Radio frequency map estimation method for industrial indoor scene
By combining ray tracing technology and CGAN to generate RF map samples and using ViT model to extract features, high-precision RF map estimation for industrial indoor scenarios is achieved, solving the shortcomings in accuracy and efficiency of existing methods.
Patent Information
- Application Number
- CN202411537810.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-31
- Publication Date
- 2025-02-11
AI Technical Summary
Existing deep learning-based radio frequency map estimation methods are difficult to effectively characterize complex environments and radio propagation phenomena in industrial indoor scenarios, resulting in insufficient estimation accuracy and computational efficiency.
A radio frequency map estimation method for industrial indoor scenarios is proposed. By combining ray tracing technology and conditional generation adversarial network (CGAN), a large number of new radio frequency map samples are generated, and the Vision Transformer (ViT) model is used to extract the most relevant path loss. Features, ultimately implement online rapid estimation of high-precision RF maps.
It realizes rapid and high-precision estimation of RF maps in industrial indoor scenarios, solves the shortcomings of traditional methods in environmental complexity and computing efficiency, and provides an effective data enhancement tool.
Smart Images
Figure CN119421171A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to industrial wireless network technology, and in particular to a radio frequency map estimation method for industrial indoor scenes. Background Art
[0002] Smart factories are highly dependent on industrial wireless networks that can guarantee ultra-high reliability and low-latency communications. The design of future industrial wireless networks urgently needs to obtain accurate information on spectrum activity and radio channels in the factory environment. Factory environments usually contain a lot of metal, large equipment, and rotating and moving equipment, resulting in extremely dynamic channel states. RF maps, as a visual description of the radio propagation environment, are essential for the planning and resource management of industrial wireless networks. Therefore, RME has become one of the most popular research directions in the field of industrial wireless networks.
[0003] Traditional RME methods either rely on empirical channel models, which have limited estimation accuracy, or rely on ray tracing simulations, which are too time-consuming to implement in highly dynamic wireless networks. In recent years, deep learning-based RME methods have performed well in terms of estimation accuracy and computational efficiency, and are widely believed to replace traditional RME methods. Existing deep learning-based RME methods are mainly aimed at commercial wireless applications, such as macro cells in urban or rural scenarios, indoor or outdoor WiFi scenarios, etc., and cannot fully characterize the important features in industrial indoor scenarios. This is because industrial indoor scenarios are significantly different from general scenarios in terms of environmental complexity and radio propagation phenomena. Specifically, industrial indoor scenes contain a large number of obstacles, including walls, windows, and static or moving machines (heavy machinery, rotating robotic arms, automatic guided vehicles, etc.), which lead to a large number of reflection, diffraction and scattering phenomena. To address the above challenges, the present invention proposes for the first time a radio frequency map estimation method for industrial indoor scenarios, which can quickly estimate the corresponding high-precision radio frequency map online based on a given workshop floor plan. Summary of the invention
[0004] This paper proposes a radio frequency map estimation method for industrial indoor scenes. Specifically, this method first uses ray tracing technology to construct a radio frequency map simulation data set to train a conditional generative adversarial network (CGAN) to efficiently generate a large number of new radio frequency map samples; then a RME method based on Vision Transformer (ViT) is proposed. This method uses a large number of radio frequency map samples generated by CGAN to train the ViT model, which can extract the most important features related to path loss. Finally, the corresponding high-precision radio frequency map can be quickly estimated online based on the input industrial indoor scene plan.
[0005] The technical solution adopted by the present invention to achieve the above-mentioned purpose is: a radio frequency map estimation method for industrial indoor scenes, comprising the following steps:
[0006] Construct a simulated RF map dataset based on ray tracing to train the CGAN model, and use the CGAN model to generate new RF map samples;
[0007] Adopting the ViT-based RF map estimation method, the simulated RF map dataset is expanded with new RF map samples generated by the CGAN model, and the expanded dataset is used to train the ViT model;
[0008] Given a floor plan of an industrial indoor scene, it is input into the trained ViT model to estimate the corresponding RF map in real time.
[0009] The simulated radio frequency map data set based on ray tracing includes N×M radio frequency maps of size H×W, and each radio frequency map stores a path loss value between the pixel position and the transmitter at each pixel;
[0010] N is the number of workshop floor plans, each workshop floor plan is configured with a corresponding simulated RF map, where each workshop floor plan shares the same M transmitter locations;
[0011] Divide each workshop plan into grid pixels according to a set resolution and convert it into a two-dimensional image of H×W pixels; the two-dimensional image contains workshop building materials with different electromagnetic properties;
[0012] In each simulation, for each workshop floor plan, only one of the M transmitters is in working state.
