Deep learning-driven multi-base-station radio map rapid construction method and deep learning-driven multi-base-station radio map rapid construction system

Abstract

The invention discloses a deep learning-driven multi-base-station radio map rapid construction method and system. The method comprises the steps of generating a radio map original data set of a current environment; processing the radio map original data set to obtain a radio map sample data set; and constructing a neural network model, and training the neural network model according to the radio map sample data set to obtain a radio map construction model. According to the method, the learning ability and generalization performance of the model are enhanced through residual connection and batch normalization by adopting a double-path neural network architecture, so that comprehensive processing of signals of multiple base stations and rapid and high-precision radio map prediction are realized, and the requirements of real-time network planning and optimization are met. Furthermore, by processing the radio map original data set, each pixel data in the image includes path gain information, thereby providing an accurate signal propagation path and improving identification accuracy.

Description

Technical Field

[0001] The present invention belongs to the field of wireless network communications, and in particular relates to a deep learning-driven multi-base station radio map rapid construction method and system. Background Art

[0002] In modern wireless communication networks, accurate radio maps are essential for understanding signal propagation characteristics, which directly impacts multiple applications such as network planning, resource allocation, interference control, and positioning. Radio maps are usually expressed as a path loss function, which describes the signal attenuation from the transmit point (TX) to the receive point (RX). In an urban multi-base station environment, signal propagation is affected by buildings, street canyons, and objects of different heights, making accurate estimation of path loss more complicated.

[0003] Existing radio map construction methods mainly rely on the following techniques:

[0004] 1. Statistical models: These methods approximate path loss based on the distance attenuation function, but they usually cannot accurately capture the propagation characteristics in complex urban environments.

[0005] 2. Data-driven interpolation methods: such as Kriging interpolation and radial basis function (RBF) interpolation, which use measurement data at known locations to estimate path loss at unknown locations. Although these methods are simple, they often ignore the complex three-dimensional spatial propagation patterns in urban environments.

[0006] 3. Model-based data fitting methods: For example, tomography techniques estimate the path loss at unmeasured locations by deriving the spatial loss field (SLF) from the measured values. Although these methods take more physical phenomena into account, they usually require a lot of computing resources and have limited adaptability to dynamic environments.

[0007] 4. Model-based prediction methods: such as ray tracing and dominant path models, which are completely based on physical models to predict path loss. These methods can provide high-precision predictions in theory, but in practical applications they are often limited by computational complexity and uncertainty in environmental changes.

[0008] However, existing radio map construction methods often suffer from low computational efficiency, inability to update in real time, and insufficient prediction accuracy when dealing with multi-base station environments. Summary of the invention

[0009] In order to solve the above problems existing in the prior art, the present invention provides a method and system for quickly building a multi-base station radio map driven by deep learning. The technical problem to be solved by the present invention is achieved through the following technical solutions:

[0010] A deep learning driven method for fast construction of multi-base station radio maps, comprising:

[0011] Generate a raw dataset of radio maps of the current environment;

[0012] Processing the radio map original data set to obtain a radio map sample data set;

[0013] A neural network model is constructed, and the neural network model is trained according to the radio map sample data set to obtain a radio map construction model.

[0014] In a specific embodiment, processing the radio map raw data set includes: performing path loss estimation on the radio map raw data set generated by a ray tracing mode, so that each pixel data in the image includes path gain information.

[0015] In a specific embodiment, processing the radio map original data set further includes: cropping the map data in the data set so that each image is divided into at least M×N blocks of segmented images, upsampling each segmented image so that each segmented image is restored to the size of the image in the original data set, wherein M and N represent the number of segmentations along the pixel length and pixel width directions, respectively.

[0016] In a specific embodiment, processing the radio map raw data set further includes:

[0017] The cropped images are rotated, and each rotated image is stored as a radio map sample dataset.

