Model training method, radio map reconstruction method, device, equipment and medium

Abstract

The embodiment of the present application provides a model training method, a radio map reconstruction method, an apparatus, a device and a medium, which belongs to the field of wireless communication technology. The method includes: initializing multiple three-dimensional Gaussian splash models according to the three-dimensional point cloud of the area covered by the base station, projecting each three-dimensional Gaussian splash model to the reference signal receiving power space model based on the grid position coordinates of the grid in the coverage area, obtaining multiple projection coefficients, rendering each three-dimensional Gaussian splash model according to each projection coefficient, grid position coordinates and corresponding base station coordinate embedding vector, obtaining the predicted reference signal receiving power of the grid, determining the target loss according to the predicted reference signal receiving power of each grid and the corresponding reference reference signal receiving power, updating each three-dimensional Gaussian splash model, projection coefficient and reference signal receiving power space model according to the target loss, obtaining a radio map reconstruction model, which can improve the accuracy of radio map reconstruction.

Description

Technical Field

[0001] The present application relates to the field of wireless communication technology, and in particular to a model training method, a radio map reconstruction method, a device, a equipment and a medium. Background Art

[0002] Wireless networks, such as the fifth and sixth generation standard cellular networks, can provide large-scale connectivity for smart terminal devices, various ultra-wideband services, and high-precision perception. These functions require the support of high-precision radio maps to accurately assess the radio environment. A radio map is a heat map that describes the spatial distribution of signal strength. It records the distribution of power information in geometric position, time, and frequency. In related technologies, radio maps are reconstructed based on sparse observation data of wireless signals collected by sensors and mobile devices. However, sparse observation data is short of data and it is difficult to fully represent fine-grained environmental details, resulting in a decrease in the accuracy of radio map reconstruction. Summary of the invention

[0003] The main purpose of the embodiments of the present application is to propose a model training method, a radio map reconstruction method, an apparatus, a device and a medium, aiming to improve the accuracy of radio map reconstruction.

[0004] To achieve the above objectives, a first aspect of an embodiment of the present application proposes a model training method, the method comprising:

[0005] Acquire a three-dimensional point cloud of the coverage area of ​​the base station, and initialize multiple three-dimensional Gaussian splash models according to the three-dimensional point cloud;

[0006] Dividing the coverage area into a plurality of grids, and obtaining grid position coordinates and a reference signal receiving power of each grid;

[0007] For each grid, projecting each three-dimensional Gaussian spreading model onto a reference signal received power spatial model based on the grid position coordinates to obtain a projection coefficient of each three-dimensional Gaussian spreading model;

[0008] Determine a base station coordinate embedding vector of each three-dimensional Gaussian splash model according to the base station coordinates of the base station;

[0009] Rendering each three-dimensional Gaussian spreading model according to the projection coefficient of each three-dimensional Gaussian spreading model, the grid position coordinates and the corresponding base station coordinate embedding vector to obtain the predicted reference signal received power of the grid;

[0010] Perform loss calculation based on the predicted reference signal received power of each grid and the corresponding reference signal received power to obtain a target loss;

[0011] Each three-dimensional Gaussian spreading model, the projection coefficient, and the reference signal received power spatial model are updated according to the target loss to obtain a radio map reconstruction model.

[0012] In some embodiments, the three-dimensional Gaussian splash model has a three-dimensional covariance matrix and a three-dimensional spatial position coordinate, the reference signal received power spatial model includes a first multi-layer perceptron and a second multi-layer perceptron, and the projecting of each three-dimensional Gaussian splash model onto the reference signal received power spatial model based on the grid position coordinate to obtain the projection coefficient of each three-dimensional Gaussian splash model includes:

[0013] Based on the grid position coordinates, projecting the three-dimensional covariance matrix of each three-dimensional Gaussian spreading model through the first multi-layer perceptron to obtain the projection covariance feature of each three-dimensional Gaussian spreading model;

[0014] Projecting the three-dimensional spatial position coordinates of each three-dimensional Gaussian splash model through the second multi-layer perceptron to obtain projection coordinate features of each three-dimensional Gaussian splash model;

[0015] The projection coefficient of each three-dimensional Gaussian spreading model is determined according to the projection covariance feature of each three-dimensional Gaussian spreading model, the corresponding projection coordinate feature, the preset continuous tensor and the preset projection scale.

[0016] In some embodiments, the three-dimensional Gaussian splash model has three-dimensional spatial position coordinates, and determining the base station coordinate embedding vector of each three-dimensional Gaussian splash model according to the base station coordinates of the base station includes:

[0017] Determine a gap vector of each three-dimensional Gaussian splash model according to the base station coordinates and the three-dimensional spatial position coordinates of each three-dimensional Gaussian splash model;

[0018] Position encoding is performed on the gap vector of each three-dimensional Gaussian splash model to obtain an encoding vector of each three-dimensional Gaussian splash model;

[0019] Perform vector projection on the encoding vector of each three-dimensional Gaussian splash model to obtain the base station coordinate embedding vector of each three-dimensional Gaussian splash model.

[0020] In some embodiments, the three-dimensional Gaussian splash model has three-dimensional spatial position coordinates, opacity and spherical harmonic functions, and rendering each three-dimensional Gaussian splash model according to the projection coefficient of each three-dimensional Gaussian splash model, the grid position coordinates and the corresponding base station coordinate embedding vector to obtain the predicted reference signal received power of the grid includes:

[0021] Projecting the opacity of each three-dimensional Gaussian splash model according to the projection coefficient of each three-dimensional Gaussian splash model and the corresponding base station coordinate embedding vector to obtain the projected opacity of each three-dimensional Gaussian splash model;

[0022] Performing function calculation on the corresponding projection coefficient, the grid position coordinates and the three-dimensional space position coordinates through the spherical harmonic function of each three-dimensional Gaussian spreading model to obtain the expected reference signal received power of each three-dimensional Gaussian spreading model;

[0023] The projection opacity of each three-dimensional Gaussian splash model is screened to obtain a reference opacity;

[0024] Rendering the expected reference signal received power of each three-dimensional Gaussian splash model according to the projection opacity of each three-dimensional Gaussian splash model and the reference opacity to obtain multiple rendering powers;

[0025] Each rendering power is summed to obtain the predicted reference signal received power of the grid.

[0026] In some embodiments, the performing loss calculation according to the predicted reference signal received power of each grid and the corresponding reference reference signal received power to obtain the target loss includes:

[0027] Perform loss calculation according to the predicted reference signal received power of each grid and the corresponding reference reference signal received power to obtain a first loss;

[0028] Calculate the grid distance between any two grids;

[0029] For each grid, select the grid with the smallest grid distance from the grid as the reference grid;

[0030] Perform loss calculation according to the predicted reference signal received power of each grid and the predicted reference signal received power of the corresponding reference grid to obtain a second loss;

[0031] The first loss and the second loss are summed to obtain the target loss.