[0013] The method of using the CGAN model to generate a new radio frequency map sample includes the following steps:
[0014] 1) Each workshop plan is discretized into H×W grids, and each workshop i has a plane image with material information For the jth transmitter position, j = 1, 2, ···, M, represents the RF map generated according to the 3GPP InF model; when the jth transmitter in workshop i is working, scenario X ij Expressed as
[0015]
[0016] Among them, concatenate means concatenating single-channel images of the same size F i and T j X is connected in series along the channel dimension to form a dual channel ij ; Given X ij , Yij Realistic RF maps generated for ray tracing simulations; [X ij ,Y ij ] constitutes a labeled training sample; based on the labeled training sample [X ij ,Y ij ]Learn the network model f θ (·), to estimate the entire RF map Right now
[0017]
[0018] Among them, θ represents the set of network parameters;
[0019] 2) Construct a CGAN model consisting of a generator and a discriminator, and use the CGAN generator to represent f θ (·); where the conditional information input into the CGAN model is used to represent the position of the working transmitter in the workshop The location information of the transmitter in the scene is encoded using one-hot encoding. After training the CGAN model using a simulated RF map dataset, a RF map is generated using the CGAN model.
[0020] The ViT-based radio frequency map estimation method uses new radio frequency map samples generated by the CGAN model to expand the simulated radio frequency map dataset, and uses the expanded dataset to train the ViT model, including the following steps:
[0021] (1) Image serialization: In the ViT model, we first convert the input image of size 2×H×W into Crop to image blocks of size 2×P×P, and expand each image block into a one-dimensional vector, and finally obtain
[0022] (2) Image block embedding and position embedding: Use the linear projection operator E to embed the image block sequence X p Mapped to the D-dimensional feature space, the image block embedding and position embedding together form the D-dimensional z0 as the input of the ViT encoder:
[0023]
[0024] in, To achieve the linear projection matrix, represents position embedding, Represents the Nth image block sequence; the ViT model uses sine and cosine functions of different frequencies to achieve position embedding:
[0025]
[0026]
[0027] Where pos represents the position of the image block in the image, i∈[0,1,···,D / 2] is the index used to calculate the channel dimension, and D represents the dimension of the feature space; PE (pos,2i) Represents the position embedding at the 2i-th dimension pos; the position embedding needs to have the same dimension as the image block embedding;
[0028] (3) Constructing the ViT encoder: The ViT encoder is composed of L identical encoding layers stacked together. Each encoding layer includes a multi-head self-attention sublayer and a multi-layer perceptron sublayer. Before the data enters each sublayer, it is normalized using layer normalization. After passing through each sublayer, the data is directly fused with the input using residual connections.
[0029] In order to achieve residual connection, the output dimension of each layer of the ViT encoder is designed to be D-dimensional; the calculation process of the lth layer is as follows:
[0030] z l ′ =MSA(LN(z l-1 ))+z l-1 ,l=1,2,···,L
[0031] z l =MLP(LN(z l ′ ))+z l ′ ,l=1,2,···,L
[0032] in, represents the feature vector obtained after L encoding layers; MSA and MLP represent multi-head self-attention mechanism and multi-layer perceptron respectively, and LN represents layer normalization;
[0033] (4) RF map output: Get z L After that, reshape it into two dimensions with size The feature map of D = P 2 , the feature map P 2 The channels are rearranged to the same dimension to obtain a grayscale RF map of size 1×H×W.
[0034] A radio frequency map estimation method for industrial indoor scenes, comprising:
[0035] The sample construction module is used to construct a simulated RF map dataset based on ray tracing to train the CGAN model and use the CGAN model to generate new RF map samples;
[0036] A model building module for adopting a ViT-based RF map estimation method, augmenting a simulated RF map dataset with new RF map samples generated by a CGAN model, and training a ViT model using the augmented dataset;
[0037] The RF map estimation module is used to input the trained ViT model based on the given industrial indoor scene floor plan and estimate the corresponding RF map in real time.
[0038] A radio frequency map estimation device for industrial indoor scenes comprises a memory and a processor; the memory is used to store a computer program; the processor is used to implement a radio frequency map estimation method for industrial indoor scenes when executing the computer program.
[0039] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, a radio frequency map estimation method for industrial indoor scenes is implemented.
