A radio path loss dynamic prediction method, system, device and medium

Abstract

This application relates to a method, system, device, and medium for dynamic prediction of radio path loss. The method includes: performing three-dimensional scene analysis based on a stereo image sequence and a motion pose sequence to obtain a three-dimensional scene sequence; predicting the motion trajectory of each dynamic object based on the three-dimensional scene sequence to obtain a predicted three-dimensional state set; obtaining a future virtual three-dimensional scene based on the three-dimensional scene sequence and the predicted three-dimensional state set, and extracting three-dimensional ray projection features based on the future virtual three-dimensional scene to obtain a deterministic loss value and a wireless propagation visual feature vector; inputting the wireless propagation visual feature vector into a cross-modal neural network model for electromagnetic attenuation mapping to obtain a predicted electromagnetic fading factor; and calculating a predicted path loss value based on the predicted electromagnetic fading factor and the deterministic loss value. This method can capture random losses caused by dynamic obstructions, improving prediction accuracy.

Description

Technical Field

[0001] This invention belongs to the field of intelligent prediction, and in particular relates to a method, system, device and medium for dynamic prediction of radio path loss. Background Technology

[0002] As wireless communication technology continues to evolve towards higher frequencies, higher densities, and lower latency, unprecedented demands are being placed on the ability to accurately perceive and predict wireless channel quality. Against this backdrop, channel prediction technology has emerged, aiming to predict future channel states through prior information, thereby providing a crucial basis for adaptive resource scheduling. This technology is characterized by its forward-looking and proactive nature, effectively addressing link instability issues in high-speed mobile scenarios. Traditional techniques for path loss prediction in complex urban environments primarily rely on two types of methods: one is deterministic modeling based on ray tracing, which simulates radio wave propagation through precise 3D maps and electromagnetic calculations; the other is statistical modeling based on a large amount of measured data, using empirical formulas or historical distributions to estimate loss. The former attempts to perform accurate simulations based on physical principles, while the latter derives statistical patterns from historical data. However, current traditional methods have significant problems when dealing with dynamic and complex real-world urban environments. Ray tracing methods heavily rely on high-precision 3D digital maps, have high computational complexity, and struggle to simulate the random occlusion effects of dynamic objects such as vehicles and pedestrians in real time, causing predictions to lag behind environmental changes. While statistical models are computationally simple, they rely too heavily on measurement data from specific scenarios, have poor generalization ability, and cannot accurately reflect instantaneous depth fading caused by specific dynamic objects. Neither approach effectively addresses the core challenge of random loss caused by dynamic obstructions, making it difficult to meet the accurate prediction requirements of high-reliability communication. Summary of the Invention

[0003] Therefore, it is necessary to provide a method, system, device, and medium for dynamic prediction of radio path loss that can simulate the random occlusion effect of dynamic objects and ensure prediction accuracy, in order to address the above-mentioned technical problems.

[0004] In a first aspect, this application provides a method for dynamic prediction of radio path loss, including:

[0005] Acquire stereo image sequences and motion pose sequences, and perform 3D scene analysis based on stereo image sequences and motion pose sequences to obtain 3D scene sequences;

[0006] Based on the 3D scene sequence, the motion trajectory of each dynamic object is predicted to obtain a set of predicted 3D states.

[0007] Based on the three-dimensional scene sequence and the predicted three-dimensional state set, a future virtual three-dimensional scene is constructed. Based on the future virtual three-dimensional scene, three-dimensional ray projection feature extraction is performed to obtain deterministic loss value and wireless propagation visual feature vector.

[0008] The wireless propagation visual feature vector is input into a cross-modal neural network model to perform electromagnetic attenuation mapping and obtain the predicted electromagnetic fading factor.

[0009] Based on the predicted electromagnetic fading factor and the deterministic loss value, the predicted path loss value is calculated; the predicted path loss value is used to characterize the predicted loss of the radio path between communication nodes.

[0010] Furthermore, the cross-modal neural network model was obtained through the following method:

[0011] Based on the original time-series data stream, 3D scene analysis is performed to obtain dense 3D scene and object dynamic history trajectory. Based on the dense 3D scene and object dynamic history trajectory, a standard virtual 3D scene is constructed.

[0012] Three-dimensional ray projection feature extraction is performed on a standard virtual three-dimensional scene to obtain standard deterministic loss value and standard visual feature vector;

[0013] Based on the channel impulse response sequence and standard deterministic loss value, the path loss value is calculated, and the standard visual feature vector and the path loss value are paired to obtain feature loss paired samples.

[0014] Using standard visual feature vectors as input features and path loss values ​​as labels, we integrate feature loss paired samples to obtain a standardized dataset.

[0015] Based on a standardized dataset, the cross-modal neural network model framework is iteratively trained to obtain the cross-modal neural network model.

[0016] Furthermore, based on the channel impulse response sequence and the standard deterministic loss value, the path loss value is calculated, including:

[0017] Based on the channel impulse response sequence, the total received power is calculated, and based on the device calibration parameters, the total received power is calibrated to obtain the received power value.

[0018] Calculate the difference between the received power value and the transmitted power value to obtain the measured value of the path loss;

[0019] Based on a dense 3D scene, the free space propagation distance of a standard radio path is extracted, and the free space fundamental loss value is calculated based on the free space propagation distance.

[0020] The free space base loss value is calculated using the following formula:

[0021]

[0022] in, d is the free space fundamental loss value, f is the free space propagation distance, c is the frequency of the radio wave, and c is the speed of light in a vacuum.

[0023] The path loss value is calculated based on the measured path loss value, the free space basic loss value, and the standard deterministic loss value.

[0024] Furthermore, the formula for calculating the path loss value is as follows:

[0025]

[0026] in, This is the path loss value. This is the measured value of path loss. This represents the basic loss value in free space. This is the standard deterministic loss value.

[0027] Furthermore, based on the predicted electromagnetic fading factor and the deterministic loss value, the predicted path loss value is calculated, including:

[0028] Extract the predicted 3D coordinates of communication nodes from a future virtual 3D scene, and calculate the predicted straight-line distance between communication nodes based on the predicted 3D coordinates;

[0029] Based on the predicted straight-line distance and communication parameters, the predicted free-space basis loss value of the radio path between the corresponding communication nodes is calculated.

[0030] The intermediate predicted loss value is obtained by algebraically adding the deterministic loss value and the predicted free space basic loss value.

[0031] The predicted electromagnetic fading factor and the intermediate predicted loss value are algebraically added together to obtain the complete predicted loss value.

[0032] Based on the real-time environmental calibration factor, the complete predicted loss value is adjusted to obtain the environmentally adjusted loss value. The system margin parameter is then superimposed on the environmentally adjusted loss value to obtain the predicted path loss value.

[0033] Furthermore, based on the future virtual 3D scene, 3D ray projection feature extraction is performed to obtain deterministic loss values ​​and wireless propagation visual feature vectors, including:

[0034] Based on a future virtual 3D scene, simulated detection ray emission is performed between communication nodes to obtain the detection ray. The interaction state between the detection ray and all geometric objects is analyzed to obtain the main propagation path interaction report.