[0032] In some embodiments, initializing a plurality of three-dimensional Gaussian splash models according to the three-dimensional point cloud comprises:

[0033] Randomly selecting a plurality of target points from the three-dimensional point cloud;

[0034] Initialize multiple 3D Gaussian splash models based on multiple target points.

[0035] To achieve the above objective, a second aspect of an embodiment of the present application proposes a radio map reconstruction method, the method comprising:

[0036] Obtain target base station coordinates of the target base station;

[0037] Dividing the coverage area of ​​the target base station into a plurality of target grids, and obtaining the target position coordinates of each target grid;

[0038] Predicting the target reference signal received power of each target grid based on the target base station coordinates and the target position coordinates of each target grid through a radio map reconstruction model; wherein the radio map reconstruction model is trained according to the model training method described in the first aspect;

[0039] The radio map is reconstructed according to each target reference signal received power to obtain a target radio map.

[0040] To achieve the above-mentioned purpose, a third aspect of an embodiment of the present application provides a radio map reconstruction device, the device comprising:

[0041] An acquisition module, used to acquire the target base station coordinates of the target base station;

[0042] A division module, used to divide the coverage area of ​​the target base station into a plurality of target grids, and obtain the target position coordinates of each target grid;

[0043] A prediction module, configured to predict the target reference signal received power of each target grid based on the target base station coordinates and the target position coordinates of each target grid through a radio map reconstruction model; wherein the radio map reconstruction model is trained according to the model training method described in the first aspect;

[0044] The reconstruction module is used to reconstruct the radio map according to each target reference signal receiving power to obtain a target radio map.

[0045] To achieve the above-mentioned purpose, the fourth aspect of an embodiment of the present application proposes an electronic device, which includes a memory and a processor, the memory stores a computer program, and the processor implements the model training method described in the first aspect or the radio map reconstruction method described in the second aspect when executing the computer program.

[0046] To achieve the above-mentioned purpose, the fifth aspect of an embodiment of the present application proposes a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the model training method described in the first aspect or the radio map reconstruction method described in the second aspect.

[0047] The model training method, radio map reconstruction method, radio map reconstruction device, electronic device and computer-readable storage medium of the embodiments of the present application obtain a three-dimensional point cloud of the coverage area of ​​the base station to accurately perceive the environment based on the three-dimensional point cloud, thereby making up for the defect that sparse observation data cannot fully express the details of the environment. Multiple three-dimensional Gaussian sprinkling models are initialized according to the three-dimensional point cloud to accurately depict the spatial layout, material characteristics and other detailed features of objects in the environment using the three-dimensional Gaussian sprinkling model, generate high-quality three-dimensional scene representation, and alleviate the problem of data shortage. There are differences in the signal strength of the area covered by the base station. In order to achieve a refined analysis of the signal strength of the area covered by the base station, the coverage area of ​​the base station is divided into multiple grids. The grid position coordinates and the benchmark reference signal receiving power of each grid are obtained to construct an accurate radio map reconstruction model based on the grid position coordinates and the benchmark reference signal receiving power. In order to predict the signal strength of each grid, for each grid, each three-dimensional Gaussian splash model is projected onto the reference signal received power space model based on the grid position coordinates to obtain the projection coefficient of each three-dimensional Gaussian splash model, so as to represent the material properties of the three-dimensional Gaussian splash model such as opacity based on the projection coefficient, and then accurately characterize the wireless environment according to the three-dimensional Gaussian splash model. The signal strength is related to the location where the base station is set. In order to determine the signal strength of each grid, the base station coordinates of the base station are obtained. In order to extract the features that act on the three-dimensional Gaussian splash model from the base station coordinates, the base station coordinate embedding vector of each three-dimensional Gaussian splash model is determined according to the base station coordinates. In order to enable the three-dimensional Gaussian splash model to accurately model the wireless environment, each three-dimensional Gaussian splash model is rendered according to the projection coefficients, grid position coordinates and corresponding base station coordinate embedding vectors of each three-dimensional Gaussian splash model to predict the signal strength of the grid and obtain the predicted reference signal received power of the grid. The loss is calculated according to the predicted reference signal received power of each grid and the corresponding reference reference signal received power to obtain the target loss, so as to guide the model optimization process based on the target loss. Each three-dimensional Gaussian spreading model, projection coefficient and reference signal received power space model are updated according to the target loss to perform model optimization and obtain a radio map reconstruction model, thereby improving the accuracy of radio map reconstruction. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 is a flow chart of the model training method provided in an embodiment of the present application;

[0049] Figure 2 yes Figure 1 Flow chart of step S110 in FIG.

[0050] Figure 3 yes Figure 1 Flow chart of step S130 in FIG.

[0051] Figure 4 yes Figure 1 Flow chart of step S140 in FIG.

[0052] Figure 5 yes Figure 1 Flow chart of step S150 in FIG.

[0053] Figure 6 yes Figure 1 Flow chart of step S160 in FIG.

[0054] Figure 7 is a flow chart of a radio map reconstruction method provided by an embodiment of the present application;

[0055] Figure 8 is a schematic diagram of the structure of a radio map reconstruction device provided in an embodiment of the present application;

[0056] Fig. 9 It is a schematic diagram of the hardware structure of the electronic device provided in the embodiment of the present application. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0058] It should be noted that, although the functional modules are divided in the device schematic diagram and the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first", "second", etc. in the specification, claims and the above drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence.

[0059] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which this application belongs. The terms used herein are only for the purpose of describing the embodiments of this application and are not intended to limit this application.

[0060] Wireless networks, such as the fifth and sixth generation standard cellular networks, can provide large-scale connectivity for smart terminal devices, various ultra-wideband services, and high-precision perception. These functions require the support of high-precision radio maps to accurately assess the radio environment. A radio map is a heat map that describes the spatial distribution of signal strength. It records the distribution of power information in geometric position, time, and frequency. In related technologies, radio maps are reconstructed based on sparse observation data of wireless signals collected by sensors and mobile devices. However, sparse observation data is short of data and it is difficult to fully represent fine-grained environmental details, resulting in a decrease in the accuracy of radio map reconstruction.

[0061] Based on this, the embodiments of the present application provide a model training method, a radio map reconstruction method, a radio map reconstruction device, an electronic device and a computer-readable storage medium, aiming to improve the accuracy of radio map reconstruction.

[0062] The model training method, radio map reconstruction method, radio map reconstruction device, electronic device and computer-readable storage medium provided in the embodiments of the present application are specifically described through the following embodiments. First, the model training method in the embodiments of the present application is described.

[0063] The model training method provided in the embodiment of the present application relates to the field of wireless communication technology. The model training method provided in the embodiment of the present application can be applied to a terminal, can be applied to a server side, or can be software running in a terminal or a server side. In some embodiments, the terminal can be a smart phone, a tablet computer, a laptop computer, a desktop computer, etc.; the server side can be configured as an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms; the software can be an application that implements the model training method, etc., but is not limited to the above forms.