[0040] The present invention has the following beneficial effects and advantages:
[0041] 1. For the first time, this paper designs and trains a CGAN model that can efficiently generate industrial RF maps based on a ray tracing simulation dataset, providing a powerful data enhancement tool for the data-scarce RME problem;
[0042] 2. The present invention proposes a ViT-based RME method, which can quickly estimate the corresponding high-precision radio frequency map online according to the input industrial indoor scene plan. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] Figure 1 This is a schematic diagram of the generator network of CGAN;
[0044] Figure 2 Schematic diagram of the discriminator network of CGAN;
[0045] Figure 3 This is a schematic diagram of the overall architecture of CGAN;
[0046] Figure 4 Flowchart of the ViT-based RME method. DETAILED DESCRIPTION
[0047] In order to make the above-mentioned objectives, technical solutions and advantages of the present invention more clear, the specific implementation modes of the present invention will be further described in detail below with reference to the accompanying drawings.
[0048] A radio frequency map estimation method for industrial indoor scenes includes the following steps:
[0049] (1) Using ray tracing technology to construct a radio frequency map simulation dataset to train CGAN and efficiently generate a large number of new radio frequency map samples;
[0050] (2) A ViT-based RME method is proposed. A large number of RF map samples generated by CGAN are used to train the ViT model to extract the important features most relevant to path loss. Finally, the corresponding high-precision RF map can be quickly estimated online based on the input industrial indoor scene floor plan.
[0051] In the step (1), the simulated radio frequency map data set based on ray tracing includes N workshop floor plans (each plan share the same M transmitter positions) and corresponding simulated radio frequency maps. In order to perform ray tracing simulation, the present invention divides each workshop floor plan into grid pixels at a resolution of 0.1 meters, and converts it into a two-dimensional image of H×W pixels, which contains various workshop building materials (such as concrete, bricks, glass and metal) with different electromagnetic properties. In each simulation, only one of the M transmitters is in working state, so the interference between transmitters is not considered. Therefore, the data set includes a total of N×M radio frequency maps of size H×W. Each radio frequency map stores the path loss (PL) value between the pixel position and the transmitter at each pixel.
[0052] In step (1), a new radio frequency map sample is generated using the CGAN model, as follows:
[0053] (1.1) Problem Modeling
[0054] The RME problem of industrial indoor scenes (such as workshops) can be considered as an image-to-image learning task. In the context of the present invention, RME aims to estimate the PL value in the region of interest using the workshop floor plan, transmitter locations, and the 3GPP Indoor Factory (InF) model. In the workshop scenario considered in the present invention, each workshop floor plan is discretized into H×W grids. All workshop floor plans share the same m transmitter locations, and the locations of the transmitters are known.
[0055] Each workshop i (i=1,2,···,N) has its own plane image with material information For the jth transmitter position (j=1,2,···,M), represents the RF map generated according to the 3GPP InF model. Therefore, when the jth transmitter in workshop i is working, scenario X ij It can be expressed as
[0056]
[0057] Where concatenate means concatenating single-channel images of the same size i and T j X is connected in series along the channel dimension to form a dual channel ij Given X ij , remember Y ij Realistic RF maps generated for ray tracing simulations. Therefore, [X ij ,Y ij ] constitutes a labeled training sample. The goal of the RME problem is to ij ,Y ij ]Learn a network model f θ (·), to estimate the entire RF map Right now
[0058]
[0059] Where θ represents the set of network parameters.
[0060] (1.2) Train CGAN to generate new RF map samples
[0061] The present invention designs a CGAN model consisting of a generator and a discriminator, wherein the generator is a fully convolutional network based on the UNet architecture, and the discriminator is a binary convolutional neural network whose final output is 0 (false) or 1 (true). The present invention uses the CGAN generator to approximate f in the RME problem (9) θ (·), where the conditional information of CGAN is input The location information of the transmitter in the scene is encoded by one-hot encoding. After training the CGAN model using the RF map dataset obtained in step (1.1), CGAN can be used as a data enhancement tool to output a large number of generated RF maps, significantly increasing the size of the existing simulation dataset and requiring less computing resources than ray tracing simulation.
[0062] The RME method based on the ViT model in step 2) specifically includes the steps of image serialization, image block embedding, position embedding, ViT encoder encoding and radio frequency map output.
[0063] In this embodiment, the present invention mainly includes two parts. One is to train the CGAN model based on ray tracing simulation data to generate a radio frequency map, and the other is to train the ViT model based on the generated radio frequency map to realize RME.