[0035] Based on the main propagation path interaction report, the type of geometry is analyzed to obtain regular geometry and irregular geometry;

[0036] Based on the main propagation path interaction report, the blade-shaped diffraction parameters of the regular geometry are extracted, and the deterministic loss value is calculated based on the blade-shaped diffraction parameters.

[0037] Based on the main propagation path interaction report, feature extraction is performed on the irregular geometry to obtain the wireless propagation visual feature vector.

[0038] Secondly, this application also provides a dynamic prediction system for radio path loss, comprising:

[0039] The modeling module is used to acquire stereo image sequences and motion pose sequences, and based on the stereo image sequences and motion pose sequences, to perform 3D scene analysis to obtain 3D scene sequences.

[0040] The trajectory module is used to predict the motion trajectory of each dynamic object based on a 3D scene sequence, and obtain a set of predicted 3D states.

[0041] The feature module is used to construct a future virtual 3D scene based on the 3D scene sequence and the predicted 3D state set, and to extract 3D ray projection features based on the future virtual 3D scene to obtain deterministic loss values ​​and wireless propagation visual feature vectors.

[0042] The prediction module is used to input the wireless propagation visual feature vector into the cross-modal neural network model, perform electromagnetic attenuation mapping, and obtain the predicted electromagnetic fading factor.

[0043] The loss module is used to calculate the predicted path loss value based on the predicted electromagnetic fading factor and the deterministic loss value; the predicted path loss value is used to characterize the predicted loss of the radio path between communication nodes.

[0044] Thirdly, this application also provides a computer device including a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement any step of the method provided in the first aspect of this application.

[0045] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any step of the method provided in the first aspect of this application.

[0046] The aforementioned method, system, device, and medium for dynamic prediction of radio path loss acquire a sequence of stereo images and a sequence of moving poses. Based on these sequences, a 3D scene is analyzed to obtain a 3D scene sequence. Based on this 3D scene sequence, the trajectory of each dynamic object is predicted to obtain a set of predicted 3D states. A future virtual 3D scene is constructed based on the 3D scene sequence and the set of predicted 3D states. Based on this virtual 3D scene, 3D ray projection features are extracted to obtain a deterministic loss value and a wireless propagation visual feature vector. This wireless propagation visual feature vector is input into a cross-modal neural network model for electromagnetic attenuation mapping to obtain a predicted electromagnetic fading factor. Based on the predicted electromagnetic fading factor and the deterministic loss value, a predicted path loss value is calculated. This predicted path loss value characterizes the predicted loss of the radio path between communication nodes. It incorporates a visual modality to perceive the dynamic environment in real time, providing direct evidence for understanding the instantaneous causes of signal obstruction and reflection. A hybrid modeling framework is adopted, introducing a classical physical model to handle losses caused by regular obstacles and using a neural network to handle losses caused by irregular obstacles, reducing the model learning difficulty and improving the model's generalization ability. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 This is a schematic diagram of the process of a dynamic prediction method for radio path loss provided in an embodiment of the present invention;

[0049] Figure 2 This is a schematic diagram of the structure of a dynamic prediction system for radio path loss provided in an embodiment of the present invention. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0051] In one embodiment, such as Figure 1As shown, a method for dynamic prediction of radio path loss is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, and to a system including both a terminal and a server, and implemented through interaction between the terminal and the server. The communication node includes a transmitting node and a receiving node, which can be a mobile platform equipped with stereo vision equipment, such as a vehicle, drone, or other mobile device, or it can be without stereo vision equipment. This embodiment uses a mobile platform equipped with stereo vision equipment and the communication node as examples. In this embodiment, the method includes the following steps:

[0052] Step 101: Obtain the stereo image sequence and the motion pose sequence, and perform 3D scene analysis based on the stereo image sequence and the motion pose sequence to obtain the 3D scene sequence.

[0053] Stereo image sequence refers to a time series of paired images acquired by a stereo vision device, such as a binocular camera, at multiple consecutive time points. Each pair of images contains scenes captured at the same moment from two slightly different perspectives, used to perceive 3D information. Motion pose sequence refers to a time series of the position and orientation of a mobile platform equipped with a stereo vision device in 3D space at each corresponding time point in the stereo image sequence acquisition. The mobile platform can be a drone or a vehicle. Pose is usually represented by 3D coordinates and rotation angles. 3D scene analysis refers to the process of recovering and constructing the 3D environment structure from 2D images using visual geometry principles and computer vision algorithms. A 3D scene sequence refers to a series of 3D environment models arranged in chronological order after analyzing the stereo images at each time point. Each model describes the 3D geometric structure of the static background in the scene at that moment.

[0054] The terminal acquires paired images continuously over time using sensors and records its own position and orientation data at the time of image capture. The terminal performs 3D scene analysis, combining the parallax information between the stereo image pairs (i.e., the pixel position difference of the same object in two images) and the known camera movement trajectory. This is the motion pose sequence. Depth information for each frame is calculated using a stereo matching algorithm. Then, using the same algorithm, the sequence depth and pose data are fused. Using visual inertial odometry (VIO), a globally consistent dense 3D point cloud map is constructed and maintained in real time, resulting in a set arranged by timestamps—the 3D scene sequence. Essentially, the 3D scene sequence is a collection of snapshots showing the changes in the 3D structure of the environment over time.

[0055] Step 102: Based on the 3D scene sequence, predict the motion trajectory of each dynamic object to obtain a predicted 3D state set.

[0056] Specifically, dynamic objects refer to objects in a 3D scene sequence whose position or shape changes over time, such as pedestrians and vehicles. Motion trajectory prediction refers to inferring the object's future path and state based on its past and current motion states. The predicted 3D state set refers to the set of predicted states for each identified dynamic object at a series of future points in time. Each state typically includes information such as the object's predicted 3D position, velocity, and orientation at a future point in time.

[0057] The terminal identifies and segments dynamic objects and static background from a 3D scene sequence, predicting the trajectory of each dynamic object. By analyzing the positional changes of each dynamic object in the historical 3D scene sequence, the terminal uses motion models and machine learning prediction algorithms to infer the most likely path the object will take in the future and its state at each moment. After predicting all dynamic objects, the terminal aggregates the trajectory information to obtain a predicted 3D state set containing the predicted future states of all objects.

[0058] Step 103: Based on the three-dimensional scene sequence and the predicted three-dimensional state set, a future virtual three-dimensional scene is constructed, and based on the future virtual three-dimensional scene, three-dimensional ray projection feature extraction is performed to obtain deterministic loss value and wireless propagation visual feature vector.

[0059] Specifically, a future virtual 3D scene refers to a synthetic 3D model constructed by integrating a static 3D model of the environment with the predicted future states of dynamic objects, used to represent the overall environmental state at a future moment. 3D ray casting refers to a computer graphics technique that projects a virtual ray from one point to another in 3D space and calculates the interaction between the ray and the surfaces of all objects in the scene. Deterministic loss refers to the partial signal power attenuation caused by known and deterministic geometric occlusion and reflection paths in wireless propagation. Deterministic loss can be directly calculated using geometric optics principles. Wireless propagation visual feature vectors refer to high-dimensional numerical vectors extracted by analyzing and quantifying the 3D ray casting process, representing the visual and geometric characteristics of the radio electromagnetic wave propagation path. These vectors can include encoding information such as the type of material the ray passes through, the angle of incidence, the path length, and the number of reflections.

