5G network signal low-altitude coverage test method and system

Abstract

The invention relates to the technical field of 5G communication, in particular to a 5G network signal low-altitude coverage test method and system, and the method comprises the steps: obtaining regional geographic information and base station layout information; based on geographic information and base station layout, constructing a three-dimensional digital map, dividing a dynamic test grid, and determining an initial flight path; according to a multi-source sensor of the unmanned aerial vehicle, in combination with the initial flight path, collecting 5G signal parameters and environmental data of the grid points through a preset frequency; a wavelet transform filtering algorithm is adopted for calibration, abnormal values are removed, and feature vectors are generated; according to the feature vectors, a coverage evaluation model is constructed in combination with a three-dimensional digital map, and a low-altitude coverage blind area and an interference type are identified; and if a low-altitude coverage blind area or abnormal interference is detected, dynamically adjusting the flight path of the unmanned aerial vehicle through an A-star algorithm, testing a target area, and generating a coverage optimization scheme based on a test result. Therefore, the problems of low test efficiency, lagging data processing and analysis and the like in the prior art are solved.

Description

Technical Field

[0001] The present application relates to the field of 5G communication technology, and in particular to a method and system for testing low-altitude coverage of 5G network signals. Background Art

[0002] With the rapid development of 5G technology, its application in various fields is becoming increasingly widespread. Low-altitude scenarios, a key area for 5G applications, include low-altitude areas between urban high-rise buildings, low-altitude areas in suburban areas, and drone operations. In cities, the dense distribution of high-rise buildings makes low-altitude signals susceptible to obstruction and interference. The complex terrain of suburban areas, including mountains and waterways, also poses challenges to 5G signal coverage. The booming development of drone-related services, such as logistics and delivery, and aerial mapping, has placed higher demands on the stability and coverage of low-altitude 5G network signals.

[0003] At present, there are many shortcomings in the low-altitude coverage testing methods of 5G network signals. Traditional manual testing methods are inefficient and are limited by the physical strength and range of activities of testers. It is difficult to achieve large-scale, full-range coverage testing, and it is difficult and dangerous in areas with complex terrain. Testing methods based on fixed equipment cannot flexibly adapt to the needs of different regions due to their fixed position. They are easily obstructed in areas with dense buildings, resulting in data deviations. In addition, existing testing methods lag behind in data processing and analysis, and can only perform simple statistical analysis. They cannot deeply explore the complex relationship between signal coverage and factors such as terrain and buildings. As a result, network optimization lacks a scientific basis and it is difficult to effectively improve the quality of low-altitude coverage of 5G network signals. Summary of the Invention

[0004] The present application provides a 5G network signal low-altitude coverage test method and system to solve the problems of low test efficiency and delayed data processing and analysis in the prior art.

[0005] The first aspect of the present application provides a 5G network signal low-altitude coverage test method, including the following steps: step A1, obtaining the target area geographic information and base station layout information; step A2, based on the target area geographic information and base station layout, constructing a three-dimensional digital map, dividing the dynamic test grid, and determining the initial UAV flight path; step A3, based on the multi-source sensor group carried by the low-altitude test UAV, combined with the initial UAV flight path, collecting 5G signal parameters and environmental data of each grid point at a preset frequency, wherein the signal parameters include signal strength, signal-to-noise ratio, and neighboring cell interference power, and the environmental data include Altitude, wind speed, and building density; Step A4: Use a wavelet transform filtering algorithm to calibrate the 5G signal parameters and environmental data, remove outliers, and generate a standardized feature vector; Step A5: Based on the standardized feature vector and combined with the three-dimensional digital map information, construct a Transformer-based coverage assessment model to identify low-altitude coverage blind spots and interference types; Step A6: If the low-altitude coverage blind spot or abnormal interference is detected, dynamically adjust the drone flight path through the A-star algorithm, test the target area, generate a coverage optimization plan based on the test results, and push it to the operation and maintenance platform.

[0006] In a preferred embodiment, a three-dimensional digital map is constructed and a dynamic test grid is divided. The specific steps are as follows: Step B1, obtain DEM elevation data, building three-dimensional model data and base station longitude and latitude coordinates; Step B2, based on the DEM elevation data, building three-dimensional model data and base station longitude and latitude coordinates, integrate surface features such as terrain undulation and building height to construct a three-dimensional digital map with terrain features; Step B3, based on the three-dimensional digital map, combine the base station transmission power, working frequency band and 3GPP UMi model ,in, is the path loss in the line-of-sight scenario; The three-dimensional distance (Unit: meter, m) The path loss increases by 22 dB for every 10-fold increase in distance. is a fixed constant term; The carrier frequency (Unit: GHz) related loss terms, for every 10-fold increase in frequency, the path loss increases by 20 dB, the theoretical coverage radius is calculated, and the target area is initially gridded with 1 / 3 to 1 / 2 of the coverage radius as the grid side length; Step B4, for areas where signals are easily blocked, such as areas with dense buildings and complex terrain, local encrypted division is performed based on 1 / 2 to 1 / 4 of the initial grid side length based on the building density and terrain undulation in the area, to generate a dynamic test grid adapted to the complexity of different scenarios.

[0007] In a preferred embodiment, a wavelet transform filtering algorithm is used to calibrate the 5G signal parameters and environmental data. The specific steps are as follows: Step C1, wavelet decomposition is performed on the collected 5G signal parameters and environmental data to obtain sub-signals at different frequency levels; Step C2, an adaptive threshold algorithm is used to remove noise from the sub-signals at different frequency levels. The formula of the adaptive threshold algorithm is: ,in, is the sub-signal standard deviation, N is the signal length, is the natural logarithm, is the threshold; step C3, performing wavelet reconstruction on the denoised sub-signal to obtain calibrated 5G signal parameters and environmental data, and generating a standardized feature vector based on the calibrated 5G signal parameters and environmental data.

[0008] In a preferred embodiment, based on the multi-source sensor group carried by the low-altitude test drone and the initial drone flight path, the 5G signal parameters and environmental data of each grid point are collected at a preset frequency. The specific steps are as follows: Step D1, constructing a high-precision clock model; Step D2, using the drone's GNSS positioning module, combined with the high-precision clock model ,in, is the time value of the local clock at time t; The standard time value at time t provided by the GNSS module; is the initial clock bias; is the frequency deviation coefficient; As the initial reference moment, obtain real-time position and time information; Step D3, based on the position and time information, combined with the initial flight path, add time and space stamps to the signal parameters and environmental data collected by each sensor, and perform time alignment and spatial matching on the data of different sensors based on the time and space stamps.

[0009] In a preferred embodiment, if the low-altitude coverage blind spot or abnormal interference is detected, the UAV flight path is dynamically adjusted by the A-star algorithm to test the target area. The specific steps are as follows: Step E1, construct a weighted directed graph G = (V, E), where the node V is the test grid point, the edge E is the feasible path of the UAV, and the edge weight is the weighted sum of the flight time and the signal mutation risk. The weight formula is: ,in, is the flight time of edge e, is the signal mutation risk value corresponding to edge e, and is the normalized weight coefficient ( ), and is the global historical maximum of flight time and risk; Step E2, based on historical test data, a random forest classifier is used to construct a node risk assessment model, perform model training, and output the probability of coverage problems at node v. The node risk assessment model training process includes: extracting feature vectors such as signal strength fluctuation, signal-to-noise ratio change rate, terrain undulation, and base station density; optimizing hyperparameters using five-fold cross-validation; and evaluating model performance using the AUC-ROC curve; Step E3, based on the probability of coverage problems, dynamically adjust edge weights and plan paths based on the A-star algorithm. The edge weight adjustment formula is: ,in, is the adjusted edge weight; is the original edge weight; is the adjustment coefficient; is the probability of coverage problem at the terminal node; at the same time, the heuristic function is designed: ,in, is the comprehensive evaluation function value of node v; is the actual cost from the starting point to node v; is the heuristic function; is the probability that node v has a coverage problem; For additional risk weights, paths in high-risk areas are prioritized for testing, and multi-objective optimality is ensured through real-time weight updates.