[0064] The present application can be used in many general or special computer system environments or configurations. For example: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, etc. The present application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. The present application can also be practiced in distributed computing environments, in which tasks are performed by remote processing devices connected through a communication network. In a distributed computing environment, program modules can be located in local and remote computer storage media including storage devices.

[0065] Figure 1 It is an optional flowchart of the model training method provided in an embodiment of the present application. The model training method may include but is not limited to steps S110 to S170.

[0066] Step S110, obtaining a three-dimensional point cloud of the coverage area of ​​the base station, and initializing multiple three-dimensional Gaussian splash models according to the three-dimensional point cloud;

[0067] Step S120, dividing the coverage area into a plurality of grids, and obtaining the grid position coordinates and the reference signal received power of each grid;

[0068] Step S130, for each grid, projecting each three-dimensional Gaussian spreading model onto the reference signal received power space model based on the grid position coordinates to obtain a projection coefficient of each three-dimensional Gaussian spreading model;

[0069] Step S140, determining a base station coordinate embedding vector of each three-dimensional Gaussian splash model according to the base station coordinates of the base station;

[0070] Step S150, rendering each three-dimensional Gaussian spreading model according to the projection coefficient of each three-dimensional Gaussian spreading model, the grid position coordinates and the corresponding base station coordinate embedding vector to obtain the predicted reference signal received power of the grid;

[0071] Step S160, performing loss calculation according to the predicted reference signal received power of each grid and the corresponding reference signal received power to obtain a target loss;

[0072] Step S170 , updating each three-dimensional Gaussian spreading model, projection coefficient, and reference signal received power spatial model according to the target loss to obtain a radio map reconstruction model.

[0073] In step S110 of some embodiments, a downlink multiple-input multiple-output wireless system is considered, the multiple-input multiple-output wireless system includes a base station, the base station is equipped with an antenna array for receiving and transmitting wireless signals, the antenna array includes antennas. The base station uses the codebook The reference signal is sent using different beamforming vectors. is the i-th beamforming vector in the codebook to scan the entire angle space and determine the number of spatial angles after dividing the entire angle three-dimensional space. The three-dimensional point cloud of the area covered by the base station is collected through laser radar, depth camera, structured light sensor and other equipment. The coverage area is the coverage range of the base station network. The three-dimensional point cloud is a data structure used to describe the wireless environment. The three-dimensional point cloud includes multiple three-dimensional coordinate points, each point corresponds to a sampling position in the wireless environment. The three-dimensional point cloud can accurately perceive the complex wireless environment. Initialize multiple three-dimensional Gaussian splash models according to the number of spatial angles and the three-dimensional point cloud. The number of three-dimensional Gaussian splash models is the same as the number of spatial angles.

[0074] The three-dimensional Gaussian splash model is represented by a three-dimensional anisotropic Gaussian sphere in the angle space. Each Gaussian sphere has learnable parameters or property characteristics, such as the three-dimensional spatial position coordinates. , Opacity , three-dimensional covariance matrix , color value c, expected reference signal received power (RSRP) rsrp, etc., to describe the spatial layout and material characteristics of objects in the wireless environment through these properties. That is, the three-dimensional Gaussian splash model has properties such as three-dimensional spatial position coordinates, opacity, three-dimensional covariance matrix, and expected reference signal received power. In the task of radio map construction, high-quality real-time rendering of the wireless environment can be achieved by updating these properties of the three-dimensional Gaussian splash model, which improves the accuracy of radio map construction, thereby providing technical support for various ultra-wideband services and high-precision perception requirements of the new generation of wireless communication systems.

[0075] The three-dimensional covariance matrix can be decomposed into a rotation matrix and a scaling matrix. The three-dimensional covariance matrix is ​​expressed as:

[0076] ,

[0077] in, represents the three-dimensional covariance matrix; R represents the rotation matrix; S represents the scaling matrix; T represents the transpose operation.

[0078] The three-dimensional covariance matrix satisfies the properties of a semi-positive definite matrix, that is, all eigenvalues ​​in the matrix are greater than or equal to 0, otherwise numerical instability may occur. This anisotropic covariance representation suitable for optimization allows the optimization of a three-dimensional Gaussian sphere to adaptively capture different geometric shapes in the environmental scene, thereby producing a compact feature representation for the wireless environment. Compared with the acquisition of image data, the measurement data of wireless signals, such as the received power of the reference signal, is often very small, and the data shortage problem needs to be considered when constructing a radio map reconstruction model. In the three-dimensional covariance matrix, the rotation matrix R can be represented by a quaternion, i.e., four parameters, which has fewer parameters and more stable numerical characteristics, and does not require additional orthogonalization steps, thereby reducing the number of parameters, reducing the computational cost during model training and inference, and alleviating the problem of shortage of training data required for the model.

[0079] See also Figure 2 In some embodiments, step S110 may include but is not limited to steps S210 to S220:

[0080] Step S210, randomly selecting multiple target points from the three-dimensional point cloud;

[0081] Step S220 , initializing multiple three-dimensional Gaussian splash models according to multiple target points.

[0082] In step S210 of some embodiments, the number of three-dimensional Gaussian spreading models is fixed in advance to N according to the number of spatial angles N, and N points are randomly selected from the three-dimensional point cloud describing the wireless environment as target points.

[0083] In step S220 of some embodiments, the three-dimensional spatial position coordinates of N three-dimensional Gaussian splash models are initialized according to the three-dimensional coordinates of N target points, so as to fully utilize the multimodal information of the three-dimensional point cloud to accurately perceive the wireless environment, and accurately capture the spatial layout and material characteristics of objects in the environment through the three-dimensional Gaussian sphere.

[0084] In the above steps S210 to S220, the three-dimensional Gaussian spreading model is initialized through the three-dimensional point cloud, so that the three-dimensional Gaussian spreading model can accurately perceive the wireless environment based on the multimodal information of the three-dimensional point cloud, thereby making up for the shortage of wireless signal measurement data and improving the accuracy of radio map reconstruction.

[0085] In step S120 of some embodiments, the coverage area of ​​the base station is divided into L grids, and the grid position coordinates and the benchmark reference signal receiving power of each grid in the coverage area are obtained. The grid position coordinates are the real geographic three-dimensional coordinates of the grid, and the benchmark reference signal receiving power is the real reference signal receiving power measured at the grid at the tth moment. The radio map is reconstructed by fusing the measurement data of the wireless signal and the three-dimensional environmental point cloud data, so as to realize wireless channel modeling with multi-source data fusion.

[0086] No. A single-antenna user device in a grid can measure the channel impulse response at the tth time, and determine the reference signal received power information based on the channel impulse response, beamforming vector, and base station transmit power. This information includes the incident angle and channel gain information from the base station to the user device. The reference signal received power for the mth beamforming vector measured by the grid at the tth time is expressed as:

[0087] ,

[0088] Where P is the transmission power of the base station; For the The channel impulse response measured by the grid at the tth time; H represents the conjugate transpose operation; , N is the number of spatial angles, is the array steering vector, , and represent the elevation incident angle, azimuth incident angle and channel gain of the nth potential path respectively.

