is located in the center of the tunnel ( Figure 2The receiver moves along the tunnel and gradually approaches the transmitter.

[0014] Step 1.2: After the parameter setting is completed, ray tracing simulation is performed to obtain the channel transfer function (ChannelTransfer Function, CTF), and the channel impulse response (CIR) can be written as: in, represents the channel impulse response, represents the inverse Fourier transform, Represents the channel transfer function. From the channel impulse response, we can get the two parameters of received power and root mean square delay spread, where the received power can be written as: in, represents the received power, represents the channel impulse response. The RMS delay spread can be written as: in, represents the RMS delay spread, Indicates delay, represents the average delay, Represents the power at a certain delay point, where the average delay It can be written as: in, It can be written as: In addition to CIR, ray tracing can also obtain the power, horizontal azimuth and zenith angle of each ray, so the three channel characteristic parameters of Rice K factor, root mean square horizontal azimuth and root mean square zenith angle can be obtained. The Rice K factor can be written as: in, represents the power of the direct path, represents the sum of the powers of other paths. The RMS horizontal angle spread can be written as: in, represents the power of the nth ray, represents the horizontal azimuth angle of arrival of the nth ray, represents the average horizontal azimuth of arrival, which can be written as: The root mean square arrival zenith angle can be written as: in, represents the horizontal azimuth angle of arrival of the nth ray, represents the average horizontal azimuth of arrival, which can be written as: This allows us to construct dataset labels .

[0015] Step 1.3: Generate the data set and features required for ensemble learning ,in represents the tunnel bottom width, represents the tunnel center height, Indicates the horizontal coordinate of the tunnel section where the receiver is located, Indicates the vertical coordinate of the tunnel section where the receiver is located, Indicates the horizontal coordinate of the tunnel section where the transmitter is located, Indicates the vertical coordinate of the tunnel section where the transmitter is located, Indicates the distance between the transmitter and the receiver. Indicates the center frequency of the signal. ,in represents the received power, represents the RMS delay spread, represents the Rice K factor, represents the RMS horizontal azimuth angle of arrival, represents the RMS zenith angle of arrival.

[0016] Step 2: If Figure 3 As shown, the channel model is pre-trained based on MLP to obtain the pre-training result After that, Concatenate with the input X of MLP as the new input of XGBoost , and then further train the channel model based on XGBoost to get the final output , the specific steps are as follows: Step 2.1: Read dataset features With label , divide the data set into training set and validation set, read the training set data and normalize the data.

[0017] Step 2.2: After normalizing the data, As input, using supervised learning method, the loss function is defined as mean square error: in, is the output dimension, is the training label, is the prediction result of MLP. represents the difference between the target value and the MLP predicted value in all dimensions, The smaller it is, the better the prediction effect is. The MLP structure is 6 layers, each layer has 512 neurons, the learning rate is set to 0.001, the optimizer uses the Adam optimizer, the number of iterations is 1000, and the pre-training results are obtained after the MLP training is completed. .

[0018] Step 2.3: and Splicing to get the input of the XGBoost model , configure XGBoost model hyperparameters, label , MSE is also used as the loss function, and its objective function can be written as: in Expressed as the objective function, is the mean square error, is a regularization term used to control the complexity of the model and prevent overfitting. The XGBoost training process is to continuously reduce its objective function. After the XGBoost training is completed, the prediction results can be obtained by inputting the environmental feature parameters. The final performance of the model is evaluated by RMSE and MAE.

[0019] MAE measures the average deviation between the predicted values ​​and the true values, while RMSE is used to measure the larger error present in the model.

[0020] Step 3: Apply the model to tunnel environment channel modeling. It supports directly generating five channel characteristic parameters, including received power, root mean square delay spread, Ricean K factor, root mean square horizontal azimuth angle of arrival, and root mean square zenith angle of arrival, by inputting eight parameters, including tunnel bottom width, tunnel height, horizontal coordinate of the tunnel section where the transmitter is located, vertical coordinate of the tunnel section where the transmitter is located, horizontal coordinate of the tunnel section where the receiver is located, vertical coordinate of the tunnel section where the receiver is located, distance between the transmitter and the receiver, and signal center frequency.