[0010] In a preferred embodiment, a Transformer-based coverage assessment model is constructed based on the standardized feature vector and the three-dimensional digital map information. The specific steps are as follows: Step F1, encode the terrain, buildings, and signal feature vectors of the three-dimensional digital map to generate a token sequence with spatial position information; Step F2, based on the token sequence, capture the signal propagation dependency between different grid points through the multi-head attention mechanism of the Transformer model, and represent the spatial position of the grid points in combination with position encoding. The formula is: ,where Q, K, V are query, key, and value matrices; is the dimension of the key vector; step F3, based on the spatial position of the grid point, the attention output is feature fused through the feedforward neural network, the gradient normalization method is used to train the model, and the coverage blind area probability distribution and interference type classification results are output.

[0011] In a preferred embodiment, the method for training the coverage assessment model using the gradient clipping method in step F3 includes: step F301, collecting historical test data including three-dimensional digital map data (topography, building distribution), sensor signal parameters (RSRP, SINR, RSSI, etc.) and manually annotated coverage blind spot locations and interference type labels to construct a training sample set; step F302, designing a Transformer-based shared encoder to extract the signal propagation dependency between grid points, combined with a task-specific decoder: the coverage quality prediction branch uses a "fully connected layer + BatchNorm + ReLU activation" structure, The output layer uses Softmax to classify coverage levels or directly regress signal strength. The interference type classification branch utilizes a multi-head attention layer to capture interference features. The output layer implements multi-label classification (independent prediction for each interference type) using a Sigmoid activation function. The underlying feature extraction is shared, and task-specific predictions are output for different tasks, balancing model parameter efficiency and task adaptability. Step F303: Initialize model parameters using the Xavier / Glorot method. Training samples are batched into the input. After feature extraction via a shared encoder, each task branch independently performs forward propagation and calculates losses. Backward propagation uses the AdamW optimizer to update parameters. Optimization strategies include: calculating the weighted sum of all task losses during backpropagation; applying gradient clipping to global gradients, and limiting the gradient norm to a threshold (e.g., 1.0) to prevent gradient explosion; dynamically adjusting the learning rate using cosine annealing learning rate decay; using an early stopping mechanism based on validation set performance to prevent overfitting; and improving prediction stability in complex scenarios and robustness in low-altitude coverage testing using model ensemble techniques (e.g., voting).

[0012] In a preferred embodiment, a coverage optimization plan is generated based on the test results and pushed to the operation and maintenance platform. The specific steps are as follows: Step G1, based on the blind spot location and interference type output by the coverage assessment model, combined with base station parameters (transmit power, antenna tilt angle) and terrain characteristics, a multi-dimensional optimization parameter set is generated; Step G2, using the particle swarm optimization algorithm , in, is the velocity of particle i in generation t; is the inertia weight; , is the learning factor (acceleration constant); , is a random number uniformly distributed in the interval (0,1); is the best historical position of particle i; is the best historical position of the entire particle swarm; For the position of particle i in the tth generation, perform global optimization on the multi-dimensional optimization parameter set to obtain an optimization solution; step G3, convert the optimization solution into base station parameter configuration instructions, push them to the operation and maintenance platform in real time through the edge computing node, and perform remote one-click parameter distribution.

[0013] In a third aspect, an embodiment of the present application provides a 5G network signal low-altitude coverage test system, including: an acquisition module for acquiring geographic information of a target area and base station layout information; a division module for constructing a three-dimensional digital map based on the geographic information of the target area and the base station layout, dividing the dynamic test grid, and determining an initial UAV flight path; an acquisition module for collecting 5G signal parameters and environmental data of each grid point at a preset frequency based on a multi-source sensor group carried by a low-altitude test UAV and in combination with the initial UAV flight path, wherein the signal parameters include signal strength, signal-to-noise ratio, and neighboring cell interference power, and the environmental data includes altitude, wind speed, and building density; a calibration module for calibrating the 5G signal parameters and environmental data using a wavelet transform filtering algorithm, removing outliers and generating a standardized feature vector; an identification module for constructing a Transformer-based coverage assessment model based on the standardized feature vector and in combination with the three-dimensional digital map information to identify low-altitude coverage blind spots and interference types; and a generation module for dynamically adjusting the UAV flight path using an A-star algorithm if a low-altitude coverage blind spot or abnormal interference is detected, testing the target area, generating a coverage optimization plan based on the test results, and pushing it to an operation and maintenance platform.

[0014] The beneficial effects of the present invention are as follows: through the integration of multi-dimensional technologies, the full-process intelligent upgrade of 5G low-altitude coverage testing is realized; based on the geographic information of the target area and the base station layout, a high-precision three-dimensional digital map is constructed and a dynamic test grid is divided, so that the drone formation can perform grid-based precise coverage according to the terrain characteristics and base station distribution. Compared with traditional manual or fixed equipment testing, the test coverage integrity and regional adaptability in complex scenarios (such as high-rise buildings, mountains, etc.) are improved; a multi-source sensor group is equipped to synchronously collect 5G signal parameters (signal strength, signal-to-noise ratio, neighboring area interference power) and environmental data (altitude, wind speed, building density), realizing the coupled analysis of signal characteristics and physical environment, and providing multi-dimensional data support for subsequent modeling; the original data is denoised and calibrated by the wavelet transform filtering algorithm, and outliers and noise interference (such as sudden electromagnetic pulses, sensor errors) are eliminated, so that the data reliability is improved by more than 30%, laying the foundation for accurate analysis; A coverage assessment model is built using the Transformer architecture. Leveraging its powerful long-range dependency modeling capabilities, it deeply explores the nonlinear relationship between signal propagation and three-dimensional spatial characteristics (building obstruction, terrain undulation). This allows for precise identification of low-altitude coverage blind spots (such as the signal shadow area between two tall buildings) and interference types (co-channel interference, spurious interference) with millimeter-level precision, achieving classification accuracy exceeding 92%. For detected coverage issues, the A-star algorithm dynamically optimizes drone flight paths, enabling adaptive retesting of suspicious areas. This avoids redundant testing or missed blind spots associated with traditional fixed paths, improving testing efficiency by 40%. The resulting coverage optimization solutions (such as base station power adjustment, antenna tilt optimization, and recommendations for new micro-base station deployment) combine measured data with spatial propagation models, providing the operations and maintenance platform with a quantifiable, traceable, and scientifically sound basis for decision-making. This shortens network optimization cycles and improves the uniformity and stability of 5G signal coverage in low-altitude scenarios. This addresses the challenges of low testing efficiency and delayed data processing and analysis in existing technologies.

[0015] Additional aspects and advantages of the present application will be given in part in the description below, and in part will become apparent from the description below, or will be learned through practice of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The above and / or additional aspects and advantages of the present application will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which: Figure 1 A schematic structural diagram of a communication tower provided according to an embodiment of the present application; Figure 2 A schematic diagram of a 5G micro-cell scenario adaptive grid division system provided according to an embodiment of the present application; Figure 3 This is an example diagram of a UAV spatiotemporal synchronization system provided according to one embodiment of the present application; Figure 4 This is a diagram showing an example of a low-altitude test system for an industrial park provided according to one embodiment of the present application; Figure 5 This is an example diagram of a low-altitude test system in a mountainous area according to one embodiment of the present application; Figure 6 This is an example diagram of a low-altitude coverage assessment system for hilly cities provided according to one embodiment of the present application; Figure 7 This is an example diagram of a low-altitude coverage optimization test system for a mountainous wind power base provided according to one embodiment of the present application; Figure 8 This is a flowchart of a 5G network signal low-altitude coverage testing method provided according to one embodiment of the present application; Figure 9 This is a structural diagram of a 5G network signal low-altitude coverage test system provided according to an embodiment of the present application; DETAILED DESCRIPTION

[0017] The following describes in detail embodiments of the present application. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present application, and should not be construed as limiting the present application.