[0089] Specifically, the base station is divided into three cells, each of which is equipped with a set of antennas pointing in a specific direction. The test phone is connected to the base station via electromagnetic waves with a center frequency of 3.6 GHz, and the data measured by the test phone is recorded by installing special computer software. The data is the RSRP measurement value in the beam domain. The base station sends two beams, the Synchronization Signal Block (SSB) beam and the Channel State Information-Reference Signal (CSI-RS) beam, to the test phone alternately at a certain period to measure the RSRP measurement values ​​corresponding to the beams. These RSRP measurement values ​​will be used to build a radio map reconstruction model.

[0090] The coverage area of ​​the base station is discretized into multiple geographic grids (grids), each of which can be regarded as a 10m x 10m rectangle. The measurement time is about three minutes. For the measurement activities, a large file is downloaded using a mobile phone to simulate full-load traffic. The storage space occupied by the large file can be greater than 100 GB. At the same time, the mobile phone is carried in the coverage area of ​​the base station network to collect RSRP measurement values ​​in different beam directions. For each cell, there are synchronization signal block beams and channel state information reference signal beams. Each beam has a beam index and a corresponding RSRP measurement value.

[0091] The propagation of wireless signals is more complex than that of visible light, and we need to reconsider how to map the three-dimensional Gaussian sphere in three-dimensional space to the one-dimensional multi-beam reference signal received power space model. The specific mapping process of the three-dimensional Gaussian sphere is described in detail below.

[0092] See also Figure 3 In some embodiments, the reference signal received power space model includes a first multilayer perceptron and a second multilayer perceptron, the first multilayer perceptron and the second multilayer perceptron are two parallel multilayer perceptrons, and step S130 may include but is not limited to steps S310 to S330:

[0093] Step S310, based on the grid position coordinates, projecting the three-dimensional covariance matrix of each three-dimensional Gaussian spreading model through a first multi-layer perceptron to obtain the projection covariance feature of each three-dimensional Gaussian spreading model;

[0094] Step S320, projecting the three-dimensional spatial position coordinates of each three-dimensional Gaussian splash model through a second multi-layer perceptron to obtain projection coordinate features of each three-dimensional Gaussian splash model;

[0095] Step S330, determining the projection coefficient of each three-dimensional Gaussian splatting model according to the projection covariance feature of each three-dimensional Gaussian splatting model, the corresponding projection coordinate feature, the preset continuous tensor and the preset projection scale.

[0096] In step S310 of some embodiments, for each grid, the distance between the grid and each three-dimensional Gaussian splash model is calculated according to the grid position coordinates of the grid and the three-dimensional spatial position coordinates of each three-dimensional Gaussian splash model, and each three-dimensional Gaussian splash model is sorted in order from small to large according to the distance to determine the model order corresponding to the grid. The grid position coordinates of the grid are , the three-dimensional spatial position coordinates of the nth three-dimensional Gaussian splash model are , No. The distance between a grid and the nth three-dimensional Gaussian splash model is expressed as , 2 represents the 2-norm.

[0097] According to the model order, the first multi-layer perceptron performs feature projection on the three-dimensional covariance matrix of each three-dimensional Gaussian splatting model in turn to obtain the projection covariance feature of each three-dimensional Gaussian splatting model. That is:

[0098] ,

[0099] in, is the three-dimensional covariance matrix of the i-th three-dimensional Gaussian splatter model, , Represents the real number space, the matrix dimension of the three-dimensional covariance matrix is ​​3×3; represents the first multi-layer perceptron; is the projection covariance feature of the i-th three-dimensional Gaussian splatter model, , the feature dimension of the projection covariance feature is one dimension.

[0100] In step S320 of some embodiments, the outer layer of the second multilayer perceptron is provided with a sigmoid activation function, and the second multilayer perceptron performs feature projection on the three-dimensional spatial position coordinates of each three-dimensional Gaussian splatting model in turn according to the model order to obtain the projection coordinate features of each three-dimensional Gaussian splatting model, and the projection coordinate features are continuous tensors with values ​​in the interval [0,31]. That is:

[0101] ,

[0102] in, is the projection coordinate feature of the i-th three-dimensional Gaussian splatter model; sigmoid represents the sigmoid activation function; represents the second multi-layer perceptron; represents the three-dimensional spatial position coordinates of the i-th three-dimensional Gaussian splash model, ; * indicates multiplication operation.

[0103] In step S330 of some embodiments, the preset continuous tensor is a continuous tensor whose value is in the range of [0,31], and its dimension is the same as the dimension of the projection coordinate feature. The preset projection scale is the scaling ratio of the three-dimensional Gaussian splash model mapped to the reference signal received power space model, which can be set to 0.01. According to the projection coordinate features, the preset continuous tensor and the preset projection scale of each three-dimensional Gaussian splash model, the distance features of each three-dimensional Gaussian splash model are determined. According to the distance features and the corresponding projection covariance features of each three-dimensional Gaussian splash model, the projection coefficient of each three-dimensional Gaussian splash model is determined. The projection coefficient of the i-th three-dimensional Gaussian splash model is expressed as:

[0104] ,

[0105] ,

[0106] Wherein, Coord_RSRP is a preset continuous tensor, and Coord_RSRP is expressed as [0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31]; is the distance feature of the i-th three-dimensional Gaussian splash model; is the projection coordinate feature of the i-th three-dimensional Gaussian splash model; * represents multiplication operation; scaling is the preset projection scale; is the projection covariance feature of the i-th three-dimensional Gaussian splash model; is the projection coefficient of the i-th three-dimensional Gaussian splash model; exp represents the exponential calculation with base e.

[0107] Through the above steps S310 to S330, projection coefficients can be obtained to project the three-dimensional Gaussian spreading model based on the projection coefficients, thereby extracting features useful for the radio map reconstruction task from the three-dimensional Gaussian spreading model.

[0108] See also Figure 4 In some embodiments, step S140 may include but is not limited to steps S410 to S430:

[0109] Step S410, determining a gap vector of each three-dimensional Gaussian spreading model according to the base station coordinates and the three-dimensional spatial position coordinates of each three-dimensional Gaussian spreading model;

[0110] Step S420, position encoding is performed on the gap vector of each three-dimensional Gaussian splash model to obtain the encoding vector of each three-dimensional Gaussian splash model;

[0111] Step S430, performing vector projection on the encoding vector of each three-dimensional Gaussian spreading model to obtain the base station coordinate embedding vector of each three-dimensional Gaussian spreading model.