[0018] In a specific embodiment, the neural network model includes two layers of convolution modules, each layer of convolution modules includes a convolution unit, a normalization unit and an activation function unit connected in sequence, wherein a residual connection unit is included between the two layers of convolution modules, and the calculation formula of the residual connection is: H(x)=F(x)+Wx, wherein H(x) represents the expected output of the layer, F(x) represents the output of the layer after nonlinear transformation, and Wx represents the residual mapping.

[0019] In a specific implementation, when the neural network model is trained according to the radio map sample data set, the model is evaluated by a normalized mean square error function, and when the normalized mean square error meets a preset condition, the iteration is stopped to obtain a radio map construction model.

[0020] The present invention also provides a deep learning-driven multi-base station radio map rapid construction system, including:

[0021] A raw data set generation module, used to generate a raw data set of a radio map of the current environment;

[0022] A sample data set processing module, used for processing the radio map original data set to obtain a radio map sample data set;

[0023] The neural network model processing module is used to construct a neural network model and train the neural network model according to the radio map sample data set to obtain a radio map construction model.

[0024] In a specific embodiment, processing the radio map raw data set includes: performing path loss estimation on the radio map raw data set generated by a ray tracing mode, so that each pixel data in the image includes path gain information.

[0025] In a specific embodiment, processing the radio map original data set further includes: cropping the map data in the data set so that each image is divided into at least M×N blocks of segmented images, upsampling each segmented image so that each segmented image is restored to the size of the image in the original data set, wherein M and N represent the number of segmentations along the pixel length and pixel width directions, respectively.

[0026] In a specific embodiment, the neural network model includes two layers of convolution modules, each layer of convolution modules includes a convolution unit, a normalization unit and an activation function unit connected in sequence, wherein a residual connection unit is included between the two layers of convolution modules, and the calculation formula of the residual connection is: H(x)=F(x)+Wx, wherein H(x) represents the expected output of the layer, F(x) represents the output of the layer after nonlinear transformation, and Wx represents the residual mapping.

[0027] Beneficial effects of the present invention:

[0028] The deep learning-driven multi-base station radio map rapid construction method of the present invention adopts a dual-path neural network architecture to enhance the learning ability and generalization performance of the model through residual connection and batch normalization, thereby realizing the comprehensive processing of multi-base station signals and fast, high-precision radio map prediction, meeting the needs of real-time network planning and optimization. In addition, by processing the original data set of the radio map, each pixel data in the image includes path gain information, thereby providing an accurate signal propagation path and improving recognition accuracy.

[0029] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is a schematic flow chart of a method for rapidly constructing a multi-base station radio map driven by deep learning provided by an embodiment of the present invention;

[0031] Figure 2 This is a module block diagram of a deep learning-driven multi-base station radio map rapid construction system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0032] The present invention is further described in detail below with reference to specific embodiments, but the embodiments of the present invention are not limited thereto.

[0033] Embodiment 1

[0034] See also Figure 1 , Figure 1 The present invention provides a method for rapidly constructing a multi-base station radio map driven by deep learning, including:

[0035] S1. Generate a radio map raw data set of the current environment; the radio map raw data set of this embodiment takes into account the path propagation loss of the radio signal, so each pixel data in the image includes path gain information. Generally, path propagation loss processing can be performed by ray tracing software WinProp. In addition, the radio map raw data set also includes network environment parameters, such as the size of the area to be detected, pixel length, carrier frequency, bandwidth, transmission power, transmitter antenna type, and transmitter waveform.

[0036] S2. Process the radio map original data set to obtain a radio map sample data set.