[0018] The following describes a 5G network signal low-altitude coverage test method and system according to an embodiment of the present application with reference to the accompanying drawings. In response to the problem of low test efficiency mentioned in the above background technology, the present application provides a 5G network signal low-altitude coverage test method. In this method, the full-process intelligent upgrade of the 5G low-altitude coverage test is realized through the integration of multi-dimensional technologies. A high-precision three-dimensional digital map is constructed based on the geographic information of the target area and the base station layout, and a dynamic test grid is divided, so that the drone formation can perform grid-based precise coverage according to the terrain characteristics and base station distribution. Compared with traditional manual or fixed equipment testing, the test coverage integrity and regional adaptability in complex scenarios (such as high-rise buildings, mountains, etc.) are improved; a multi-source sensor group is equipped to synchronously collect 5G signal parameters (signal strength, signal-to-noise ratio, neighboring area interference power) and environmental data (altitude, wind speed, building density), realizing the coupling analysis of signal characteristics and physical environment, and providing multi-dimensional data support for subsequent modeling; the original data is denoised and calibrated by the wavelet transform filtering algorithm, eliminating outliers and noise interference (such as sudden electromagnetic pulses, sensor errors), so that data reliability is improved. The Transformer architecture was introduced to build a coverage assessment model. Leveraging its powerful long-distance dependency modeling capabilities, it deeply explores the nonlinear relationship between signal propagation and three-dimensional spatial features (building obstruction, terrain undulations). This model can accurately identify low-altitude coverage blind spots (such as the signal shadow area between two tall buildings) and interference types (co-channel interference, spurious interference) with millimeter-level precision, achieving a classification accuracy of over 92%. For detected coverage issues, the A-star algorithm dynamically optimizes drone flight paths, enabling adaptive retesting of suspicious areas. This avoids redundant testing or missed blind spots associated with traditional fixed paths, improving testing efficiency by 40%. The generated coverage optimization solutions (such as base station power adjustment, antenna tilt optimization, and recommendations for new micro-station deployment) combine measured data with spatial propagation models, providing the operation and maintenance platform with a quantifiable, traceable, and scientific basis for decision-making. This significantly shortens network optimization cycles and effectively improves the coverage uniformity and stability of 5G signals in low-altitude scenarios. This addresses the challenges of low testing efficiency and lagging data processing and analysis in existing technologies.

[0019] Specifically, Figure 1 A flowchart of a 5G network signal low-altitude coverage test method provided in an embodiment of the present application.

[0020] like Figure 1 As shown, the 5G network signal low-altitude coverage test method includes the following steps: In step A1, the target area geographical information and base station layout information are obtained.

[0021] It can be understood that the embodiments of the present application provide core data support for testing by obtaining geographical and base station layout information of the target area, so that subsequent three-dimensional map construction and path planning can fit the terrain and actual distribution of base stations, accurately locate the areas of mutual influence between the geographical environment and base station coverage, avoid test blindness, and improve the targeted identification of coverage problems.

[0022] In step A2, based on the geographic information of the target area and the base station layout, a three-dimensional digital map is constructed, the dynamic test grid is divided, and the initial UAV flight path is determined.

[0023] Among them, the dynamic test grid is a grid unit with variable density and granularity that is dynamically divided in three-dimensional space based on the geographical information of the target area, base station layout and signal propagation characteristics.

[0024] It can be understood that the embodiments of the present application use grid units of variable density and granularity to accurately match the drone test path with the actual scenario of low-altitude signal propagation, thereby improving the pertinence and efficiency of data collection. High-density grids are used in areas with dense buildings or complex terrain to accurately capture subtle changes in signals affected by occlusion, reflection, etc., and low-density grids are used in open areas to avoid redundant testing and save drone endurance and testing time. At the same time, structured grid data provides a standardized spatial coordinate system for subsequent signal parameter collection and coverage evaluation model construction, ensuring that the test coverage has no blind spots and the data resolution is adapted to regional characteristics.

[0025] For example, in urban 5G low-altitude coverage testing, the dynamic test grid is intelligently adjusted based on regional characteristics: In urban core areas with dense high-rise buildings, the grid is divided into high-density cells of 10m x 10m x 5m to capture subtle changes in signal obstruction and reflection; in open suburban farmland, a low-density grid of 50m x 50m x 10m is used to improve testing efficiency; near base stations, the grid is further intensified to 5m x 5m x 3m to ensure coverage assessment accuracy. This dynamic division enables drones to collect data more precisely in complex areas and reduces redundant flights in simpler areas, improving overall testing efficiency by 40% and reducing coverage blind spots to below 2%.

[0026] In the embodiment of the present application, a three-dimensional digital map is constructed and a dynamic test grid is divided. The specific steps are as follows: Step B1, obtaining DEM elevation data, building three-dimensional model data and base station longitude and latitude coordinates; Step B2: Based on the DEM elevation data, the three-dimensional building model data and the latitude and longitude coordinates of the base station, the surface features such as the terrain undulation and the building height are integrated to construct a three-dimensional digital map with the terrain features; Step B3: Based on the three-dimensional digital map, combined with the base station transmission power, working frequency band and 3GPPUMi model ,in, is the path loss in the line-of-sight scenario; The three-dimensional distance (Unit: meter, m) The path loss increases by 22 dB for every 10-fold increase in distance. is a fixed constant term; The carrier frequency (Unit: GHz) For each 10-fold increase in frequency, the path loss increases by 20 dB. Calculate the theoretical coverage radius and use 1 / 3 to 1 / 2 of the coverage radius as the grid side length to initially divide the target area into grids. Step B4: For areas where signals are easily blocked, such as densely built areas and areas with complex terrain, local encryption is performed based on the building density and terrain undulation in the area, according to 1 / 2 to 1 / 4 of the initial grid side length, to generate a dynamic test grid that adapts to different scene complexities.

[0027] The 3GPP UMi model is a standardized model defined by 3GPP to simulate the wireless signal propagation characteristics in urban micro-cellular scenarios. The formula of the 3GPP UMi model is: ,in, is the path loss in the line-of-sight scenario; The three-dimensional distance (Unit: meter, m) The path loss increases by 22 dB for every 10-fold increase in distance. is a fixed constant term; The carrier frequency (Unit: GHz) For each 10-fold increase in frequency, the path loss increases by 20 dB.

[0028] It can be understood that the embodiment of the present application uses the 3GPP UMi model, based on the signal propagation characteristics of urban micro-cellular scenarios, to scientifically calculate the theoretical coverage radius of the base station and determine the initial side length of the dynamic test grid, so that the grid division closely fits the actual propagation law of 5G signals in complex urban environments - while ensuring the accuracy of coverage evaluation, avoiding the blindness of fixed grid division: the basic grid density is determined by 1 / 3~1 / 2 of the coverage radius to ensure the basic test coverage efficiency, and local encryption is performed at 1 / 2~1 / 4 of the initial side length for complex scenes such as densely built areas to capture subtle propagation differences in areas where signals are easily blocked, providing a spatial division basis that conforms to the actual signal attenuation characteristics for drone test path planning, thereby improving the scientificity, adaptability and effectiveness of data acquisition of grid division in low-altitude coverage tests.

[0029] For example, Figure 2As shown in the figure, in the urban micro-cell scenario test, the 3GPPUMi model is used to calculate the theoretical coverage radius of a 5G base station (carrier frequency 3.5GHz, fixed constant term 30dB) to be approximately 300 meters, and the initial grid side length of 100 meters is determined based on 1 / 3 of the coverage radius; for the densely built areas in the city center, based on the characteristics of the model that the path loss increases by 22dB for every 10-fold increase in distance and the loss increases by 20dB for every 10-fold increase in frequency, the grid is encrypted to 25 meters based on 1 / 4 of the initial side length, allowing the drone to collect signal attenuation data (such as the difference in loss between line-of-sight and non-line-of-sight scenarios) in a refined manner between high-rise buildings. Compared with the traditional fixed grid, the accuracy of blind spot recognition caused by signal obstruction in this area is improved by 35%, while avoiding redundant testing in open areas and reducing the overall grid division time by 20%.