[0021] Figure 4The results of MLP, XGBoost and MLP-XGBoost ensemble learning models on the validation set are compared. Compared with using MLP or XGBoost alone, the model based on MLP-XGBoost has higher accuracy, and both RMSE and MAE have been improved to a certain extent. Compared with the ray tracing method, the execution time of the method of the present invention is shortened from 605 seconds to 3 seconds.

[0022] The above description is only a description of the preferred embodiments of the present application, and is not intended to limit the scope of the present application. Any changes or modifications made by any person skilled in the art based on the above disclosed technical contents shall be deemed as equivalent effective embodiments and shall fall within the scope of protection of the technical solution of the present application.

Claims

1. A tunnel environment channel modeling method based on ray tracing and ensemble learning, characterized in that: The following steps are involved: Step 1: Establish multiple 3D models of tunnel environments of different scales, set different center frequencies and transmitter and receiver positions, perform ray tracing simulation, and generate a channel characteristic parameter data set; Step 2: Pre-train the channel model based on MLP to obtain the pre-training results After that, Concatenate with the input X of MLP as the new input of XGBoost , and then further train the channel model based on XGBoost to get the final output ; Step 3: Apply the model to tunnel environment channel modeling. It supports directly generating five channel characteristic parameters, including received power, root mean square delay spread, Ricean K factor, root mean square horizontal azimuth angle of arrival, and root mean square zenith angle of arrival, by inputting eight parameters, including tunnel bottom width, tunnel height, horizontal coordinate of the tunnel section where the transmitter is located, vertical coordinate of the tunnel section where the transmitter is located, horizontal coordinate of the tunnel section where the receiver is located, vertical coordinate of the tunnel section where the receiver is located, distance between the transmitter and the receiver, and signal center frequency.

2. The method according to claim 1, characterized in that Step 1 includes the following steps: Step 1.1: Create multiple 3D models of arched tunnels with different bottom widths and heights and set the materials of the tunnel walls, unify the reflection, scattering and diffraction modes, set different center frequencies, transmitter coordinates and receiver coordinates. The tunnel section is perpendicular to the y-axis, the transmitter is located at a random position within a semicircle 10 cm away from the tunnel wall, and the receiver is located at a random position in the tunnel center. The receiver moves along the tunnel direction and gradually approaches the transmitter; Step 1.2: After the parameter setting is completed, ray tracing simulation is performed to obtain the channel transfer function and the arrival angle and power of each ray; the channel impulse response can be obtained through the channel transfer function, and the two channel characteristic parameters of received power and root mean square delay spread can be obtained through the channel impulse response; the Rice K factor is obtained through the power of each ray, and the root mean square arrival horizontal azimuth angle and root mean square arrival zenith angle are obtained through the incident angle and power of the ray; Step 1.3: Generate the data set and features required for ensemble learning ,in represents the tunnel bottom width, represents the tunnel center height, Indicates the horizontal coordinate of the tunnel section where the receiver is located, Indicates the vertical coordinate of the tunnel section where the receiver is located, Indicates the horizontal coordinate of the tunnel section where the transmitter is located, Indicates the vertical coordinate of the tunnel section where the transmitter is located, Indicates the distance between the transmitter and the receiver. Indicates the center frequency of the signal; label ,in represents the received power, represents the RMS delay spread, represents the Rice K factor, represents the RMS horizontal azimuth angle of arrival, represents the RMS zenith angle of arrival.

3. The method according to claim 1, characterized in that Step 2 includes the following steps: Step 2.1: Read dataset features With label , divide the data set into training set and validation set, read the training set data and normalize the data; Step 2.2: After normalizing the data, As input, using supervised learning method, the loss function is defined as mean square error: in, is the output dimension, is the training label, The prediction result of MLP. The pre-training result is obtained after MLP training. ; Step 2.3: and Splicing to get the input of the XGBoost model , configure XGBoost model hyperparameters, label , MSE is also used as the loss function, and its objective function can be written as: in Expressed as the objective function, is the mean square error, It is a regularization term used to control the complexity of the model and prevent overfitting. The XGBoost training process is to continuously reduce its objective function. After the XGBoost training is completed, the prediction result can be obtained by inputting the environmental feature parameters. ; The final performance of the model is evaluated by RMSE and MAE: Among them, MAE measures the average deviation between the predicted value and the true value, while RMSE is used to measure the larger error in the model.

Images