[0037] It should be noted that in ray tracing simulations, due to free space power loss, the light from a given TX antenna gradually loses power as the distance increases. Further attenuation occurs when these rays intersect vegetation, and during reflection or transmission from buildings. Following these physical principles, ray tracing simulates the behavior of each multipath component (MPC) between TX and RX, including free space power loss and interaction with different interacting objects (IOs), so that the absolute amplitude |α|, departure Ω and arrival Ψ directions, and delay τ of each MPC can be calculated. The contribution of the mth MPC can be expressed as:

[0038] h m (τ,Ω,Ψ)=|α m |δ(τ-τ m )δ(Ω-Ω m )δ(Ψ-Ψ m ),

[0039] The total contribution of all MPCs is given by:

[0040]

[0041] The average path gain (the inverse of the path loss) for small-scale fading can be calculated as:

[0042]

[0043] The radio map uses path gain information (expressed in dB scale), while other angles and delay information are not used. Considering the actual geographical and morphological maps, the preferred simulation is performed at 2.5GHz to simulate the most widely used cellular band, but similar simulations can be performed in other bands with only slight parameter adjustments. The original dataset for the ray tracing simulation is preprocessed using interpolation and data augmentation methods, and then divided into training dataset, validation dataset and test dataset. The ray tracing simulations are performed on a set of discrete RX positions, and it is computationally challenging to perform these simulations for each position (pixel). Therefore, some pixels may lack channel information. In order to interpolate the missing data of the map pixels, the present invention utilizes bilinear interpolation to effectively approximate the values ​​between adjacent positions while maintaining computational efficiency.

[0044] In a specific embodiment, processing the radio map original data set further includes: cropping the map data in the data set so that each image is divided into at least M×N blocks of segmented images, upsampling each segmented image so that each segmented image is restored to the size of the image in the original data set, wherein M and N represent the number of segmentations along the pixel length and pixel width directions, respectively.

[0045] Specifically, the map with channel information (i.e., image data in the original dataset of the radio map) is converted into a binary grayscale 256×256 image. For each pixel, the received power PRX (in [dBm]) is converted to a value between 0 and 1 using Min-Max normalization. The minimum value is selected as -150dB because smaller path gain values ​​are irrelevant to system considerations. It should be noted that this embodiment sets the pixels where the building is located to 0, that is, indoor coverage is not considered. Then, the ground truth map becomes a channel 256×256 image. Similarly, the pixels located in the building are filled with 0, while the region of interest (RoI) is filled with values ​​between 0 and 1, which corresponds to PRX. In general, a larger dataset is required to achieve better neural network training performance. In this embodiment, two enhancement methods are mainly used: cropping and rotation. The entire map data is cropped into an image of approximately one-quarter size with TX as the anchor point, increasing the size of the dataset by 96 times. The image is first cropped to an image of 64×64 size and then upsampled to an image of 256×256 size. Some cropped images (those excluding TX) will be skipped because the TX position will be used as our second input channel. After cropping, the image set is rotated by 90°, 180°, and 270°, further increasing the size of the dataset by a factor of 4. Each rotated image is stored as the radio map sample dataset.

[0046] S3. Construct a neural network model, and train the neural network model according to the radio map sample data set to obtain a radio map construction model. An improved version of the UNet neural network. By adding residual connections and batch normalization after each convolutional layer, RED-UNet accelerates the convergence of the model, enhances the generalization ability of the model, promotes gradient flow, and alleviates the problem of gradient disappearance, thereby improving the accuracy of radio map prediction.

[0047] In traditional deep convolutional neural networks, as the network depth increases, gradient vanishing and gradient exploding problems often occur, which makes the network difficult to train and optimize. The residual connection of this embodiment alleviates these problems by adding direct connections between layers, thereby improving the training efficiency and performance of the network. In this embodiment, the form of the residual connection can be expressed as:

[0048] H(x)=F(x)+x

[0049] Where H(x) represents the expected output of the layer, F(x) represents the output of the layer after nonlinear transformation, and x represents the input signal. This form intuitively expresses the idea of ​​residual connection: by adding the input signal x to the output F(x) after nonlinear transformation, the expected output H(x) of the layer is obtained, thus establishing a direct jump connection in the network. In practice, in order to ensure effective learning in the network, a residual mapping Wx is usually introduced, and the calculation of the residual connection is further expressed as:

[0050] H(x)=F(x)+Wx

[0051] This design enables the network to learn the residual more effectively, thereby improving the generalization ability and training speed of the network.