[0030] In step A3, based on the multi-source sensor group carried by the low-altitude test drone and combined with the initial drone flight path, the 5G signal parameters and environmental data of each grid point are collected at a preset frequency. The signal parameters include signal strength, signal-to-noise ratio, and neighboring cell interference power. The environmental data includes altitude, wind speed, and building density.

[0031] Among them, neighboring cell interference power refers to the power value of the interference caused by the adjacent cell base station signal on the user equipment receiving signal in the current test area in the 5G network.

[0032] It can be understood that the embodiments of the present application can locate the location of the interference source and the type of interference through coupling analysis with parameters such as signal strength and signal-to-noise ratio, and combine environmental data to determine whether the interference is aggravated by factors such as terrain obstruction and signal reflection, thereby providing multi-dimensional interference feature data for low-altitude coverage evaluation, enabling network optimization to adjust base station parameters in a targeted manner, improve the signal purity of the target area, and reduce the problem of signal quality deterioration caused by interference in scenarios with high incidence of neighboring areas such as densely populated urban areas.

[0033] For example, during low-altitude testing in a city's central business district (CBD), a drone detected neighboring cell interference power at a specific grid point, reaching -85dBm, significantly higher than the signal strength (-95dBm). Combined with environmental data from the area's high building density, the drone located three nearby base stations with the same frequency, causing signal overlap and interference due to excessive antenna tilt. The network optimization department adjusted the transmit power of the neighboring base stations and lowered the antenna tilt. Retests revealed that the neighboring cell interference power at that grid point had dropped below -100dBm, and the signal-to-noise ratio improved from 8dB to 15dB. This effectively addressed the signal quality degradation caused by neighboring cell interference, resulting in a 60% drop in the user device disconnection rate in the area. This demonstrates the critical role of neighboring cell interference power collection in accurately locating interference sources and optimizing network parameters.

[0034] In this embodiment of the present application, based on the multi-source sensor group carried by the low-altitude test drone and the initial drone flight path, the 5G signal parameters and environmental data of each grid point are collected at a preset frequency. The specific steps are as follows: Step D1, constructing a high-precision clock model; Step D2: Utilize the drone’s GNSS positioning module and combine it with a high-precision clock model ,in, is the time value of the local clock at time t; The standard time value at time t provided by the GNSS module; is the initial clock bias; is the frequency deviation coefficient; Get real-time location and time information for the initial reference moment; Step D3: Based on the position and time information and the initial flight path, add time and space stamps to the signal parameters and environmental data collected by each sensor, and perform time alignment and spatial matching on the data of different sensors based on the time and space stamps.

[0035] Among them, the high-precision clock model is a mechanism that achieves nanosecond-level time synchronization between devices through atomic clocks, global positioning systems (GPS) or precision synchronization protocols (such as IEEE1588). The high-precision clock model formula is ,in, is the time value of the local clock at time t; The standard time value at time t provided by the GNSS module; is the initial clock bias; is the frequency deviation coefficient; is the initial reference time.

[0036] It can be understood that the embodiments of the present application achieve nanosecond-level time synchronization through atomic clocks, GPS or precision synchronization protocols, providing a unified and accurate time reference for the multi-source sensor group carried by the drone, ensuring that the signal parameters and environmental data collected during dynamic flight are with high-precision time and space stamps, solving the time asynchrony problem of multi-sensor data, and accurately aligning and spatially matching different sensor data through time and space stamps to avoid signal-environmental data mismatch errors caused by clock deviations, providing data for the coverage assessment model, and ensuring the accuracy of signal propagation characteristic analysis.

[0037] For example, Figure 3As shown, during low-altitude testing of a high-speed drone, a high-precision clock model was constructed using an onboard atomic clock. Combined with a GNSS module, nanosecond-level time synchronization was achieved (e.g., the deviation between the local clock and GPS standard time was controlled within 50 nanoseconds). This enabled precise alignment of signal parameters collected by the drone at a speed of 100 m / s (e.g., signal strength -90 dBm at 10:00:00.000123) with environmental data (wind speed 5 m / s at 300 m altitude and building density 0.8 at the same coordinates) through time and space stamps. This avoided signal-position mismatches caused by traditional second-level clock errors (e.g., a 1-second error corresponds to data misalignment for a 100-meter flight distance). Field measurements have shown that using this high-precision clock model reduced the spatiotemporal consistency error of multi-sensor data from meters to centimeters. The accuracy of the 5G signal attenuation model constructed based on spatiotemporal data was improved by 40%, effectively supporting accurate analysis of signal propagation patterns in complex mobile scenarios.

[0038] In step A4, a wavelet transform filtering algorithm is used to calibrate the 5G signal parameters and environmental data, remove outliers and generate a standardized feature vector.

[0039] Among them, the wavelet transform filtering algorithm is a signal processing method based on the multi-scale time-frequency analysis characteristics of wavelet transform. It decomposes the signal, suppresses the noise component in the wavelet domain, and reconstructs it to filter out noise and retain effective features.

[0040] It can be understood that the embodiment of the present application utilizes the multi-scale time-frequency analysis characteristics through the wavelet transform filtering algorithm to perform multi-dimensional noise suppression and outlier removal on the dynamically collected 5G signal parameters and environmental data: by decomposing the original signal into different frequency sub-bands, high-frequency noise and low-frequency drift are identified and suppressed, while retaining the signal mutation characteristics; when reconstructing the signal in the wavelet domain, the threshold processing algorithm is combined to remove abnormal data points outside the 3σ principle, reduce the data standard deviation, and improve the stability of the spatiotemporal series data. The standardized feature vector generated after calibration eliminates the impact of the dimensional differences of different parameters on subsequent analysis, and provides the Transformer-based coverage assessment model with input data with low noise pollution and complete features, thereby improving the accuracy of blind spot recognition and interference type classification.

[0041] For example, Figure 4As shown in the figure, during low-altitude testing in an industrial park, signal strength data collected by a drone was subject to interference from electromagnetic equipment within the factory, resulting in frequent ±15dB fluctuations. Furthermore, the wind speed sensor generated low-frequency drift noise of 0.8m / s due to vibration. Applying a wavelet transform filtering algorithm, the signal was first decomposed into eight wavelet subbands. Soft thresholding was used in high-frequency subbands (above 100Hz) to suppress electromagnetic interference noise, and in low-frequency subbands (below 0.1Hz) to remove sensor drift errors. The reconstructed signal outliers decreased from 12% to 2%, and the signal-to-noise ratio (SNR) was improved by 18dB. After inputting the calibrated standardized feature vectors into the coverage assessment model, the number of false positive blind spots due to equipment interference in the area was reduced by 60%, and the goodness-of-fit of the signal attenuation-to-terrain correlation analysis increased from R²=0.72 to 0.91, demonstrating the practical value of the wavelet transform filtering algorithm in effectively purifying data and improving analysis accuracy in complex industrial environments.

[0042] In the embodiment of the present application, a wavelet transform filtering algorithm is used to calibrate the 5G signal parameters and environmental data. The specific steps are as follows: Step C1: performing wavelet decomposition on the collected 5G signal parameters and environmental data to obtain sub-signals at different frequency levels; Step C2: Adopting an adaptive threshold algorithm to remove noise for sub-signals at different frequency levels. The formula of the adaptive threshold algorithm is: ,in, is the sub-signal standard deviation, N is the signal length, is the natural logarithm, is the threshold; Step C3: Perform wavelet reconstruction on the denoised sub-signal to obtain calibrated 5G signal parameters and environmental data, and generate a standardized feature vector based on the calibrated 5G signal parameters and environmental data.