[0112] In step S410 of some embodiments, a three-dimensional Gaussian sphere is used as a virtual transmission source to replace the base station used to transmit wireless signals in the real scene. The strength of the wireless signal emitted by the virtual transmission source is related to the relative position and direction of the base station. Therefore, the base station coordinates of the base station can be obtained, and the difference vector between the base station coordinates and the three-dimensional spatial position coordinates of each three-dimensional Gaussian splash model is calculated to obtain the gap vector of each three-dimensional Gaussian splash model, so as to express the relative position and direction of the three-dimensional Gaussian sphere and the base station based on the gap vector, thereby realizing the implicit expression of the wireless signal strength. The gap vector of the i-th three-dimensional Gaussian splash model is expressed as:

[0113] ,

[0114] in, is the gap vector of the i-th three-dimensional Gaussian splash model; represents the three-dimensional spatial position coordinates of the i-th three-dimensional Gaussian splash model; represents the base station coordinates of the base station; N is the number of three-dimensional Gaussian splashing models.

[0115] In step S420 of some embodiments, for each three-dimensional Gaussian splash model, the gap vector of the three-dimensional Gaussian splash model is position-encoded based on the encoding function to obtain an encoding vector. The encoding function can convert the gap vector from Map to , the encoding function is expressed as:

[0116] ,

[0117] Among them, p represents the input parameter, which uses a gap vector; M is the encoding scale of the encoding function, which can be 10; sin is the sine function; and cos is the cosine function.

[0118] In step S430 of some embodiments, the encoding vector of each three-dimensional Gaussian splash model is projected by an explicit base station coordinate embedding module to extract features in the base station coordinates that affect the opacity of each three-dimensional Gaussian sphere, thereby obtaining a base station coordinate embedding vector of each three-dimensional Gaussian splash model. The explicit base station coordinate embedding module may use a multi-layer perceptron.

[0119] Through the above steps S410 to S430, the features that affect each three-dimensional Gaussian sphere can be extracted from the difference vector between the base station position and the position of each three-dimensional Gaussian sphere, so as to update the three-dimensional Gaussian sphere based on the features, thereby accurately characterizing the spatial layout and material characteristics of objects in the wireless environment and achieving accurate modeling of the radio map.

[0120] See also Figure 5 In some embodiments, step S150 may include but is not limited to steps S510 to S550:

[0121] Step S510, projecting the opacity of each three-dimensional Gaussian splash model according to the projection coefficient of each three-dimensional Gaussian splash model and the corresponding base station coordinate embedding vector to obtain the projected opacity of each three-dimensional Gaussian splash model;

[0122] Step S520, performing function calculation on the corresponding projection coefficient, grid position coordinates and three-dimensional space position coordinates through the spherical harmonic function of each three-dimensional Gaussian spreading model to obtain the expected reference signal received power of each three-dimensional Gaussian spreading model;

[0123] Step S530, screening the projection opacity of each three-dimensional Gaussian splash model to obtain a reference opacity;

[0124] Step S540, rendering the expected reference signal received power of each three-dimensional Gaussian splash model according to the projection opacity and reference opacity of each three-dimensional Gaussian splash model to obtain multiple rendering powers;

[0125] Step S550 , summing each rendering power to obtain the predicted reference signal received power of the grid.

[0126] In step S510 of some embodiments, for each three-dimensional Gaussian splash model, the opacity is projected according to the projection coefficient and the base station coordinate embedding vector to obtain the projected opacity. The projected opacity is the opacity after projection, which is expressed as:

[0127] ,

[0128] in, , , and are respectively the projection opacity, projection coefficient, opacity and base station coordinate embedding vector of the i-th three-dimensional Gaussian splash model, and the dimension of the projection opacity is 32.

[0129] In step S520 of some embodiments, the three-dimensional Gaussian splash model also has spherical harmonic functions, which are a set of orthogonal basis functions defined on a sphere. The order of the spherical harmonic function can be set in advance, such as a fourth-order spherical harmonic function. For each three-dimensional Gaussian splash model, the projection coefficient, grid position coordinates, and three-dimensional space position coordinates are input into the spherical harmonic function for function calculation to obtain the expected reference signal received power. The expected reference signal received power is the expected RSRP value of the three-dimensional Gaussian splash model, which is the weighted sum of the spherical harmonic function and its coefficients. The expected reference signal received power of the i-th three-dimensional Gaussian splash model is It is expressed as:

[0130] ,

[0131] Wherein, SH represents spherical harmonic function, which can be a fourth-order spherical harmonic function; is the projection coefficient of the i-th three-dimensional Gaussian splash model, including coefficients of different orders of spherical harmonic functions, is the highest order of spherical harmonics; Indicates Spherical harmonics of order; For the Coefficients of spherical harmonics of order; represents the three-dimensional spatial position coordinates of the i-th three-dimensional Gaussian splash model; Indicates The grid position coordinates of the grid.

[0132] The dimension of each coefficient of the spherical harmonic function is aligned with the dimension of the expected reference signal received power, that is, the dimension of each coefficient of the spherical harmonic function is 32, which ensures that the expected output rsrp of the spherical harmonic function is non-negative.

[0133] In step S530 of some embodiments, according to the model order obtained in step S310, the projection opacity of each 3D Gaussian splash model before the current 3D Gaussian splash model is used as the reference opacity. If there is no 3D Gaussian splash model before the current 3D Gaussian splash model, the reference opacity is zero.

[0134] In step S540 of some embodiments, for each three-dimensional Gaussian splash model, one is subtracted from multiple reference opacities corresponding to the three-dimensional Gaussian splash model, and multiple subtraction results are multiplied to obtain an intermediate parameter. The projection opacity, the intermediate parameter and the expected reference signal received power are multiplied to obtain the rendering power.

[0135] In step S550 of some embodiments, the rendering power of each three-dimensional Gaussian splash model is summed to obtain the predicted reference signal received power of the grid. The predicted reference signal received power is expressed as:

[0136] ,

[0137] in, is the expected reference signal received power of the i-th three-dimensional Gaussian splash model; is the projection opacity of the i-th three-dimensional Gaussian splash model; N is the number of models of the three-dimensional Gaussian splash model; Represents the j-th reference opacity.

[0138] Through the above steps S510 to S550, the predicted reference signal received power can be obtained, so as to construct a radio map reconstruction model based on the predicted reference signal received power.

[0139] See also Figure 6 In some embodiments, step S160 may include but is not limited to steps S610 to S650:

[0140] Step S610, performing loss calculation according to the predicted reference signal received power of each grid and the corresponding reference signal received power to obtain a first loss;

[0141] Step S620, calculating the grid distance between any two grids;

[0142] Step S630, for each grid, selecting the grid with the smallest grid distance between the grids as the reference grid;

[0143] Step S640, performing loss calculation according to the predicted reference signal received power of each grid and the predicted reference signal received power of the corresponding reference grid to obtain a second loss;

[0144] Step S650, summing the first loss and the second loss to obtain a target loss.