[0060] The terminal combines the static environment represented by the 3D scene sequence with the future positions of dynamic objects represented by the predicted 3D state set to form a complete, predicted 3D digital scene for the future. Within this virtual scene, the terminal extracts 3D ray projection features, simulating the propagation of wireless signals from the transmitting node to the receiving node. It emits numerous virtual rays between the two points, tracking the collisions and reflections of each ray with objects in the scene. From the ray tracing results, the terminal calculates and extracts deterministic loss values—the base loss calculated based on the ray's determined propagation path—and wireless propagation visual feature vectors, which encode complex ray interaction patterns to form mathematical vectors rich in visual geometric information.

[0061] Step 104: Input the wireless propagation visual feature vector into the cross-modal neural network model to perform electromagnetic attenuation mapping and obtain the predicted electromagnetic fading factor.

[0062] The cross-modal neural network model refers to a trained deep learning model specifically designed to process and correlate data from different modalities or domains. In this embodiment, it involves visual geometric feature modality and electromagnetic propagation modality, learning the complex mapping relationship between them. Electromagnetic attenuation mapping refers to the process of converting feature information representing the visual geometric environment into corresponding electromagnetic wave attenuation characteristics. The predicted electromagnetic fading factor refers to the portion of signal attenuation predicted by the model that cannot be captured by deterministic geometric calculations. It mainly characterizes the random or difficult-to-model fading caused by complex scattering, diffraction, and subtle differences in the electromagnetic properties of materials.

[0063] The terminal uses the wireless propagation visual feature vector as input data and feeds it into a cross-modal neural network model. This model has been pre-trained on a large amount of data, learning the complex correspondence between scene visual features and electromagnetic attenuation. Internally, the cross-modal neural network model performs electromagnetic attenuation mapping calculations, inferring the corresponding electromagnetic attenuation characteristics based on the input feature vector. After calculation, the model obtains and outputs a numerical value, namely the predicted electromagnetic fading factor. The electromagnetic fading factor supplements the radio path attenuation portion that was not covered by geometrically deterministic calculations.

[0064] Step 105: Based on the predicted electromagnetic fading factor and the deterministic loss value, the predicted path loss value is calculated; the predicted path loss value is used to characterize the predicted loss of the radio path between communication nodes.

[0065] Among them, the predicted path loss value refers to the predicted total power attenuation value of the radio signal along the entire path from the transmitting node to the receiving node, measured in decibels. It is a key parameter for network planning, link budget, etc.

[0066] The terminal calculates the predicted path loss value based on the predicted electromagnetic fading factor and the deterministic loss value. For example, the deterministic loss value caused by geometric obstruction is added to the predicted electromagnetic fading factor caused by complex propagation effects to obtain a complete prediction of the total path loss. The resulting predicted path loss value is used to quantitatively characterize the predicted loss of the radio path between communication nodes, providing a direct basis for the performance evaluation and optimization of wireless communication systems.

[0067] This embodiment provides a method for dynamic prediction of radio path loss. It acquires a stereo image sequence and a motion pose sequence, and performs 3D scene analysis based on these sequences to obtain a 3D scene sequence. Based on this 3D scene sequence, it predicts the motion trajectory of each dynamic object, obtaining a predicted 3D state set. Based on the 3D scene sequence and the predicted 3D state set, it constructs a future virtual 3D scene, and performs 3D ray projection feature extraction based on this virtual scene to obtain a deterministic loss value and a wireless propagation visual feature vector. The wireless propagation visual feature vector is input into a cross-modal neural network model for electromagnetic attenuation mapping to obtain a predicted electromagnetic fading factor. Based on the predicted electromagnetic fading factor and the deterministic loss value, it calculates the predicted path loss value. This predicted path loss value characterizes the predicted loss of the radio path between communication nodes. Through these methods, a visual modality can be introduced to perceive the dynamic environment in real time, providing direct evidence for understanding the instantaneous causes of signal obstruction and reflection. A hybrid modeling framework is adopted, introducing a classical physical model to handle losses caused by regular obstacles, and using a neural network to handle losses caused by irregular obstacles, reducing the model learning difficulty and improving the model's generalization ability.

[0068] In one embodiment, the cross-modal neural network model is obtained through the following method:

[0069] Step 201: Based on the original time-series data stream, perform 3D scene analysis to obtain a dense 3D scene and the dynamic historical trajectory of objects, and construct a standard virtual 3D scene based on the dense 3D scene and the dynamic historical trajectory of objects.

[0070] The original time-series data stream refers to the set of raw sensor data collected during the training phase and arranged in chronological order. It typically includes stereo image sequences acquired by stereo vision devices and pose sequences of the mobile platform synchronously acquired through devices such as inertial navigation systems. 3D scene resolution refers to the process of recovering and constructing the 3D environment structure from a 2D image sequence using computer vision algorithms, and identifying the motion of objects within it. A dense 3D scene refers to a high-precision, high-completeness static 3D environment model reconstructed through in-depth processing of the original time-series data stream. Dense 3D scenes contain richer and more accurate geometric and texture details than 3D scene sequences obtained through real-time resolution. Dynamic historical trajectory of objects refers to the set of actual motion paths of all moving objects in the scene over a past period, identified through analysis of the original time-series data stream. Each trajectory records the actual 3D position, velocity, and other information of an object at different points in time. A standard virtual 3D scene refers to a digital copy of a real scene specifically constructed for model training. It is formed by fusing a precise static background represented by a dense 3D scene with the actual position and state of an object at a specific historical moment, recorded in the dynamic historical trajectory of the object, as a known, labeled standard 3D environment.

[0071] The terminal acquires the raw time-series data stream for training, namely, stereo images containing timestamps and corresponding device pose data. The terminal performs 3D scene parsing. Utilizing more powerful offline computing capabilities, the entire data stream is processed to generate a detailed, globally consistent static environment model, i.e., a dense 3D scene. All moving objects in the data stream are detected and tracked, and their actual motion paths are recorded, thus obtaining the objects' dynamic historical trajectories. The terminal constructs a standard virtual 3D scene. At a specific historical moment, the dense 3D scene is fused with the actual positional states of all objects in the objects' dynamic historical trajectories at that moment, forming a complete and accurate 3D digital scene. This 3D digital scene will serve as the standard reference environment for subsequent ray tracing and feature extraction.

[0072] Step 202: Extract three-dimensional ray projection features from the standard virtual three-dimensional scene to obtain the standard deterministic loss value and the standard visual feature vector.

[0073] Specifically, 3D ray casting refers to a simulation technique that emits a virtual ray from a designated emission point to a receiving point in 3D space and precisely calculates the interaction path between the ray and the surface of the scene model. Standard deterministic loss value refers to the signal power attenuation value caused by a defined reflection and transmission path, obtained through geometrical optics calculations of 3D ray casting in a standard virtual 3D scene. It represents a portion of the loss that can be directly calculated using physical laws at a specific scene and node location. Standard visual feature vector refers to a high-dimensional numerical vector extracted from the 3D ray casting process of a standard virtual 3D scene to describe the visual and geometric characteristics of the ray propagation path. The standard visual feature vector encodes visually relevant information such as material properties and geometric interactions along the path.