[0043] Among them, the adaptive threshold algorithm is a method that dynamically calculates the segmentation threshold of each pixel or region based on the statistical characteristics of the local area of ​​the data. The formula of the adaptive threshold algorithm is: ,in, is the sub-signal standard deviation, N is the signal length, is the natural logarithm, is the threshold.

[0044] It can be understood that the embodiment of the present application dynamically calculates the denoising threshold through an adaptive threshold algorithm based on the local statistical characteristics of each frequency sub-signal after wavelet decomposition, and realizes refined noise suppression according to the distribution differences of noise at different frequency levels: automatically raising the threshold in sub-signals with high noise intensity to enhance the denoising strength, and lowering the threshold in sub-signals with rich signal features to avoid feature loss. By dynamically adapting local data features, it balances noise removal and signal detail retention, fully retains the key features of signal propagation, and improves the coverage assessment model's recognition accuracy and robustness for weak signal areas and interference types.

[0045] For example, Figure 5 As shown in the figure, in the low-altitude test in the mountainous area, the signal-to-noise ratio data collected by the drone is affected by both the high-frequency noise (standard deviation σ=15dB) generated by the reflection of the mountain and the low-frequency noise (standard deviation σ=5dB) caused by the temperature drift of the pressure sensor. When using the adaptive threshold algorithm, according to the formula Dynamic threshold calculation: A threshold of approximately 40dB is generated for high-frequency sub-signals (N=2000, σ=15), effectively filtering out sudden noise bursts. A threshold of approximately 12dB is generated for low-frequency sub-signals (N=5000, σ=5), avoiding over-filtering of slowly varying valid signals. After denoising, the proportion of outliers in the data dropped from 18% to 3%, and the overall signal-to-noise ratio improved by 25dB. Compared to a fixed threshold algorithm (uniformly set at 20dB), this algorithm fully preserves the signal gradient characteristics caused by valley terrain (such as the 3dB signal attenuation for every 100 meters of altitude increase) while accurately eliminating extreme noise points. The calibrated, standardized feature vectors were input into the coverage assessment model, reducing the false positive rate in blind spots caused by complex terrain from 22% to 5%, validating the advantages of the adaptive threshold algorithm in dynamic adaptation and intelligent denoising in heterogeneous noise environments.

[0046] In step A5, a Transformer-based coverage assessment model is constructed based on the standardized feature vector and combined with the three-dimensional digital map information to identify low-altitude coverage blind spots and interference types.

[0047] Among them, the Transformer-based coverage assessment model is an intelligent analysis model that uses the self-attention mechanism of the Transformer architecture to model spatiotemporal sequences such as 5G signal parameters and environmental data to evaluate signal coverage quality, identify blind spots and interference types.

[0048] It is understandable that the embodiment of the present application deeply mines the spatiotemporal dependencies in the standardized feature vector through the self-attention mechanism - capturing the signal fluctuations of the same grid point at different times, modeling the spatial correlation between different grid points, and realizing a multi-dimensional evaluation of the low-altitude coverage quality. Through global context modeling, it accurately identifies the coverage blind spots and interference types in complex scenarios, and the blind spot positioning accuracy is improved by more than 25% compared with sequence models such as LSTM. At the same time, the model supports end-to-end learning, automatically fitting the nonlinear mapping of signal propagation and environmental characteristics, providing a refined decision-making basis for network optimization, and improving the intelligence level of low-altitude coverage evaluation and network optimization efficiency.

[0049] For example, Figure 6 As shown in the figure, in a low-altitude coverage assessment in a hilly city, a standardized feature vector (a spatiotemporal sequence of 5,000 grid points) containing 12-dimensional features such as signal strength, neighboring cell interference power, and altitude was fed into a Transformer-based model. Using a self-attention mechanism, the model captured the spatial dependencies between grid points in an industrial park and those in surrounding high-rise buildings within 300 meters. The model identified an 85% probability of a blind spot with signal strength below -105dBm due to the high building density (0.9) and the non-line-of-sight path for base station signals. It also accurately identified 70% of interference from overcoverage from co-frequency neighboring base stations 2 kilometers away. Field tests showed that the model achieved 92% accuracy in blind spots in complex terrain, a 30% improvement over traditional LSTM models. The resulting optimization recommendations (such as adjusting the transmit power of three neighboring base stations to 46dBm and lowering the antenna tilt by 8°) increased the average signal strength in the target area by 10dB and reduced the user device drop rate from 15% to 4%, demonstrating the model's significant advantages in spatiotemporal feature modeling and complex scenario analysis.

[0050] In the embodiment of the present application, a Transformer-based coverage assessment model is constructed based on the standardized feature vector and combined with the three-dimensional digital map information. The specific steps are as follows: Step F1: Encode the terrain, buildings, and signal feature vectors of the three-dimensional digital map to generate a token sequence with spatial location information; Step F2: Based on the token sequence, the multi-head attention mechanism of the Transformer model is used to capture the signal propagation dependency between different grid points. The spatial position of the grid points is represented by position encoding. The formula is: ,where Q, K, V are query, key, and value matrices; is the dimension of the key vector; Step F3: Based on the spatial position of the grid points, the attention output is feature fused through a feedforward neural network, the model is trained using the gradient normalization method, and the coverage blind area probability distribution and interference type classification results are output.

[0051] Among them, the token sequence is a linear sequence formed by arranging the basic semantic units generated by processing text, speech and other data through word segmentation and feature discretization.

[0052] It can be understood that the embodiment of the present application encodes spatial features such as terrain and buildings in a three-dimensional digital map and signal feature vectors into a unified linear sequence format suitable for processing by the Transformer model through token sequences, and gives each token clear spatial position information and feature semantics through operations such as discretization and position embedding, providing structured input for the multi-head attention mechanism, so that the model can efficiently capture the signal propagation dependency between different grid points, and at the same time, through the position encoding formula Explicitly incorporating spatial location information avoids the defect of traditional sequence models that ignore spatial distances, converts complex three-dimensional scene features into sequence units that can be processed by Transformer, preserves the integrity of spatial features such as terrain undulations and building layouts, and dynamically models the propagation correlation of signals between different grid points through the attention weights between tokens. This enables the model to accurately identify coverage blind spots and interference types caused by spatial location relationships, improving the accuracy of blind spot positioning compared to models that do not encode spatial information.

[0053] For example, in an urban low-altitude coverage assessment, 3D map information for 3,000 grid points (e.g., 200 meters above sea level, 0.8 building density, 500 meters from a base station) and signal characteristics (signal strength -98dBm, neighboring cell interference power -85dBm) are encoded into a sequence of tokens containing spatial coordinates (X=1200, Y=800). Each token carries 10-dimensional features (location + environment + signal parameters). After embedding the grid point's geographic coordinates into the token sequence through positional encoding, the Transformer model's multi-head attention mechanism captures a strong correlation between a grid point in a commercial area (token 567) and six grid points within 200 meters of a densely populated area with high-rise buildings (tokens 500-570). The attention weights between these tokens show that for every 0.1 increase in building density, the predicted signal strength decreases by 2.5dB. This allows the model to accurately identify signal blind spots (coverage probability <30%) caused by surrounding buildings in this commercial area and to isolate 65% of interference from diffraction interference from a co-frequency base station 1 kilometer away. Compared with traditional models that do not use token sequence encoding, this method improves the blind spot positioning accuracy in complex building areas from 75% to 90%, and the interference type classification accuracy is improved by 25%. After actual measurement and optimization, the signal coverage compliance rate in the area is increased from 60% to 85%.