[0145] In step S610 of some embodiments, the power difference between the predicted reference signal received power of each grid and the corresponding reference signal received power is calculated, and the power difference value of each grid is summed to obtain a first loss. In the radio map interpolation task, the first loss can be used as the target loss. The formula for loss calculation is expressed as:

[0146] ,

[0147] in, Indicates the first loss; represents the predicted reference signal received power of grid b; represents the benchmark reference signal received power of grid b; B is the number of grids into which the coverage area of ​​the base station is divided; 2 represents the 2-norm.

[0148] In steps S620 to S630 of some embodiments, in the radio map extrapolation task, the first loss cannot capture the correlation of RSRP values ​​between grids, so a physical prior penalty is designed to describe this correlation, so as to use the penalty as a regularization term to improve the generalization ability of the model. The distance between each two grids is calculated according to the grid position coordinates of each two grids to obtain the grid distance. For each grid, the grid with the smallest grid distance with the grid is selected as the reference grid. If the coverage area of ​​the base station includes B grids, the reference grid corresponding to grid b is the grid with the smallest grid distance with grid b among the remaining B-1 grids except grid b.

[0149] In step S640 of some embodiments, based on the physical a priori penalty term, the power difference value between the predicted reference signal received power of each grid and the predicted reference signal received power of the corresponding reference grid is calculated to capture the correlation of the RSRP values ​​between the grids, and the power difference value of each grid is added to obtain the second loss. The second loss can ensure that the RSRP changes of adjacent grids are relatively smooth, and the second loss is defined as:

[0150] ,

[0151] in, For the second loss; represents the predicted reference signal received power of grid b; is the reference grid corresponding to grid b The predicted reference signal received power.

[0152] In step S650 of some embodiments, a loss weight of the second loss is set, the second loss and the loss weight are multiplied, and the multiplication result is added to the first loss to obtain the target loss of the radio map extrapolation task. The loss weight can be set according to actual conditions, such as 0.01.

[0153] Through the above steps S610 to S650, the target loss can be obtained to guide the model training process in a supervised learning manner based on the target loss.

[0154] In step S170 of some embodiments, the target loss is minimized, and the three-dimensional spatial position coordinates, opacity, three-dimensional covariance matrix (rotation matrix and scaling matrix), projection coefficient, model parameters of the reference signal received power space model, and model parameters of the explicit base station coordinate embedding module of each three-dimensional Gaussian splash model are updated to obtain a radio map reconstruction model. By fusing the measurement data of the wireless signal with the three-dimensional environmental point cloud data, the problems of complex wireless signal propagation and data shortage are solved, and the accuracy of radio map reconstruction is improved.

[0155] Figure 7 is a flowchart of a radio map reconstruction method provided in an embodiment of the present application. The radio map reconstruction method is applied to a network layer data processing module of a target base station, and may include but is not limited to steps S710 to S740:

[0156] Step S710, obtaining the target base station coordinates of the target base station;

[0157] Step S720, dividing the coverage area of ​​the target base station into a plurality of target grids, and obtaining the target position coordinates of each target grid;

[0158] Step S730, predicting the target reference signal received power of each target grid based on the target base station coordinates and the target position coordinates of each target grid through a radio map reconstruction model; wherein the radio map reconstruction model is trained according to the above-mentioned model training method;

[0159] Step S740 , reconstructing a radio map according to each target reference signal received power to obtain a target radio map.

[0160] In step S710 of some embodiments, target base station coordinates of a target base station are acquired, where the target base station is a base station for which radio map reconstruction is to be performed, and the target base station coordinates are three-dimensional base station coordinates of the target base station.

[0161] In step S720 of some embodiments, the coverage area of ​​the target base station is divided into a plurality of target grids, and the three-dimensional space coordinates of each target grid are acquired to obtain the target position coordinates.

[0162] In step S730 of some embodiments, the radio map reconstruction model includes N three-dimensional Gaussian splash models, a reference signal received power spatial model, and an explicit base station coordinate embedding module. For each target grid, each three-dimensional Gaussian splash model is projected onto the reference signal received power spatial model based on the target position coordinates of the target grid to obtain the projection coefficient of each three-dimensional Gaussian splash model. Through the explicit base station coordinate embedding module, the base station coordinate embedding vector of each three-dimensional Gaussian splash model is extracted from the difference vector between the three-dimensional spatial position coordinates of each three-dimensional Gaussian splash model and the target base station coordinates. Each three-dimensional Gaussian splash model is rendered according to the projection coefficient of each three-dimensional Gaussian splash model, the target position coordinates of the target grid, and the corresponding base station coordinate embedding vector to obtain the target reference signal received power of the target grid.

[0163] In step S740 of some embodiments, a radio map is reconstructed according to the target position coordinates of the target grid and the target reference signal received power of the target grid at each time to obtain a target radio map.

[0164] Through the above steps S710 to S740, a target radio map can be obtained, so as to provide strong technical support for various ultra-wideband services and high-precision sensing requirements of the new generation wireless communication system based on the target radio map.

[0165] See also Figure 8 The embodiment of the present application further provides a radio map reconstruction device, which can implement the above radio map reconstruction method, and the radio map reconstruction device includes:

[0166] An acquisition module 810 is used to acquire the target base station coordinates of the target base station;

[0167] A division module 820 is used to divide the coverage area of ​​the target base station into a plurality of target grids and obtain the target position coordinates of each target grid;

[0168] A prediction module 830 is used to predict the target reference signal received power of each target grid based on the target base station coordinates and the target position coordinates of each target grid through a radio map reconstruction model; wherein the radio map reconstruction model is trained according to the above-mentioned model training method;

[0169] The reconstruction module 840 is configured to reconstruct the radio map according to each target reference signal received power to obtain a target radio map.

[0170] The specific implementation of the radio map reconstruction device is substantially the same as the specific implementation of the radio map reconstruction method described above, and will not be described in detail herein.

[0171] The present application also provides a model training device, which can implement the above-mentioned model training method, and the model training device includes:

[0172] A point cloud acquisition module is used to acquire a three-dimensional point cloud of the coverage area of ​​the base station and initialize multiple three-dimensional Gaussian splash models according to the three-dimensional point cloud;

[0173] An area division module is used to divide the coverage area into a plurality of grids and obtain the grid position coordinates and the reference signal receiving power of each grid;

[0174] A projection module, for projecting each three-dimensional Gaussian spreading model onto a reference signal received power space model based on the grid position coordinates for each grid, to obtain a projection coefficient of each three-dimensional Gaussian spreading model;

[0175] A determination module, configured to determine a base station coordinate embedding vector of each three-dimensional Gaussian splash model according to the base station coordinates of the base station;

[0176] A rendering module, used to render each three-dimensional Gaussian spreading model according to the projection coefficient of each three-dimensional Gaussian spreading model, the grid position coordinates and the corresponding base station coordinate embedding vector to obtain the predicted reference signal received power of the grid;

[0177] A calculation module, used to calculate the loss according to the predicted reference signal received power of each grid and the corresponding reference signal received power to obtain a target loss;

[0178] The updating module is used to update each three-dimensional Gaussian spreading model, projection coefficient and reference signal receiving power space model according to the target loss to obtain a radio map reconstruction model.