[0052] Preferably, in the network architecture of the present embodiment, a batch normalization layer is inserted after each convolutional layer to normalize the input distribution of each convolutional layer, which helps to accelerate the convergence of the network. This normalization process reduces the impact of internal covariate offset, making the network more robust to parameter initialization, thereby accelerating convergence and improving the generalization ability of the model. In addition, the batch normalization layer helps to alleviate the problem of gradient vanishing or exploding by dynamically adjusting the gradient range of each layer during training, making the network easier to optimize. It stabilizes the training process of the network based on normalizing the mean and variance of each feature channel in each training batch, and reduces the sensitivity to parameter initialization and learning rate adjustment. Therefore, batch normalization not only improves the training efficiency of the network, but also improves its generalization ability, making the network more suitable for complex real-world tasks.

[0053] In a specific implementation, when the neural network model is trained according to the radio map sample data set, the model is evaluated by a normalized mean square error function, and when the normalized mean square error meets a preset condition, the iteration is stopped to obtain a radio map construction model.

[0054] Normalized mean square error (NMSE) emphasizes the degree of misclassification in the entire dataset. Lower NMSE values ​​indicate better performance, and vice versa. The expression of NMSE is as follows:

[0055]

[0056]

[0057] For the radio map prediction task, each pixel value (representing the power value at that specific coordinate) is the main indicator of accuracy. Therefore, the NMSE work considers the normalized value of each pixel to calculate the accuracy. Normalized Depth Loss (NDL) is a good metric for comparing the changes in deep structures in an image. NDL calculates the logarithm of the depth error, which is sensitive to the shift in the depth direction. Let The loss is calculated as follows:

[0058]

[0059] Among them, F(x)=ln(x+γ),γ≥0.

[0060] This function is used to analyze and evaluate the model to determine whether to train again or save the model.

[0061] The deep learning-driven multi-base station radio map rapid construction method of this embodiment adopts a dual-path neural network architecture to enhance the learning ability and generalization performance of the model through residual connection and batch normalization, thereby realizing the comprehensive processing of multi-base station signals and fast, high-precision radio map prediction, meeting the needs of real-time network planning and optimization. In addition, by processing the original data set of the radio map, each pixel data in the image includes path gain information, thereby providing an accurate signal propagation path and improving recognition accuracy.

[0062] The present invention also provides a deep learning driven multi-base station radio map rapid construction system, see Figure 2 ,include:

[0063] A raw data set generation module, used to generate a raw data set of a radio map of the current environment;

[0064] A sample data set processing module, used for processing the radio map original data set to obtain a radio map sample data set;

[0065] The neural network model processing module is used to construct a neural network model and train the neural network model according to the radio map sample data set to obtain a radio map construction model.

[0066] In a specific embodiment, processing the radio map raw data set includes: performing path loss estimation on the radio map raw data set generated by a ray tracing mode, so that each pixel data in the image includes path gain information.

[0067] In a specific embodiment, processing the radio map original data set further includes: cropping the map data in the data set so that each image is divided into at least M×N blocks of segmented images, upsampling each segmented image so that each segmented image is restored to the size of the image in the original data set, wherein M and N represent the number of segmentations along the pixel length and pixel width directions, respectively.

[0068] In a specific embodiment, the neural network model includes two layers of convolution modules, each layer of convolution modules includes a convolution unit, a normalization unit and an activation function unit connected in sequence, wherein a residual connection unit is included between the two layers of convolution modules, and the calculation formula of the residual connection is: H(x)=F(x)+Wx, wherein H(x) represents the expected output of the layer, F(x) represents the output of the layer after nonlinear transformation, and Wx represents the residual mapping.

[0069] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine different embodiments or examples described in this specification.