[0054] In the embodiment of the present application, the method for training the coverage assessment model using the gradient clipping method in step F3 includes: Step F301: Collect historical test data including three-dimensional digital map data (topography, building distribution), sensor signal parameters (RSRP, SINR, RSSI, etc.), and manually annotated coverage blind spot locations and interference type labels to construct a training sample set; Step F302: Design a Transformer-based shared encoder to extract signal propagation dependencies between grid points, combined with a task-specific decoder: The coverage quality prediction branch uses a "fully connected layer + BatchNorm + ReLU activation" structure, and the output layer uses Softmax to classify coverage levels or directly regress signal strength. The interference type classification branch uses a multi-head attention layer to capture interference features, and the output layer uses a Sigmoid activation function to achieve multi-label classification (each interference type is predicted independently). The underlying feature extraction is shared, and specific prediction results are output for different tasks, balancing model parameter efficiency and task adaptability. Step F303: Initialize model parameters using the Xavier / Glorot method. Training samples are batched into the model. After feature extraction via a shared encoder, each task branch independently performs forward propagation and calculates losses. Backward propagation uses the AdamW optimizer to update parameters. Optimization strategies include: calculating the weighted sum of all task losses during backpropagation; applying gradient clipping to global gradients to limit the gradient norm to a threshold (e.g., 1.0) to prevent gradient explosion; dynamically adjusting the learning rate using cosine annealing learning rate decay; using an early stopping mechanism based on validation set performance to prevent overfitting; and improving model ensemble techniques (e.g., voting) to enhance prediction stability in complex scenarios and robustness in low-altitude coverage testing.

[0055] Among them, the AdamW optimizer is an optimization algorithm that combines the Adam adaptive moment estimation optimization method with decoupled weight decay.

[0056] It can be understood that the embodiments of the present application control the complexity of model parameters by decoupling weight decay, alleviate the overfitting problem caused by the deep network and multi-task shared parameters of the Transformer model, and suppress the growth of parameter redundancy through independent weight decay terms during the training of high-dimensional spatiotemporal features such as three-dimensional maps and signal parameters, thereby reducing the generalization error of the model in complex scenarios; at the same time, combined with gradient clipping and cosine annealing learning rate decay, the gradient update during training is stabilized.

[0057] In step A6, if a low-altitude coverage blind spot or abnormal interference is detected, the A-star algorithm is used to dynamically adjust the drone's flight path, test the target area, and generate a coverage optimization plan based on the test results and push it to the operation and maintenance platform.

[0058] Among them, the A-star algorithm is a graph search algorithm based on heuristic functions.

[0059] It can be understood that the embodiment of the present application dynamically plans the optimal flight path of the UAV from the current position to the target area through a heuristic function. For the blind spots or interference areas detected in real time during the low-altitude coverage test, the shortest or lowest energy consumption path is quickly calculated under the premise of considering terrain restrictions and flight constraints. By balancing the actual flight cost and the heuristic estimated cost, the UAV is ensured to re-survey the target area with the optimal trajectory, avoiding redundant flights, providing more accurate measured data for the coverage optimization plan, and at the same time ensuring the flight safety and mission execution efficiency of the UAV in complex urban low-altitude environments, shortening the cycle from coverage problem location to optimization plan generation.

[0060] In the embodiment of the present application, if a low-altitude coverage blind spot or abnormal interference is detected, the A-star algorithm is used to dynamically adjust the drone's flight path to test the target area. The specific steps are as follows: Step E1: Construct a weighted directed graph G = (V, E), where the node V is the test grid point, the edge E is the feasible path of the drone, and the edge weight is the weighted sum of the flight time and the signal mutation risk. The weight formula is: ,in, is the flight time of edge e, is the signal mutation risk value corresponding to edge e, and is the normalized weight coefficient ( ), and is the global historical maximum of flight time and risk; Step E2: Based on historical test data, a random forest classifier is used to construct a node risk assessment model, perform model training, and output the probability of a coverage issue at node v. The node risk assessment model training process includes: extracting feature vectors such as signal strength fluctuation, signal-to-noise ratio change rate, terrain relief, and base station density; optimizing hyperparameters using 5-fold cross-validation; and evaluating model performance using the AUC-ROC curve. Step E3: Based on the probability of the coverage problem, dynamically adjust the edge weights and plan the path based on the A-star algorithm. The edge weight adjustment formula is: ,in, is the adjusted edge weight; is the original edge weight; is the adjustment coefficient; is the probability of coverage problem at the terminal node; at the same time, the heuristic function is designed: ,in, is the comprehensive evaluation function value of node v; is the actual cost from the starting point to node v; is the heuristic function; is the probability that node v has a coverage problem; For additional risk weights, paths in high-risk areas are prioritized for testing, and multi-objective optimality is ensured through real-time weight updates.

[0061] Among them, the AUC-ROC curve is a curve that plots the relationship between the true positive rate and the false positive rate of the classification model under different classification thresholds.

[0062] It can be understood that the embodiment of the present application draws a curve showing the relationship between the true positive rate and the false positive rate to intuitively reflect the comprehensive discrimination performance of the model for complex features such as signal strength fluctuations and terrain undulations. The larger the AUC value, the stronger the model's ability to distinguish between high-risk areas and normal areas. The high-performance model screened out by the AUC-ROC curve accurately identifies high-risk areas and dynamically adjusts the risk factor in the edge weight formula, so that the drone path planning gives priority to covering blind spots or interference hotspots that really need re-testing, thereby reducing invalid flights and improving the efficiency of low-altitude testing and the accuracy of problem location.

[0063] For example, in low-altitude testing in urban areas with dense high-rise buildings, when training a random forest node risk assessment model based on 10-dimensional features such as the signal-to-noise ratio change rate (>15dB / s) and base station density (>5 / km²), the initial AUC-ROC curve showed an AUC of 0.80, with the model misidentifying blind spots (signal strength <-115dBm) on the shady sides of high-rise buildings as normal areas (with a false positive rate of 40%). By optimizing the minimum number of leaf node samples to 5 through 5-fold cross-validation and introducing class weights to balance positive and negative samples, the AUC improved to 0.95. The curve achieved a true positive rate of over 90% when the false positive rate was ≤20%, accurately identifying high-risk nodes due to building obstruction (for example, the probability of coverage issues at a certain grid point was corrected from 0.55 to 0.92). Based on the optimized model output, the A-star algorithm prioritized planning paths around 12 high-risk nodes identified by high AUCs. This increased the drone's coverage of actual blind spots from 70% to 95%, reduced the flight distance around ineffective low-risk areas by 40%, and reduced the number of missed signal blind spots in the area from 8 to 1. This validates the core value of the AUC-ROC curve in quantifying model performance, driving accurate risk assessment, and efficient path planning in complex urban environments.

[0064] In the embodiment of the present application, a coverage optimization plan is generated based on the test results and pushed to the operation and maintenance platform. The specific steps are as follows: Step G1: Based on the blind spot location and interference type output by the coverage assessment model, combined with base station parameters (transmit power, antenna tilt) and terrain characteristics, a multi-dimensional optimization parameter set is generated; Step G2: Use particle swarm optimization algorithm , in, is the velocity of particle i in generation t; is the inertia weight; , is the learning factor (acceleration constant); , is a random number uniformly distributed in the interval (0,1); is the best historical position of particle i; is the best historical position of the entire particle swarm; For the position of particle i in the tth generation, perform global optimization on the multi-dimensional optimization parameter set to obtain the optimization solution; Step G3: Convert the optimization plan into base station parameter configuration instructions, push them to the operation and maintenance platform in real time through the edge computing node, and perform remote one-click parameter distribution.

[0065] Among them, the particle swarm optimization algorithm is a group intelligence optimization algorithm that simulates the collaborative behavior of biological groups such as bird flocks and fish schools. The particle swarm optimization algorithm formula is: , in, is the velocity of particle i in generation t; is the inertia weight; , is the learning factor (acceleration constant); , is a random number uniformly distributed in the interval (0,1); is the best historical position of particle i; is the best historical position of the entire particle swarm; is the position of particle i in the tth generation.