[0064] 1. Train the CGAN model based on ray tracing simulation data to generate RF maps
[0065] like Figure 1As shown, the present invention designs a two-channel input X for the generator of the CGAN ij The first channel (scene channel F i ) is a workshop plan with material information marked; at the same time, the second channel (model channel T j ) is a radio frequency map generated according to the 3GPPInF model. Figure 1 The generator network based on the UNet architecture designed in this patent is demonstrated, including an input layer, an encoder, a decoder, multiple skip connections and an output layer.
[0066] like Figure 2 As shown, the discriminator of the CGAN described in the present invention is a four-layer convolutional neural network, which accepts the generated radio frequency map and condition information as input, and outputs the probability value that the generated radio frequency map is the real radio frequency map.
[0067] like Figure 3 As shown in Figure 1, the training of the entire CGAN model is a process in which the generator and the discriminator interact in a zero-sum game, where the generator aims to generate new RF map samples similar to the simulated dataset, while the discriminator focuses on identifying real and generated map samples. The present invention uses the generator of CGAN to approximate f in the RME problem (9) θ (·), where the conditional information of CGAN is input The location information of the transmitter in the scene is encoded by one-hot encoding. After training the CGAN model using the RF map dataset obtained in step (1), CGAN can be used as a data enhancement tool to output a large number of generated RF maps, significantly increasing the size of the existing simulation dataset and requiring less computing resources compared to ray tracing simulation.
[0068] 2. Train the ViT model based on the generated RF map to achieve RME
[0069] like Figure 4 As shown, the ViT-based RME method designed by the present invention is based on the scene channel F i and model channel T j As input, the estimated RF map is output. Specifically, it includes the steps of image serialization, image block embedding, position embedding, ViT encoder encoding and RF map output:
[0070] (2.1) Image serialization
[0071] In the ViT model, in order to process a two-dimensional image, the input image of size 2×H×W is first transformed into Crop to image blocks of size 2×P×P, and expand each image block into a one-dimensional vector, and finally obtain
[0072] (2.2) Image Patch Embedding and Position Embedding
[0073] Then, we use a trainable linear projection operator E to transform the image patch sequence X p Mapped into a potential D-dimensional feature space. In order to retain the spatial position information of the image block on the image, it is necessary to learn a specific position embedding to encode the position information. The image block embedding and the position embedding together form the D-dimensional z0 as the input of the ViT encoder:
[0074]
[0075] in is the matrix that implements linear projection, Represents Position Embedding (PE). ViT uses sine and cosine functions of different frequencies to implement position embedding:
[0076]
[0077] Where pos is the position of the image block in the image, and i∈[0,1,···,D / 2] is the index used to calculate the channel dimension. The position embedding needs to have the same dimension (D dimension) as the image block embedding so that the two can be added directly.
[0078] (2.3) ViT encoder
[0079] The encoder of ViT is composed of L identical encoding layers stacked together, each of which includes a multi-head self-attention (MSA) sublayer and a multilayer perceptron (MLP) sublayer. Before the data enters each sublayer, it is normalized using layer normalization (LN). After the data passes through each sublayer, it is directly fused with the input using residual connections. In order to achieve residual connections, the output dimension of each layer of the ViT encoder is designed to be D-dimensional. The calculation process of the lth layer is as follows:
[0080] z l ′ =MSA(LN(z l-1 ))+z l-1 ,l=1,2,···,L
[0081] z l =MLP(LN(z l ′ ))+z l ′ ,l=1,2,···,L
[0082] Represents the feature vector obtained after L encoding layers.
[0083] (2.4) RF map output
[0084] Get z L After that, reshape it into two dimensions with size The feature map of , thereby restoring its spatial position information. In the present invention, D = P 2 , so the P of the feature map can be 2 The channels are rearranged to the same dimension to obtain a grayscale RF map of size 1×H×W.
Claims
1. A radio frequency map estimation method for industrial indoor scenes, characterized in that: The following steps are involved: Construct a simulated RF map dataset based on ray tracing to train the CGAN model, and use the CGAN model to generate new RF map samples; Adopting the ViT-based RF map estimation method, the simulated RF map dataset is expanded with new RF map samples generated by the CGAN model, and the expanded dataset is used to train the ViT model; Given a floor plan of an industrial indoor scene, it is input into the trained ViT model to estimate the corresponding RF map in real time.
2. The radio frequency map estimation method for industrial indoor scenes according to claim 1 is characterized in that: The simulated radio frequency map data set based on ray tracing includes N×M radio frequency maps of size H×W, and each radio frequency map stores a path loss value between the pixel position and the transmitter at each pixel; N is the number of workshop floor plans, each workshop floor plan is configured with a corresponding simulated RF map, where each workshop floor plan shares the same M transmitter locations; Divide each workshop plan into grid pixels according to a set resolution and convert it into a two-dimensional image of H×W pixels; the two-dimensional image contains workshop building materials with different electromagnetic properties; In each simulation, for each workshop floor plan, only one of the M transmitters is in working state.