[0179] The specific implementation of the model training device is basically the same as the specific implementation of the above-mentioned model training method, and will not be repeated here.

[0180] In some embodiments, the point cloud acquisition module is further used to:

[0181] Randomly select multiple target points from the 3D point cloud;

[0182] Initialize multiple 3D Gaussian splash models based on multiple target points.

[0183] In some embodiments, the three-dimensional Gaussian spreading model has a three-dimensional covariance matrix and three-dimensional spatial position coordinates, the reference signal received power spatial model includes a first multi-layer perceptron and a second multi-layer perceptron, and the projection module is further used to:

[0184] Based on the grid position coordinates, projecting the three-dimensional covariance matrix of each three-dimensional Gaussian splash model through the first multi-layer perceptron to obtain the projection covariance feature of each three-dimensional Gaussian splash model;

[0185] Projecting the three-dimensional spatial position coordinates of each three-dimensional Gaussian splash model through a second multi-layer perceptron to obtain projection coordinate features of each three-dimensional Gaussian splash model;

[0186] The projection coefficient of each three-dimensional Gaussian splatter model is determined according to the projection covariance characteristics of each three-dimensional Gaussian splatter model, the corresponding projection coordinate characteristics, the preset continuous tensor and the preset projection scale.

[0187] In some embodiments, the three-dimensional Gaussian splash model has three-dimensional spatial position coordinates, and the determination module is further used to:

[0188] Determine the gap vector of each three-dimensional Gaussian splash model according to the base station coordinates and the three-dimensional spatial position coordinates of each three-dimensional Gaussian splash model;

[0189] Position encoding is performed on the gap vector of each three-dimensional Gaussian splash model to obtain an encoding vector of each three-dimensional Gaussian splash model;

[0190] Perform vector projection on the encoding vector of each three-dimensional Gaussian splash model to obtain the base station coordinate embedding vector of each three-dimensional Gaussian splash model.

[0191] In some embodiments, the three-dimensional Gaussian splash model has three-dimensional spatial position coordinates, opacity, and spherical harmonics, and the rendering module is further used to:

[0192] Projecting the opacity of each three-dimensional Gaussian splash model according to the projection coefficient of each three-dimensional Gaussian splash model and the corresponding base station coordinate embedding vector to obtain the projected opacity of each three-dimensional Gaussian splash model;

[0193] By performing function calculation on the corresponding projection coefficient, grid position coordinates and three-dimensional space position coordinates of each three-dimensional Gaussian splash model through the spherical harmonic function, the expected reference signal receiving power of each three-dimensional Gaussian splash model is obtained;

[0194] The projection opacity of each three-dimensional Gaussian splash model is screened to obtain a reference opacity;

[0195] Rendering the expected reference signal received power of each three-dimensional Gaussian splash model according to the projection opacity and the reference opacity of each three-dimensional Gaussian splash model to obtain a plurality of rendering powers;

[0196] The power of each rendering is summed to obtain the predicted reference signal received power of the grid.

[0197] In some embodiments, the computing module is further configured to:

[0198] Perform loss calculation according to the predicted reference signal received power of each grid and the corresponding reference signal received power to obtain a first loss;

[0199] Calculate the grid distance between any two grids;

[0200] For each grid, the grid with the smallest grid distance between grids is selected as the reference grid;

[0201] Perform loss calculation according to the predicted reference signal received power of each grid and the predicted reference signal received power of the corresponding reference grid to obtain a second loss;

[0202] The first loss and the second loss are summed to get the target loss.

[0203] The embodiment of the present application also provides an electronic device, the electronic device includes a memory and a processor, the memory stores a computer program, and the processor implements the above-mentioned communication and memory efficient large model distributed training method or text classification method when executing the computer program. The electronic device can be any intelligent terminal including a tablet computer, a car computer, etc.

[0204] See also Fig. 9 , Fig. 9 The hardware structure of an electronic device of another embodiment is illustrated, and the electronic device includes:

[0205] The processor 910 may be implemented by a general-purpose central processing unit (CPU), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of the present application;

[0206] The memory 920 can be implemented in the form of a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 920 can store an operating system and other applications. When the technical solution provided in the embodiments of this specification is implemented by software or firmware, the relevant program code is stored in the memory 920, and the processor 910 calls and executes the communication and memory efficient large model distributed training method or text classification method of the embodiments of this application;

[0207] Input / output interface 930, used to implement information input and output;

[0208] Communication interface 940, used to realize communication interaction between the device and other devices, which can be realized through wired mode (such as USB, network cable, etc.) or wireless mode (such as mobile network, WIFI, Bluetooth, etc.);

[0209] bus 950 , which transmits information between various components of the device (e.g., processor 910 , memory 920 , input / output interface 930 , and communication interface 940 );

[0210] The processor 910 , the memory 920 , the input / output interface 930 , and the communication interface 940 are connected to each other in communication within the device via a bus 950 .

[0211] An embodiment of the present application also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the above-mentioned communication and memory-efficient large-model distributed training method or text classification method.

[0212] The memory, as a non-transient computer-readable storage medium, can be used to store non-transient software programs and non-transient computer executable programs. In addition, the memory may include a high-speed random access memory, and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some embodiments, the memory may optionally include a memory remotely disposed relative to the processor, and these remote memories may be connected to the processor via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0213] The embodiments described in the embodiments of the present application are intended to more clearly illustrate the technical solutions of the embodiments of the present application and do not constitute a limitation on the technical solutions provided in the embodiments of the present application. Those skilled in the art will appreciate that with the evolution of technology and the emergence of new application scenarios, the technical solutions provided in the embodiments of the present application are also applicable to similar technical problems.

[0214] Those skilled in the art will appreciate that the technical solutions shown in the figures do not constitute a limitation on the embodiments of the present application, and may include more or fewer steps than shown in the figures, or a combination of certain steps, or different steps.

[0215] The device embodiments described above are merely illustrative, and the units described as separate components may or may not be physically separated, that is, they may be located in one place or distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0216] Those skilled in the art will appreciate that all or some of the steps in the methods disclosed above, and the functional modules / units in the systems and devices may be implemented as software, firmware, hardware, or a suitable combination thereof.

[0217] The terms "first", "second", "third", "fourth", etc. (if any) in the specification of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0218] It should be understood that in the present application, "at least one (item)" means one or more, and "plurality" means two or more. "And / or" is used to describe the association relationship of associated objects, indicating that three relationships may exist. For example, "A and / or B" can mean: only A exists, only B exists, and A and B exist at the same time, where A and B can be singular or plural. The character " / " generally indicates that the objects associated before and after are in an "or" relationship. "At least one of the following" or similar expressions refers to any combination of these items, including any combination of single or plural items. For example, at least one of a, b or c can mean: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, c can be single or multiple.