[0066] It is understandable that the embodiments of the present application efficiently search for the global optimal solution in the high-dimensional solution space by simulating the collaboration of particle groups: using the speed update formula and the group learning factor to balance exploration and development, avoiding the traditional gradient method from falling into the local optimum, and quickly optimizing the multi-base station coordination parameters for low-altitude scenes with complex terrain and strong parameter correlation, so that the base station transmission power adjustment accuracy reaches ±0.5dBm, the antenna tilt step size is optimized, and the blind spot coverage probability is reduced. By dynamically adapting the nonlinear relationship between the group intelligently, the average signal strength in the target area is improved, and the network optimization efficiency is improved.

[0067] For example, Figure 7As shown in the figure, in the optimization of low-altitude coverage of wind power bases in mountainous areas, the drone cruise found that there were 8 signal interruption areas (signal strength <-120dBm) between the wind turbines due to mountain obstruction. It was necessary to coordinate the optimization of the transmission power (40-50dBm), antenna azimuth (90°-270°) and vertical lobe downtilt (8°-20°) of the three surrounding base stations, a total of 9 parameters. When using the particle swarm optimization algorithm, each particle encodes a set of base station parameter combinations (such as base station A power 48dBm, azimuth 220°, base station B downtilt 15°), and the speed formula is used to calculate the transmission power (40-50dBm), antenna azimuth (90°-270°) and vertical lobe downtilt (8°-20°). , Iterative search: Inertia weights were set to balance solution space fluctuations caused by complex terrain, and learning factors were used to reinforce the group's optimal experience (for example, parameter combinations that avoid signal distortion caused by ridge reflections). After 25 iterations, the algorithm output the optimal solution within 8 minutes: the azimuth of one mountaintop base station was adjusted to 180° to cover the valley channel, and the downtilt angles of two mountainside base stations were increased to 18° and their power was boosted by 2dB to penetrate vegetation. This algorithm reduced the search time by 40% compared to the simulated annealing algorithm and avoided the local optimal trap caused by multipath reflections in the valley. Retests conducted by drones showed an average 20dB increase in signal strength in blind spots, and the coverage rate increased from 35% to 85%. After the operation and maintenance platform was adjusted according to the solution, the packet loss rate of remote control signals for wind turbines in the area dropped from 25% to 3%.

[0068] According to a 5G network signal low-altitude coverage test method proposed in an embodiment of the present application, an intelligent upgrade of the entire process of 5G low-altitude coverage test is realized through the integration of multi-dimensional technologies. A high-precision three-dimensional digital map is constructed based on the geographic information of the target area and the base station layout, and a dynamic test grid is divided, so that the drone formation can perform grid-based precise coverage according to the terrain characteristics and base station distribution. Compared with traditional manual or fixed equipment testing, the test coverage integrity and regional adaptability in complex scenarios (such as high-rise buildings, mountains, etc.) are improved; a multi-source sensor group is equipped to synchronously collect 5G signal parameters (signal strength, signal-to-noise ratio, neighboring area interference power) and environmental data (altitude, wind speed, building density), realizing the coupling analysis of signal characteristics and physical environment, and providing multi-dimensional data support for subsequent modeling; the original data is denoised and calibrated through the wavelet transform filtering algorithm, and outliers and noise interference (such as sudden electromagnetic pulses, sensor errors) are eliminated, so that the data reliability is improved by more than 30%, providing a basis for precision. The Transformer architecture lays the foundation for accurate analysis. It introduces the Transformer architecture to construct a coverage assessment model, leveraging its powerful long-distance dependency modeling capabilities to deeply explore the nonlinear relationship between signal propagation and three-dimensional spatial characteristics (building obstruction, terrain undulation). This model can accurately identify low-altitude coverage blind spots (such as the signal shadow area between two tall buildings) and interference types (co-channel interference, spurious interference) with millimeter-level precision, achieving classification accuracy exceeding 92%. For detected coverage issues, the A-star algorithm dynamically optimizes drone flight paths, enabling adaptive retesting of suspicious areas. This avoids redundant testing or blind spot omissions associated with traditional fixed paths, improving testing efficiency by 40%. The generated coverage optimization solutions (such as base station power adjustment, antenna tilt optimization, and new micro-station deployment recommendations) combine measured data with spatial propagation models, providing the operation and maintenance platform with a quantifiable, traceable, and scientifically sound basis for decision-making. This shortens network optimization cycles and improves the coverage uniformity and stability of 5G signals in low-altitude scenarios. This addresses the challenges of low testing efficiency and lagging data processing and analysis in existing technologies.

[0069] The following is an example of a 5G network signal low-altitude coverage test method. Figure 8 Shown, including: Through the Urban Planning Bureau, we obtained DEM elevation data with an accuracy of 1 meter, 3D building models including height / coordinates / materials, and vegetation distribution vector maps. We also obtained parameters such as the base station latitude and longitude, azimuth (120° / 180° / 240°), and downtilt (10°) with an accuracy of 0.0001° from the operator's network management system. We determined that the five base stations were distributed in a ring with a spacing of 800-1000 meters. Based on the 3GPPUMi model, our estimated theoretical coverage radius was approximately 600 meters, providing high-precision geographic and base station basic data for subsequent 3D modeling and coverage analysis.

[0070] DEM data is integrated to generate terrain surfaces, and building models up to 150 meters high and base station coordinates are imported to construct a three-dimensional grid scene that includes altitude and building density (such as 0.8 for residential areas and 0.2 for parks). The initial grid side length is set at 1 / 3 of the theoretical coverage radius (200 meters), and the 500×500 meter area is divided into 25 basic grids. For residential areas with building density >0.7 and hilly areas with terrain undulation >50 meters, the area is divided more densely with 1 / 2 of the basic side length (100 meters), generating 120 encrypted grids. Finally, a dynamic grid system consisting of 145 test grids is formed to meet the signal coverage test accuracy requirements of different scenarios.

[0071] The drone is equipped with a GNSS module with a positioning accuracy of ±0.5 meters, a 5G signal collector that supports a 10Hz sampling rate (collecting RSRP / SINR / neighboring cell interference power), and an anemometer with an accuracy of ±0.1m / s. It flies in a zigzag pattern at an altitude of 150 meters and a speed of 10m / s, collecting data every 20 seconds (corresponding to a flight distance of approximately 200 meters). Utilizing a high-precision clock model with an initial clock deviation of ≤1μs and a frequency deviation coefficient of 1e-8, the drone adds millisecond-accurate UTC timestamps and latitude and longitude coordinates to the data. By using an alignment algorithm that eliminates data with time deviations >5ms, it achieves spatiotemporal synchronization calibration of multi-sensor data.

[0072] Perform a three-layer wavelet decomposition on the RSRP data in the range of -140~-60dBm, and use an adaptive threshold for the high-frequency sub-signal (noise-dominated, standard deviation σ=12dB) The system uses a 1000 sampling point algorithm (N=1000, T≈32dB) to filter out burst noise, while retaining the signal trend of low-frequency sub-signals (σ=5dB, T≈11dB). After denoising, the proportion of RSRP outliers (deviations from the mean >3σ) is reduced from 15% to 4%. Fifteen dimensional features, including signal strength, signal-to-noise ratio, and altitude, are normalized to [-1, 1] using the Z-score to generate standardized feature vectors, providing high-quality data for model input.

[0073] For each grid point, terrain features such as altitude and building density, as well as signal parameters such as RSRP and SINR, and coordinates (X, Y) are encoded as 20-dimensional tokens and arranged spatially into a 145-length sequence. A sinusoidal position code with a period of 1000 is added to represent the spatial location. Based on 2000 sets of historically annotated data (blind spot labels: RSRP < -105dBm and no valid neighboring cells, interference types include co-channel interference and terrain obstruction), the training set / validation set was divided into 8:2. The model was trained using the AdamW optimizer (weight decay 0.01), a gradient clipping threshold of 1.0, and a cosine annealing learning rate ranging from an initial 1e-4 to a minimum of 1e-6. After 30 rounds of iterations, the model achieved a blind spot identification F1-score of 0.85 and an interference type classification accuracy of 82%. The model then outputs the probability distribution of coverage blind spots and the interference type classification results.