[0074] In a standard virtual 3D scene, the terminal sets the positions of a pair of communication nodes. The terminal performs 3D ray projection feature extraction, runs a ray tracing algorithm, and simulates all major propagation paths of electromagnetic waves from the transmitting node to the receiving node. From the ray tracing results, the terminal calculates and extracts the standard deterministic loss value, which is the signal attenuation value calculated using physical formulas such as Fresnel's equations based on the determined ray path; and the standard visual feature vector, which is the feature vector obtained after quantizing and encoding the complex visual geometric information of the interaction between the ray and the scene.

[0075] Step 203: Calculate the path loss value based on the channel impulse response sequence and the standard deterministic loss value, and pair the standard visual feature vector and the path loss value to obtain the feature loss pairing sample.

[0076] Specifically, the channel impulse response sequence refers to the time-ordered channel impulse response data obtained during the training phase by deploying wireless devices in a real physical environment corresponding to a standard virtual 3D scene. It characterizes the multipath propagation characteristics of a real wireless channel. The path loss value refers to the signal attenuation value calculated from real measurement data, after removing free-space propagation loss and standard deterministic loss values. It is mainly caused by factors that are difficult to model deterministically, such as complex scattering and diffraction. It is the true label that the model needs to learn. A feature loss paired sample refers to a data pair used for training. It is composed of a standard visual feature vector extracted from the same scene and node configuration, and the corresponding path loss value calculated from real measurement data.

[0077] The terminal calculates the path loss value based on the channel impulse response sequence and standard deterministic loss values. It calculates and calibrates the total received power from the channel impulse response sequence to obtain the measured path loss value. It then calculates the free-space fundamental loss value. Finally, it subtracts the free-space fundamental loss value and the standard deterministic loss value from the measured path loss value to obtain the pure path loss value caused by complex factors. The terminal pairs standard visual feature vectors with the path loss values, associating the standard visual feature vectors describing the scene's visual geometry with the path loss values ​​derived from actual scene measurements that reflect complex electromagnetic attenuation, forming a feature-label pair. The terminal obtains feature-loss pairing samples, which establish a correspondence between the scene's visual appearance and the difficult-to-calculate electromagnetic fading.

[0078] Step 204: Using standard visual feature vectors as input features and path loss values ​​as labels, integrate feature loss paired samples to obtain a standardized dataset.

[0079] Here, input features refer to the input data received by the machine learning model during training. In this embodiment, it specifically refers to standard visual feature vectors. Labels refer to the true target value or correct answer corresponding to the input features in supervised machine learning. In this embodiment, it specifically refers to the calculated path loss value. Standardized datasets refer to a set of standardized data that can be directly used by the model after collecting a large number of feature loss paired samples and preprocessing them such as cleaning, normalization, and partitioning.

[0080] The terminal uses the standard visual feature vectors from the paired samples as the input features that the model needs to learn, and the path loss values ​​from the paired samples as the labels that the model needs to predict. It collects all paired samples generated under a large number of different scenarios and node location configurations, and performs operations such as outlier removal and numerical standardization on the paired samples, resulting in a standardized dataset. This standardized dataset contains all the training data needed to learn the mapping relationship between visual features and electromagnetic fading from diverse environments.

[0081] Step 205: Based on the standardized dataset, iteratively train the cross-modal neural network model framework to obtain the cross-modal neural network model.

[0082] A cross-modal neural network model framework refers to a specific deep learning network structure that has not yet been trained with data. This framework is designed to handle data from different modalities, namely the visual geometric modality and the electromagnetic propagation modality. Its internal structure is defined, but its parameters are randomly initialized. Iterative training refers to the repeated training process in machine learning. In each iteration, the model uses a batch of data to calculate predicted values, compares them with the true labels to obtain the error, and then adjusts the model's internal parameters based on the error to gradually reduce the prediction error. A cross-modal neural network model refers to a model whose internal parameters have been optimized to their optimal state after the cross-modal neural network model framework has been fully trained on a standardized dataset. It is a model capable of predicting electromagnetic decay factors from visual feature vectors.

[0083] The terminal iteratively trains a cross-modal neural network model framework based on a standardized dataset. Standard visual feature vectors from the dataset are input into the model framework, which outputs a predicted fading factor. The predicted value is compared with the corresponding true path loss value in the dataset to calculate the loss function. The internal parameters of the model framework are automatically adjusted based on the error using a backpropagation algorithm. This training process is repeated multiple times on all data in the standardized dataset. After a sufficient number of iterations, the parameters of the model framework gradually stabilize, and its predictive ability reaches its optimal level. At this point, the trained network with predictive capabilities is the cross-modal neural network model. The cross-modal neural network model encapsulates complex mapping knowledge from visual geometric features to electromagnetic attenuation characteristics.

[0084] In one embodiment, the path loss value is calculated based on the channel impulse response sequence and the standard deterministic loss value, including:

[0085] Step 301: Calculate the total received power based on the channel impulse response sequence, and perform power calibration on the total received power based on the device calibration parameters to obtain the received power value.

[0086] The channel impulse response sequence refers to the channel impulse response data measured at the receiver and arranged in chronological order. Each channel impulse response characterizes the temporal distribution of the signal energy after multipath propagation from the transmitter to the receiver, including the amplitude, phase, and delay information of all resolvable multipath components. Total received power refers to the total raw signal power captured at the receiving antenna, obtained directly from the channel impulse response sequence measurement data by summing the power of each multipath component. It is a preliminary calculation result without equipment characteristic correction. Equipment calibration parameters refer to the inherent system parameters determined before experimentation or deployment by individually measuring the wireless transceiver equipment used, including the RF front-end, amplifiers, filters, antennas, and cables. These parameters quantify the gain or loss introduced by the equipment itself, including additional losses in the transmit link, gain in the receive link, and antenna efficiency. Received power value refers to the actual wireless signal power received at the standard reference point, i.e., the output port of the receiving antenna, obtained after correcting the total received power using the equipment calibration parameters, eliminating the influence of the measurement equipment's own characteristics.

[0087] The terminal acquires the channel impulse response sequence measured in a real physical environment. Based on the channel impulse response sequence, the terminal calculates the total received power by processing one or more channel impulse responses. For example, for a single channel impulse response, the sum of the squares of the amplitudes of its resolvable multipath components is calculated and divided by the corresponding integration time or symbol period to obtain the instantaneous total received power corresponding to that measurement. For the channel impulse response sequence, the average power needs to be calculated. Based on the device calibration parameters, the terminal performs power calibration on the total received power. The calculated total received power is then inversely converted according to the overall gain of the receiver link recorded in the device calibration parameters, deducting or compensating for the influence introduced by the device itself. The calibrated received power value accurately reflects the signal power actually received from the space propagation channel at the standard antenna port.