[0219] In the several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the above units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0220] The units described above as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0221] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0222] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including multiple instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of various embodiments of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory, referred to as ROM), random access memory (Random Access Memory, referred to as RAM), disk or optical disk and other media that can store programs.

[0223] The preferred embodiments of the present invention are described above with reference to the accompanying drawings, but the scope of the rights of the present invention is not limited thereto. Any modification, equivalent substitution and improvement made by a person skilled in the art without departing from the scope and essence of the present invention should be within the scope of the rights of the present invention.

Claims

1. A model training method, characterized in that: The method comprises: Acquire a three-dimensional point cloud of the coverage area of ​​the base station, and initialize multiple three-dimensional Gaussian splash models according to the three-dimensional point cloud; Dividing the coverage area into a plurality of grids, and obtaining grid position coordinates and a reference signal receiving power of each grid; For each grid, projecting each three-dimensional Gaussian spreading model onto a reference signal received power spatial model based on the grid position coordinates to obtain a projection coefficient of each three-dimensional Gaussian spreading model; Determine a base station coordinate embedding vector of each three-dimensional Gaussian splash model according to the base station coordinates of the base station; Rendering each three-dimensional Gaussian spreading model according to the projection coefficient of each three-dimensional Gaussian spreading model, the grid position coordinates and the corresponding base station coordinate embedding vector to obtain the predicted reference signal received power of the grid; Perform loss calculation based on the predicted reference signal received power of each grid and the corresponding reference signal received power to obtain a target loss; Each three-dimensional Gaussian spreading model, the projection coefficient, and the reference signal received power spatial model are updated according to the target loss to obtain a radio map reconstruction model.

2. The method according to claim 1, characterized in that The three-dimensional Gaussian splash model has a three-dimensional covariance matrix and a three-dimensional spatial position coordinate, the reference signal received power spatial model includes a first multi-layer perceptron and a second multi-layer perceptron, and each three-dimensional Gaussian splash model is projected onto the reference signal received power spatial model based on the grid position coordinate to obtain the projection coefficient of each three-dimensional Gaussian splash model, including: Based on the grid position coordinates, projecting the three-dimensional covariance matrix of each three-dimensional Gaussian spreading model through the first multi-layer perceptron to obtain the projection covariance feature of each three-dimensional Gaussian spreading model; Projecting the three-dimensional spatial position coordinates of each three-dimensional Gaussian splash model through the second multi-layer perceptron to obtain projection coordinate features of each three-dimensional Gaussian splash model; The projection coefficient of each three-dimensional Gaussian spreading model is determined according to the projection covariance feature of each three-dimensional Gaussian spreading model, the corresponding projection coordinate feature, the preset continuous tensor and the preset projection scale.

3. The method according to claim 1, characterized in that The three-dimensional Gaussian splash model has three-dimensional spatial position coordinates, and determining the base station coordinate embedding vector of each three-dimensional Gaussian splash model according to the base station coordinates of the base station includes: Determine a gap vector of each three-dimensional Gaussian splash model according to the base station coordinates and the three-dimensional spatial position coordinates of each three-dimensional Gaussian splash model; Position encoding is performed on the gap vector of each three-dimensional Gaussian splash model to obtain an encoding vector of each three-dimensional Gaussian splash model; Perform vector projection on the encoding vector of each three-dimensional Gaussian splash model to obtain the base station coordinate embedding vector of each three-dimensional Gaussian splash model.

4. The method according to claim 1, characterized in that: The three-dimensional Gaussian splash model has three-dimensional spatial position coordinates, opacity and spherical harmonic functions, and each three-dimensional Gaussian splash model is rendered according to the projection coefficient of each three-dimensional Gaussian splash model, the grid position coordinates and the corresponding base station coordinate embedding vector to obtain the predicted reference signal received power of the grid, including: Projecting the opacity of each three-dimensional Gaussian splash model according to the projection coefficient of each three-dimensional Gaussian splash model and the corresponding base station coordinate embedding vector to obtain the projected opacity of each three-dimensional Gaussian splash model; Performing function calculation on the corresponding projection coefficient, the grid position coordinates and the three-dimensional space position coordinates through the spherical harmonic function of each three-dimensional Gaussian spreading model to obtain the expected reference signal received power of each three-dimensional Gaussian spreading model; The projection opacity of each three-dimensional Gaussian splash model is screened to obtain a reference opacity; Rendering the expected reference signal received power of each three-dimensional Gaussian splash model according to the projection opacity of each three-dimensional Gaussian splash model and the reference opacity to obtain multiple rendering powers; Each rendering power is summed to obtain the predicted reference signal received power of the grid.

5. The method according to claim 1, characterized in that The performing loss calculation according to the predicted reference signal received power of each grid and the corresponding reference reference signal received power to obtain a target loss includes: Perform loss calculation according to the predicted reference signal received power of each grid and the corresponding reference reference signal received power to obtain a first loss; Calculate the grid distance between any two grids; For each grid, select the grid with the smallest grid distance from the grid as the reference grid; Perform loss calculation according to the predicted reference signal received power of each grid and the predicted reference signal received power of the corresponding reference grid to obtain a second loss; The first loss and the second loss are summed to obtain the target loss.

6. The method according to any one of claims 1 to 5, characterized in that: Initializing a plurality of three-dimensional Gaussian splash models according to the three-dimensional point cloud comprises: Randomly selecting a plurality of target points from the three-dimensional point cloud; Initialize multiple 3D Gaussian splash models based on multiple target points.

7. A radio map reconstruction method, characterized in that: The method comprises: Obtain target base station coordinates of the target base station; Dividing the coverage area of ​​the target base station into a plurality of target grids, and obtaining the target position coordinates of each target grid; Predicting the target reference signal received power of each target grid based on the target base station coordinates and the target position coordinates of each target grid by a radio map reconstruction model; wherein the radio map reconstruction model is trained by the model training method according to any one of claims 1 to 6; The radio map is reconstructed according to each target reference signal received power to obtain a target radio map.

8. A radio map reconstruction device, characterized in that: The device comprises: An acquisition module, used to acquire the target base station coordinates of the target base station; A division module, used to divide the coverage area of ​​the target base station into a plurality of target grids, and obtain the target position coordinates of each target grid; A prediction module, configured to predict the target reference signal received power of each target grid based on the target base station coordinates and the target position coordinates of each target grid through a radio map reconstruction model; wherein the radio map reconstruction model is trained by the model training method according to any one of claims 1 to 6; The reconstruction module is used to reconstruct the radio map according to each target reference signal received power to obtain a target radio map.

9. An electronic device, characterized in that: The electronic device includes a memory and a processor, the memory stores a computer program, and the processor implements the model training method according to any one of claims 1 to 6 or the radio map reconstruction method according to claim 7 when executing the computer program.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the model training method according to any one of claims 1 to 6 or the radio map reconstruction method according to claim 7 is implemented.

Images