[0070] Although the present application is described herein in conjunction with various embodiments, in the process of implementing the claimed application, those skilled in the art may understand and implement other variations of the disclosed embodiments by viewing the drawings, the disclosure, and the appended claims. In the claims, the word "comprising" does not exclude other components or steps, and "one" or "an" does not exclude multiple situations. A single processor or other unit may implement several functions listed in a claim. Certain measures are recorded in different dependent claims, but this does not mean that these measures cannot be combined to produce good results.

[0071] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the protection scope of the present invention.

Claims

1. A deep learning driven method for fast construction of multi-base station radio maps, characterized in that: include: Generate a raw dataset of radio maps of the current environment; Processing the radio map original data set to obtain a radio map sample data set; A neural network model is constructed, and the neural network model is trained according to the radio map sample data set to obtain a radio map construction model.

2. The deep learning driven multi-base station radio map rapid construction method according to claim 1 is characterized in that: Processing the radio map raw data set includes: performing path loss estimation on the radio map raw data set generated by a ray tracing mode, so that each pixel data in the image includes path gain information.

3. The deep learning driven multi-base station radio map rapid construction method according to claim 1 is characterized in that: Processing the radio map original data set also includes: cropping the map data in the data set so that each image is divided into at least M×N block segmented images, upsampling each segmented image so that each segmented image is restored to the size of the image in the original data set, wherein M and N represent the number of segmentations along the pixel length and pixel width directions, respectively.

4. The deep learning driven multi-base station radio map rapid construction method according to claim 1, characterized in that: Processing the radio map raw data set further includes: The cropped images are rotated, and each rotated image is stored as a radio map sample dataset.

5. The deep learning driven multi-base station radio map rapid construction method according to claim 1, characterized in that: The neural network model includes two layers of convolution modules, each of which includes a convolution unit, a normalization unit and an activation function unit connected in sequence, wherein a residual connection unit is included between the two layers of convolution modules, and the calculation formula of the residual connection is: H(x)=F(x)+Wx, wherein H(x) represents the expected output of the layer, F(x) represents the output of the layer after nonlinear transformation, and Wx represents the residual mapping.

6. The deep learning driven multi-base station radio map rapid construction method according to claim 1, characterized in that: When the neural network model is trained according to the radio map sample data set, the model is evaluated by a normalized mean square error function, and when the normalized mean square error meets a preset condition, the iteration is stopped to obtain a radio map construction model.

7. A deep learning driven multi-base station radio map rapid construction system, characterized in that: include: A raw data set generation module, used to generate a raw data set of a radio map of the current environment; A sample data set processing module, used for processing the radio map original data set to obtain a radio map sample data set; The neural network model processing module is used to construct a neural network model and train the neural network model according to the radio map sample data set to obtain a radio map construction model.

8. The deep learning driven multi-base station radio map rapid construction system according to claim 7, characterized in that: Processing the radio map raw data set includes: performing path loss estimation on the radio map raw data set generated by a ray tracing mode, so that each pixel data in the image includes path gain information.

9. The deep learning driven multi-base station radio map rapid construction system according to claim 8, characterized in that: Processing the radio map original data set also includes: cropping the map data in the data set so that each image is divided into at least M×N block segmented images, upsampling each segmented image so that each segmented image is restored to the size of the image in the original data set, wherein M and N represent the number of segmentations along the pixel length and pixel width directions, respectively.

10. The deep learning driven multi-base station radio map rapid construction system according to claim 9, characterized in that: The neural network model includes two layers of convolution modules, each of which includes a convolution unit, a normalization unit and an activation function unit connected in sequence, wherein a residual connection unit is included between the two layers of convolution modules, and the calculation formula of the residual connection is: H(x)=F(x)+Wx, wherein H(x) represents the expected output of the layer, F(x) represents the output of the layer after nonlinear transformation, and Wx represents the residual mapping.

Images