[0088] Step 302: Calculate the difference between the received power value and the transmitted power value to obtain the measured path loss value.

[0089] Specifically, the transmit power value refers to the output power of the signal at the standard reference point of the transmitting equipment, i.e., at the input port of the transmitting antenna. It also needs to be calibrated according to the equipment's calibration parameters to eliminate the influence of inherent characteristics of the transmit link and reflect the actual net power fed into the transmitting antenna. The measured path loss value refers to the total power attenuation experienced by the signal during propagation from the transmitting antenna to the receiving antenna, calculated from actual measurement data. The value is equal to the difference between the transmit power value and the received power value, and includes the total loss caused by all propagation effects such as free-space diffusion, obstruction, reflection, scattering, and diffraction.

[0090] The terminal calculates the difference between the received power and the transmitted power. In decibels, path loss is simply the subtraction of transmitted and received power. Through this subtraction, the terminal obtains the measured path loss. The measured path loss is based on direct observations of total loss obtained from actual wireless measurements.

[0091] Step 303: Based on the dense 3D scene, extract the free space propagation distance of the standard radio path, and calculate the free space basic loss value based on the free space propagation distance.

[0092] The free space base loss value is calculated using the following formula:

[0093]

[0094] in, d is the free space fundamental loss value, f is the free space propagation distance, f is the frequency of the radio wave, and c is the speed of light in a vacuum.

[0095] Specifically, a dense 3D scene refers to a high-precision static 3D environment model obtained through offline 3D reconstruction during the model training phase, providing accurate geometric information about the scene. A standard radio path refers to a virtual, idealized straight-line propagation path determined within a dense 3D scene based on the actual deployment locations of the transmitting and receiving nodes during training data acquisition. Free-space propagation distance specifically refers to the 3D straight-line distance between the transmitting and receiving nodes along the standard radio path; it is the distance an electromagnetic wave needs to travel in free space without any obstacles. The free-space fundamental loss value refers to the basic power attenuation of an electromagnetic wave in a perfect vacuum or homogeneous medium, caused solely by wavefront diffusion with distance. It is determined by the free-space propagation distance and the radio wave frequency, and represents the theoretically lowest loss basis for any wireless propagation.

[0096] Based on the real node positions recorded in the training data, the terminal determines a straight line connecting the transmitting and receiving points—the standard radio path—within the coordinate system of a dense 3D scene. The terminal then extracts the free-space propagation distance of this standard radio path based on the dense 3D scene by calculating the Euclidean distance between the two endpoints of the path. Based on this free-space propagation distance, the terminal calculates the free-space fundamental loss value. Substituting the distance, the known radio wave frequency, and the speed of light in a vacuum into a given formula, the terminal calculates the free-space fundamental loss value, expressed in decibels (dB).

[0097] Step 304: Calculate the path loss value based on the measured path loss value, the free space basic loss value, and the standard deterministic loss value.

[0098] The standard deterministic loss value refers to the signal attenuation value caused by deterministic propagation mechanisms such as primary reflection and transmission, obtained during the model training phase through 3D ray tracing calculations on a standard virtual 3D scene. It characterizes the loss portion caused by large obstacles and major reflecting surfaces in the environment, which can be accurately modeled using geometric optics. The path loss value refers to the remaining signal attenuation value that cannot be explained by the deterministic model after deducting the free space fundamental loss value and the standard deterministic loss value from the measured path loss value. It mainly includes the loss caused by complex scattering, diffraction, and subtle differences in environmental material properties—random or difficult-to-model factors—and is the target that the cross-modal neural network model needs to learn and predict.

[0099] The terminal subtracts the theoretically inevitable free-space base loss from the total measured path loss, and then subtracts the standard deterministic loss that can be determined by ray tracing. The terminal obtains the path loss value. The physical meaning of the path loss value is the residual loss or refinement loss remaining after removing free-space diffusion loss and the main deterministic occlusion / reflection loss from the total measured loss, primarily caused by complex electromagnetic phenomena. The path loss value will be used as a label for supervised learning and paired with the corresponding visual feature vector.

[0100] In one embodiment, the path loss value is calculated using the following formula:

[0101]

[0102] in, This is the path loss value. This is the measured value of path loss. This represents the basic loss value in free space. This is the standard deterministic loss value.

[0103] Specifically, path loss refers to the signal attenuation value remaining after deducting free-space fundamental loss and standard deterministic loss from the actual measured total path loss. This attenuation is primarily caused by random or difficult-to-model electromagnetic propagation effects such as complex scattering, diffraction, and subtle differences in environmental material properties. It is a core target variable that cross-modal neural network models need to learn and predict. Measured path loss refers to the total path loss directly calculated from actual wireless measurement data. It is the difference between the calibrated transmit power and receive power, encompassing the sum of all types of losses experienced by the signal during propagation. Free-space fundamental loss refers to the basic theoretical loss value caused solely by wavefront diffusion as the propagation distance increases when electromagnetic waves propagate in ideal free space. It depends only on the straight-line distance between the transmitting and receiving nodes and the operating frequency of the radio waves. Standard deterministic loss refers to the signal attenuation value caused by deterministic propagation mechanisms such as geometric obstruction and specular reflection, calculated using 3D ray tracing technology in a known, high-precision 3D environment model. It can be directly and explicitly calculated using physical optics principles.

[0104] The total loss obtained from actual measurement is decomposed into three components with clear physical meaning. By subtracting the precisely calculable part from the total loss, the nonlinear and complex loss residuals that are difficult to describe with deterministic models are extracted. This allows the neural network model to focus on learning the most complex and nondeterministic visual feature-electromagnetic fading mapping relationship without having to learn all the basic and deterministic physical laws from scratch. This reduces the learning difficulty of the model and helps to improve the accuracy of prediction and training efficiency.

[0105] In one embodiment, the predicted path loss value is calculated based on the predicted electromagnetic fading factor and the deterministic loss value, including:

[0106] Step 501: Extract the predicted three-dimensional coordinates of the communication nodes from the future virtual three-dimensional scene, and calculate the predicted straight-line distance between the communication nodes based on the predicted three-dimensional coordinates.

[0107] In this context, the future virtual 3D scene refers to a synthetic 3D model constructed during the prediction phase by fusing a 3D model of the static environment with the predicted future states of dynamic objects, used to characterize the overall environmental state at a future moment. The predicted 3D coordinates of the communication nodes refer to the expected spatial positions of the transmitting and receiving nodes at that future moment within the coordinate system of the future virtual 3D scene. The predicted straight-line distance refers to the 3D geometrical distance between the transmitting and receiving nodes at that future moment, calculated based on the predicted 3D coordinates of the communication nodes.

[0108] The terminal accesses the pre-constructed future virtual 3D scene, extracts the predicted 3D coordinates of the communication nodes from the future virtual 3D scene, determines the specific positions of the transmitting and receiving nodes in the future scene model according to network planning or mission requirements, and reads their coordinate values. Based on the predicted 3D coordinates, the predicted straight-line distance between the communication nodes is calculated. The Euclidean distance formula between two points in 3D space is used to calculate the two extracted coordinate values, thereby obtaining the predicted straight-line distance between the nodes.