3. The radio frequency map estimation method for industrial indoor scenes according to claim 1 is characterized in that: The method of using the CGAN model to generate a new radio frequency map sample includes the following steps: 1) Each workshop plan is discretized into H×W grids, and each workshop i has a plane image with material information For the jth transmitter position, j = 1, 2, ···, M, represents the RF map generated according to the 3GPP InF model; when the jth transmitter in workshop i is working, scenario X ij Expressed as Among them, concatenate means concatenating single-channel images of the same size F i and T j X is connected in series along the channel dimension to form a dual channel ij ; Given X ij , Y ij Realistic RF maps generated for ray tracing simulations; [X ij ,Y ij ] constitutes a labeled training sample; based on the labeled training sample [X ij ,Y ij ]Learn the network model f θ (·), to estimate the entire RF map Right now Among them, θ represents the set of network parameters; 2) Construct a CGAN model consisting of a generator and a discriminator, and use the CGAN generator to represent f θ (·); where the conditional information input into the CGAN model is used to represent the position of the working transmitter in the workshop The location information of the transmitter in the scene is encoded using one-hot encoding. After training the CGAN model using a simulated RF map dataset, a RF map is generated using the CGAN model.
4. The method for estimating a radio frequency map for industrial indoor scenes according to claim 1, characterized in that: The ViT-based radio frequency map estimation method uses new radio frequency map samples generated by the CGAN model to expand the simulated radio frequency map dataset, and uses the expanded dataset to train the ViT model, including the following steps: (1) Image serialization: In the ViT model, we first convert the input image of size 2×H×W into Crop to image blocks of size 2×P×P, and expand each image block into a one-dimensional vector, and finally obtain (2) Image block embedding and position embedding: Use the linear projection operator E to embed the image block sequence X p Mapped to the D-dimensional feature space, the image block embedding and position embedding together form the D-dimensional z0 as the input of the ViT encoder: in, To achieve the linear projection matrix, represents position embedding, Represents the Nth image block sequence; the ViT model uses sine and cosine functions of different frequencies to achieve position embedding: Where pos represents the position of the image block in the image, i∈[0,1,···,D / 2] is the index used to calculate the channel dimension, and D represents the dimension of the feature space; PE (pos,2i) Represents the position embedding at the 2i-th dimension pos; the position embedding needs to have the same dimension as the image block embedding; (3) Constructing the ViT encoder: The ViT encoder is composed of L identical encoding layers stacked together. Each encoding layer includes a multi-head self-attention sublayer and a multi-layer perceptron sublayer. Before the data enters each sublayer, it is normalized using layer normalization. After passing through each sublayer, the data is directly fused with the input using residual connections. In order to achieve residual connection, the output dimension of each layer of the ViT encoder is designed to be D-dimensional; the calculation process of the lth layer is as follows: z′ l =MSA(LN(z l-1 ))+z l-1 ,l=1,2,···,L With l =MLP(LN(z′ l ))+z′ l ,l=1,2,···,L in, represents the feature vector obtained after L encoding layers; MSA and MLP represent multi-head self-attention mechanism and multi-layer perceptron respectively, and LN represents layer normalization; (4) RF map output: Get z L After that, reshape it into two dimensions with size The feature map of D = P 2 , the feature map P 2 The channels are rearranged to the same dimension to obtain a grayscale RF map of size 1×H×W.
5. A radio frequency map estimation method for industrial indoor scenes, characterized in that: include: The sample construction module is used to construct a simulated RF map dataset based on ray tracing to train the CGAN model and use the CGAN model to generate new RF map samples; A model building module, for adopting a ViT-based RF map estimation method, augmenting a simulated RF map dataset with new RF map samples generated by a CGAN model, and training a ViT model using the augmented dataset; The RF map estimation module is used to input the trained ViT model based on the given industrial indoor scene floor plan and estimate the corresponding RF map in real time.
6. A radio frequency map estimation device for industrial indoor scenes, characterized in that: It comprises a memory and a processor; the memory is used to store a computer program; the processor is used to implement a radio frequency map estimation method for industrial indoor scenes as described in any one of claims 1 to 4 when executing the computer program.
7. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by the processor, the radio frequency map estimation method for industrial indoor scenes as described in any one of claims 1 to 4 is implemented.