[0109] Step 502: Based on the predicted straight-line distance and communication parameters, calculate the predicted free-space basis loss value of the radio path between the corresponding communication nodes.

[0110] Specifically, communication parameters refer to the system operating parameters required for wireless link budgeting, the most important of which is the center frequency of the radio waves. The predicted free-space fundamental loss value refers to the fundamental propagation loss that a signal will suffer under ideal free-space propagation conditions at future times, calculated using free-space propagation model formulas based on the predicted straight-line distance and communication parameters.

[0111] The terminal substitutes the predicted straight-line distance and communication parameters into the same free-space propagation loss formula as in the training phase to obtain the predicted free-space basic loss value, which represents the unavoidable theoretical basic loss portion under future node distances.

[0112] Step 503: The deterministic loss value and the predicted free space basic loss value are algebraically added together to obtain the intermediate predicted loss value.

[0113] Specifically, the deterministic loss value refers to the signal attenuation caused by major geometric occlusion and reflection paths, calculated during the prediction phase by extracting 3D ray tracing features from the future virtual 3D scene. The intermediate prediction loss value is the result obtained by adding the predicted free space base loss value to the deterministic loss value, representing the sum of the portion of the future path loss that can be calculated using a deterministic physical model.

[0114] The terminal performs an algebraic addition operation. Since both the deterministic loss value and the predicted free-space fundamental loss value are in decibels, the addition operation is performed directly. The result of the addition is the intermediate predicted loss value. All interpretable, deterministic loss components are summarized.

[0115] Step 504: The predicted electromagnetic fading factor and the intermediate predicted loss value are algebraically added together to obtain the complete predicted loss value.

[0116] The predicted electromagnetic fading factor refers to the loss value, characterized by complex random fading, predicted by a cross-modal neural network model based on wireless propagation visual feature vectors extracted from future scenarios. The complete predicted loss value is the result obtained by adding the intermediate predicted loss value to the predicted electromagnetic fading factor, representing a preliminary complete prediction of the total loss of the future radio path, including both deterministic and predicted random fading components.

[0117] The terminal performs an algebraic addition operation again, adding the predicted electromagnetic fading factor and the intermediate predicted loss value. The result obtained after the addition is the complete predicted loss value, which is a preliminary total loss estimate based on the combination of the physical model and the artificial intelligence model.

[0118] Step 505: Based on the real-time environmental calibration factor, adjust the complete predicted loss value to obtain the environmentally adjusted loss value, and superimpose the system margin parameter on the environmentally adjusted loss value to obtain the predicted path loss value.

[0119] The real-time environmental calibration factor is a correction coefficient dynamically calculated based on real-time or recent measurement data. It is used to correct potential biases in model predictions, including additional effects from weather changes, ambient humidity, or background noise levels. It can be a multiplicative or additive factor close to 1. The environmental adjustment loss value is the loss value obtained after correcting the complete predicted loss value with the real-time environmental calibration factor. The system margin parameter refers to the additional loss margin reserved to ensure reliable operation of the communication system. It is used to address unavoidable uncertainties in predictions, equipment aging, and unmodeled extreme cases, ensuring link connectivity even in the worst-case scenario. The predicted path loss value refers to the path loss prediction result that can be used for network planning and link budgeting. It is a value that has been environmentally calibrated in real time and includes system design margins.

[0120] The terminal adjusts the complete predicted loss value based on a real-time environmental calibration factor. It multiplies or adds the complete predicted loss value to the real-time environmental calibration factor to obtain an environmentally adjusted loss value that more closely reflects the real-time environment. A system margin parameter is then superimposed on this environmentally adjusted loss value. Finally, the preset system margin parameter is added to the environmentally adjusted loss value to obtain the predicted path loss value. The predicted path loss value provides a crucial basis for performance evaluation and resource allocation of the communication system.

[0121] In one embodiment, based on a future virtual 3D scene, 3D ray projection feature extraction is performed to obtain deterministic loss values ​​and wireless propagation visual feature vectors, including:

[0122] Step 601: Based on the future virtual 3D scene, simulate the emission of probe rays between communication nodes to obtain probe rays, analyze the interaction state between the probe rays and all geometric objects, and obtain the main propagation path interaction report.

[0123] In this context, the "future virtual 3D scene" refers to a synthetic 3D model constructed during the prediction phase by fusing a 3D model of the static environment with the predicted future states of dynamic objects, used to characterize the overall environmental state at a future moment. Simulated probe ray emission refers to a computer simulation operation in which a large number of rays representing radio wave propagation paths are virtually emitted from the transmitting node towards the receiving node within the future virtual 3D scene. Each probe ray emitted in the simulation represents a possible signal propagation path, possessing a starting point, direction, and propagation characteristics. Geometry refers to all 3D object models constituting the future virtual 3D scene, including static and dynamic objects such as buildings, vehicles, and pedestrians. Interaction state refers to the physical interaction information generated when a single probe ray collides with geometry in the scene during propagation, including but not limited to: the 3D coordinates of the collision point, the normal direction of the collision surface, the incident angle, the reflection angle, the penetration depth, and the electromagnetic properties of the material of the collided geometry. The main propagation path interaction report refers to the data report generated after tracking and analyzing the interaction states of all probe rays, filtering and recording the ray paths with strong energy that contribute significantly to the final received signal, along with their detailed set of interaction states.

[0124] The terminal loads a pre-constructed virtual 3D scene of the future. Using the transmitting node as the origin, it emits thousands of probe rays towards the receiving node using a ray tracing engine to simulate all possible signal propagation paths. The terminal analyzes the interaction states between the probe rays and all geometric objects, tracing the propagation trajectory of each probe ray in the scene through computational geometry. When a ray intersects the surface of any geometric object, the terminal calculates the interaction state resulting from the collision based on the object's material properties and physical optics laws, including whether reflection, transmission, or diffraction occurs, and records its parameters. The terminal summarizes and analyzes all ray tracing and interaction results, selecting the paths with the largest energy contribution as the main propagation paths, and obtaining a main propagation path interaction report containing detailed interaction state information for each path.

[0125] Step 602: Based on the main propagation path interaction report, analyze the type of geometry to obtain regular geometry and irregular geometry.

[0126] Specifically, regular geometry refers to objects in the main propagation path interaction report that have regular shapes and smooth surfaces, and whose electromagnetic scattering characteristics can be accurately and concisely modeled using classical geometric optics theory or uniform geometric diffraction theory. Examples include flat walls, smooth floors, and standard-shaped metal cabinets. Irregular geometry refers to objects in the main propagation path interaction report that have complex shapes, rough surfaces, or rich structural details, and whose electromagnetic scattering behavior cannot be accurately described using simple geometric optics or uniform diffraction theory, exhibiting strong randomness or complexity. Examples include dense trees, rough brick walls, sculptures, and cluttered items.

[0127] Based on the main propagation path interaction report, the terminal analyzes the type of geometry, examines each geometry that has interacted with the ray recorded in the report, and judges it according to the 3D model characteristics of the geometry and its scattering characteristics in the interaction, according to the predefined classification rules. Based on the judgment results, the terminal divides all interacting geometry into two major categories and obtains a set of regular geometry and a set of irregular geometry.

[0128] Step 603: Based on the main propagation path interaction report, extract the blade diffraction parameters of the regular geometry, and calculate the deterministic loss value based on the blade diffraction parameters.

[0129] Specifically, edge diffraction parameters refer to the key physical parameters used to calculate diffraction loss when the probe ray interacts with the sharp edge of a regular geometric object, i.e., edge diffraction. These parameters include the diffraction angle, the geometry of the edge, the electromagnetic properties of the edge material, and the polarization of the incident wave. In this embodiment, the deterministic loss value specifically refers to the signal attenuation value obtained through physical modeling and formula calculation of the edge diffraction phenomenon caused by the regular geometric object; this portion of the loss is deterministic and can be precisely calculated.

[0130] The terminal filters out all path entries that interact with regular geometry and whose interaction type is identified as blade diffraction from the main propagation path interaction report. It extracts the blade diffraction parameters of the regular geometry and reads and calculates the various blade diffraction parameters required for diffraction loss calculation from the interaction status information, including specific diffraction angles and edge directions. Based on the blade diffraction parameters, the terminal calculates the deterministic loss value. It substitutes the extracted parameters into the classical electromagnetic diffraction calculation formula to perform numerical calculation and obtains the deterministic loss value on the diffraction path. For multiple paths, the contributions of all paths need to be combined. The deterministic loss value represents the portion of signal loss caused by the sharp edges of the regular object that can be accurately predicted.

[0131] For example, the terminal filters out path entries from the main propagation path interaction report that interact with regular geometry and whose interaction type is clearly specular reflection. For each such reflection path, the terminal reads the incident angle, the material type of the reflecting surface, and the wave polarization information from the interaction status in the report. Based on the specular reflection parameters, the terminal calculates the deterministic loss value using Fresnel's formula in the geometric optics reflection model. According to the extracted incident angle, wave polarization mode, and electromagnetic properties of the surface material, the terminal substitutes these into the Fresnel reflection formula to calculate the voltage reflection coefficient. The deterministic loss value is mainly reflected in the reduction of power, and the power loss caused by reflection can be calculated through the reflection coefficient.

[0132] For a path containing multiple reflections, the total deterministic loss is the sum of the deterministic losses of each reflection.

[0133] Step 604: Based on the main propagation path interaction report, perform feature extraction on the irregular geometry to obtain the wireless propagation visual feature vector.

[0134] Specifically, feature extraction refers to a data processing operation aimed at extracting numerical representations from raw data that effectively characterize its essential attributes and are suitable for subsequent machine learning models. Wireless propagation visual feature vectors refer to a fixed-dimensional numerical array obtained after feature extraction of interactive information from irregular geometric objects. From a visual or geometric perspective, the vector encodes the potential features of the complex and random effects that irregular objects may have on radio wave propagation. For example, it may include the statistical distribution characteristics of irregular object groups, the spectral characteristics of surface roughness, and the texture features of spatial occupancy.

[0135] The terminal extracts features from irregular geometries and analyzes how these irregular objects are presented in the interactive report. For example, it can statistically analyze the density distribution of interaction points of irregular objects within a unit solid angle, analyze the degree of disorder in the surface normal direction at the interaction points, and treat the projection of a group of irregular objects onto the area through which a ray passes as a special texture and extract its frequency domain features. The goal is to transform these complex geometric and electromagnetic relationships, which are difficult to describe with simple formulas, into a set of quantifiable features. The terminal sequentially combines multiple extracted feature values ​​into a one-dimensional array to obtain a wireless propagation visual feature vector. This wireless propagation visual feature vector will be used as input to a cross-modal neural network to predict complex fading caused by irregular objects.

[0136] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0137] Based on the same inventive concept, this application also provides a radio path loss dynamic prediction system for implementing the aforementioned radio path loss dynamic prediction method. The solution provided by this system is similar to the implementation described in the above method; therefore, the specific limitations in one or more embodiments of the radio path loss dynamic prediction system provided below can be found in the limitations of the radio path loss dynamic prediction method described above, and will not be repeated here.

[0138] In one exemplary embodiment, such as Figure 2 As shown, a dynamic prediction system for radio path loss 700 is provided, comprising:

[0139] The modeling module 701 is used to acquire stereo image sequences and motion pose sequences, and to perform 3D scene analysis based on the stereo image sequences and motion pose sequences to obtain 3D scene sequences.

[0140] The trajectory module 702 is used to predict the motion trajectory of each dynamic object based on a 3D scene sequence, and obtain a predicted 3D state set.

[0141] The feature module 703 is used to construct a future virtual 3D scene based on the 3D scene sequence and the predicted 3D state set, and to extract 3D ray projection features based on the future virtual 3D scene to obtain deterministic loss values ​​and wireless propagation visual feature vectors.

[0142] The prediction module 704 is used to input the wireless propagation visual feature vector into the cross-modal neural network model, perform electromagnetic attenuation mapping, and obtain the predicted electromagnetic fading factor.

[0143] The loss module 705 is used to calculate the predicted path loss value based on the predicted electromagnetic fading factor and the deterministic loss value; the predicted path loss value is used to characterize the predicted loss of the radio path between communication nodes.

[0144] Furthermore, the system also includes a training module for:

[0145] Based on the original time-series data stream, 3D scene analysis is performed to obtain dense 3D scene and object dynamic history trajectory. Based on the dense 3D scene and object dynamic history trajectory, a standard virtual 3D scene is constructed.

[0146] Three-dimensional ray projection feature extraction is performed on a standard virtual three-dimensional scene to obtain standard deterministic loss value and standard visual feature vector;

[0147] Based on the channel impulse response sequence and standard deterministic loss value, the path loss value is calculated, and the standard visual feature vector and the path loss value are paired to obtain feature loss paired samples.

[0148] Using standard visual feature vectors as input features and path loss values ​​as labels, we integrate feature loss paired samples to obtain a standardized dataset.

[0149] Based on a standardized dataset, the cross-modal neural network model framework is iteratively trained to obtain the cross-modal neural network model.

[0150] Furthermore, the training module is also used for:

[0151] Based on the channel impulse response sequence, the total received power is calculated, and based on the device calibration parameters, the total received power is calibrated to obtain the received power value.

[0152] Calculate the difference between the received power value and the transmitted power value to obtain the measured value of the path loss;

[0153] Based on a dense 3D scene, the free space propagation distance of a standard radio path is extracted, and the free space fundamental loss value is calculated based on the free space propagation distance.

[0154] The free space base loss value is calculated using the following formula:

[0155]

[0156] in, d is the free space fundamental loss value, f is the free space propagation distance, c is the frequency of the radio wave, and c is the speed of light in a vacuum.

[0157] The path loss value is calculated based on the measured path loss value, the free space basic loss value, and the standard deterministic loss value.

[0158] Furthermore, the formula for calculating the path loss value is as follows:

[0159]

[0160] in, This is the path loss value. This is the measured value of path loss. This represents the basic loss value in free space. This is the standard deterministic loss value.

[0161] Furthermore, the loss module 705 is also used for:

[0162] Extract the predicted 3D coordinates of communication nodes from a future virtual 3D scene, and calculate the predicted straight-line distance between communication nodes based on the predicted 3D coordinates;

[0163] Based on the predicted straight-line distance and communication parameters, the predicted free-space basis loss value of the radio path between the corresponding communication nodes is calculated.

[0164] The intermediate predicted loss value is obtained by algebraically adding the deterministic loss value and the predicted free space basic loss value.

[0165] The predicted electromagnetic fading factor and the intermediate predicted loss value are algebraically added together to obtain the complete predicted loss value.

[0166] Based on the real-time environmental calibration factor, the complete predicted loss value is adjusted to obtain the environmentally adjusted loss value. The system margin parameter is then superimposed on the environmentally adjusted loss value to obtain the predicted path loss value.

[0167] Furthermore, feature module 703 is also used for:

[0168] Based on a future virtual 3D scene, simulated detection ray emission is performed between communication nodes to obtain the detection ray. The interaction state between the detection ray and all geometric objects is analyzed to obtain the main propagation path interaction report.

[0169] Based on the main propagation path interaction report, the type of geometry is analyzed to obtain regular geometry and irregular geometry;

[0170] Based on the main propagation path interaction report, the blade-shaped diffraction parameters of the regular geometry are extracted, and the deterministic loss value is calculated based on the blade-shaped diffraction parameters.

[0171] Based on the main propagation path interaction report, feature extraction is performed on the irregular geometry to obtain the wireless propagation visual feature vector.

[0172] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the radio path loss dynamic prediction method as described above.

[0173] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0174] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0175] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for dynamic prediction of radio path loss, characterized in that, The method includes: A stereo image sequence and a motion pose sequence are acquired, and a three-dimensional scene is analyzed based on the stereo image sequence and the motion pose sequence to obtain a three-dimensional scene sequence. Based on the three-dimensional scene sequence, the motion trajectory of each dynamic object is predicted to obtain a predicted three-dimensional state set. Based on the three-dimensional scene sequence and the predicted three-dimensional state set, a future virtual three-dimensional scene is constructed, and based on the future virtual three-dimensional scene, three-dimensional ray projection feature extraction is performed to obtain deterministic loss value and wireless propagation visual feature vector. The wireless propagation visual feature vector is input into a cross-modal neural network model to perform electromagnetic attenuation mapping and obtain the predicted electromagnetic fading factor. Based on the predicted electromagnetic fading factor and the deterministic loss value, the predicted path loss value is calculated; the predicted path loss value is used to characterize the predicted loss of the radio path between communication nodes.

2. The method according to claim 1, characterized in that, The cross-modal neural network model was obtained through the following method: Based on the original time-series data stream, a 3D scene is parsed to obtain a dense 3D scene and the dynamic historical trajectory of objects. Based on the dense 3D scene and the dynamic historical trajectory of objects, a standard virtual 3D scene is constructed. Three-dimensional ray projection feature extraction is performed on the standard virtual three-dimensional scene to obtain standard deterministic loss value and standard visual feature vector; Based on the channel impulse response sequence and the standard deterministic loss value, the path loss value is calculated, and the standard visual feature vector and the path loss value are paired to obtain feature loss paired samples; Using the standard visual feature vector as input features and the path loss value as label, the feature loss paired samples are integrated to obtain a standardized dataset. Based on the standardized dataset, the cross-modal neural network model framework is iteratively trained to obtain the cross-modal neural network model.

3. The method according to claim 2, characterized in that, The calculation of path loss based on the channel impulse response sequence and the standard deterministic loss value includes: Based on the channel impulse response sequence, the total received power is calculated, and based on the device calibration parameters, the total received power is calibrated to obtain the received power value. The difference between the received power value and the transmitted power value is calculated to obtain the measured path loss value; Based on the dense 3D scene, the free space propagation distance of the standard radio path is extracted, and the free space fundamental loss value is calculated based on the free space propagation distance. The free space fundamental loss value is calculated using the following formula: in, d is the free space fundamental loss value, f is the free space propagation distance, c is the frequency of the radio wave, and c is the speed of light in a vacuum. The path loss value is calculated based on the measured path loss value, the free space basic loss value, and the standard deterministic loss value.

4. The method according to claim 3, characterized in that, The formula for calculating the path loss value is: in, This is the path loss value. This is the measured value of path loss. This represents the basic loss value in free space. This is the standard deterministic loss value.

5. The method according to claim 1, characterized in that, The calculation of the predicted path loss value based on the predicted electromagnetic fading factor and the deterministic loss value includes: The predicted three-dimensional coordinates of the communication nodes are extracted from the future virtual three-dimensional scene, and the predicted straight-line distance between the communication nodes is calculated based on the predicted three-dimensional coordinates. Based on the predicted straight-line distance and communication parameters, the predicted free-space fundamental loss value of the radio path between the corresponding communication nodes is calculated; The intermediate predicted loss value is obtained by algebraically adding the deterministic loss value and the predicted free space basic loss value. The predicted electromagnetic fading factor and the intermediate predicted loss value are algebraically added together to obtain the complete predicted loss value. Based on the real-time environmental calibration factor, the complete predicted loss value is adjusted to obtain the environmentally adjusted loss value, and the system margin parameter is superimposed on the environmentally adjusted loss value to obtain the predicted path loss value.

6. The method according to claim 1, characterized in that, The step of extracting three-dimensional ray projection features based on the future virtual three-dimensional scene to obtain deterministic loss values ​​and wireless propagation visual feature vectors includes: Based on the future virtual 3D scene, simulated detection ray emission is performed between the communication nodes to obtain the detection ray. The interaction state between the detection ray and all geometric objects is analyzed to obtain the main propagation path interaction report. Based on the main propagation path interaction report, the type of the geometry is analyzed to obtain regular geometry and irregular geometry; Based on the main propagation path interaction report, the blade-shaped diffraction parameters of the regular geometry are extracted, and the deterministic loss value is calculated based on the blade-shaped diffraction parameters. Based on the main propagation path interaction report, feature extraction is performed on the irregular geometry to obtain the wireless propagation visual feature vector.

7. A dynamic prediction system for radio path loss, characterized in that, The system includes: The modeling module is used to acquire a stereo image sequence and a motion pose sequence, and to perform three-dimensional scene analysis based on the stereo image sequence and the motion pose sequence to obtain a three-dimensional scene sequence. The trajectory module is used to predict the motion trajectory of each dynamic object based on the three-dimensional scene sequence, and obtain a predicted three-dimensional state set. The feature module is used to construct a future virtual 3D scene based on the 3D scene sequence and the predicted 3D state set, and to perform 3D ray projection feature extraction based on the future virtual 3D scene to obtain deterministic loss values ​​and wireless propagation visual feature vectors. The prediction module is used to input the wireless propagation visual feature vector into a cross-modal neural network model, perform electromagnetic attenuation mapping, and obtain the predicted electromagnetic fading factor. The loss module is used to calculate the predicted path loss value based on the predicted electromagnetic fading factor and the deterministic loss value; the predicted path loss value is used to characterize the predicted loss of the radio path between communication nodes.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images