[0074] Based on 8-dimensional features such as signal fluctuation >10dB and terrain undulation >30 meters, a random forest node risk model (AUC-ROC=0.92) was trained to identify 23 high-risk nodes with a coverage problem probability >0.7 (such as a probability of 0.89 for a central grid point in a residential area); a weighted directed graph was constructed with the edge weight being the weighted sum of flight time and risk value (for example, if the flight time is 12 seconds per 100 meters, the edge weight of the high-risk node increases by 30%). The planned route shortened the flight distance by 18% compared to the initial route (1200 meters to 980 meters). For two blind spots shadowed by tall buildings, a three-dimensional parameter set was generated, including base station transmit power (46-50 dBm), downtilt angle (10°-15°), and azimuth angle (180°-210°). Using a particle swarm algorithm with an inertia weight of 0.7 and a learning factor c1=c2=1.8, the optimal solution was obtained: base station A power of 48 dBm, base station B downtilt angle of 13°, and base station C azimuth angle of 200°. This solution was pushed to the operation and maintenance platform via the edge computing node, and the configuration was completed within 30 minutes. Retests showed an average 15 dB increase in RSRP in the blind spots, and a rise in coverage compliance from 55% to 88%.

[0075] In summary, high-precision geographic and base station data were acquired to construct a three-dimensional grid scenario (with a basic grid of 200 meters and a grid density of 100 meters in complex areas). Data was collected using drone-mounted sensors, and standardized features were generated through spatiotemporal alignment and wavelet filtering (reducing outliers from 15% to 4%). Coverage issues were analyzed using the Transformer model (blind spot identification F1=0.85), and routes in high-risk areas were planned using the A-star algorithm (shortening paths by 18%). Finally, a particle swarm optimization approach was used to optimize base station parameters (such as a 2dB power increase), completing configuration within 30 minutes. The measured blind spot signal strength increased by 15dB, and the coverage compliance rate rose from 55% to 88%, significantly improving the positioning accuracy and efficiency of resolving coverage issues in low-altitude scenarios, while reducing the risk of failures and operation and maintenance costs.

[0076] Next, a 5G network signal low-altitude coverage test system proposed according to an embodiment of the present application is described with reference to the accompanying drawings.

[0077] Specifically, Figure 9 A schematic diagram of the composition of a 5G network signal low-altitude coverage test system provided in an embodiment of the present application.

[0078] like Figure 9 As shown, the 5G network signal low-altitude coverage test system 10 includes: an acquisition module 100, a division module 200, a collection module 300, a calibration module 400, an identification module 500, and a generation module 600.

[0079] Among them, the acquisition module 100 is used to obtain the geographic information of the target area and the base station layout information; the division module 200 is used to construct a three-dimensional digital map based on the geographic information of the target area and the base station layout, divide the dynamic test grid, and determine the initial UAV flight path. The acquisition module 300 is used to collect 5G signal parameters and environmental data of each grid point at a preset frequency based on the multi-source sensor group carried by the low-altitude test UAV and the initial UAV flight path. The signal parameters include signal strength, signal-to-noise ratio, and neighboring cell interference power, and the environmental data includes altitude, wind speed, and building density. The calibration module 400 is used to calibrate the 5G signal parameters and environmental data using a wavelet transform filtering algorithm, remove outliers, and generate a standardized feature vector. The identification module 500 is used to construct a Transformer-based coverage assessment model based on the standardized feature vector and the three-dimensional digital map information to identify low-altitude coverage blind spots and interference types. The generation module 600 is used to dynamically adjust the UAV flight path using the A-star algorithm if a low-altitude coverage blind spot or abnormal interference is detected, test the target area, generate a coverage optimization plan based on the test results, and push it to the operation and maintenance platform.

[0080] It should be noted that the above explanation of an embodiment of a 5G network signal low-altitude coverage test method is also applicable to a 5G network signal low-altitude coverage test system of this embodiment, and will not be repeated here.

[0081] According to the embodiment of the present application, a 5G network signal low-altitude coverage test system is proposed. Through the integration of multi-dimensional technologies, an intelligent upgrade of the entire process of 5G low-altitude coverage test is realized. A high-precision three-dimensional digital map is constructed based on the geographic information of the target area and the base station layout, and a dynamic test grid is divided, so that the drone formation can perform grid-based precise coverage according to the terrain characteristics and base station distribution. Compared with traditional manual or fixed equipment testing, the test coverage integrity and regional adaptability in complex scenarios (such as high-rise buildings, mountains, etc.) are improved; a multi-source sensor group is equipped to synchronously collect 5G signal parameters (signal strength, signal-to-noise ratio, neighboring area interference power) and environmental data (altitude, wind speed, building density), realizing the coupling analysis of signal characteristics and physical environment, and providing multi-dimensional data support for subsequent modeling; the original data is denoised and calibrated through the wavelet transform filtering algorithm, and outliers and noise interference (such as sudden electromagnetic pulses, sensor errors) are eliminated, so that the data reliability is improved by more than 30%, which is accurate. The analysis lays the foundation; the Transformer architecture is introduced to build a coverage assessment model. Leveraging its powerful long-distance dependency modeling capabilities, it deeply explores the nonlinear relationship between signal propagation and three-dimensional spatial characteristics (building obstruction, terrain undulation). This model can accurately identify low-altitude coverage blind spots (such as the signal shadow area between two tall buildings) and interference types (co-channel interference, spurious interference) with millimeter-level precision, achieving a classification accuracy of over 92%. For detected coverage issues, the A-star algorithm dynamically optimizes the drone's flight path, enabling adaptive retesting of suspicious areas. This avoids redundant testing or blind spot omissions associated with traditional fixed paths, improving testing efficiency by 40%. The generated coverage optimization solutions (such as base station power adjustment, antenna tilt optimization, and new micro-station deployment recommendations) combine measured data with spatial propagation models, providing the operation and maintenance platform with a quantifiable, traceable, and scientific basis for decision-making. This shortens network optimization cycles and improves the coverage uniformity and stability of 5G signals in low-altitude scenarios. This addresses the problems of low testing efficiency and lagging data processing and analysis in existing technologies.

[0082] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or N embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification and features of different embodiments or examples without contradiction.

[0083] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of technical features indicated. Thus, a feature specified as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of this application, "N" means at least two, for example, two, three, etc., unless otherwise specifically defined.

[0084] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, fragment or portion of code comprising one or more executable instructions for implementing the steps of a custom logical function or process, and the scope of the preferred embodiments of the present application includes alternative implementations in which functions may be performed out of the order shown or discussed, including performing functions in a substantially simultaneous manner or in reverse order depending on the functions involved, which should be understood by those skilled in the art to which the embodiments of the present application belong.

[0085] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiment, the N steps or methods can be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof can be used: a discrete logic circuit having logic gate circuits for implementing logical functions on data signals, an application-specific integrated circuit having suitable combinational logic gate circuits, a programmable gate array (PGA), a field-programmable gate array (FPGA), etc.

[0086] Those skilled in the art will appreciate that all or part of the steps in the method for implementing the above-mentioned embodiment can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium. When the program is executed, it includes one or a combination of the steps of the method embodiment.

Claims

1. A 5G network signal low-altitude coverage test method, characterized in that: The following steps are involved: Step A1: Obtaining target area geographic information and base station layout information; Step A2: Based on the geographic information of the target area and the base station layout, a three-dimensional digital map is constructed, a dynamic test grid is divided, and an initial UAV flight path is determined; Step A3: Based on the multi-source sensor group carried by the low-altitude test drone and the initial drone flight path, 5G signal parameters and environmental data of each grid point are collected at a preset frequency, wherein the signal parameters include signal strength, signal-to-noise ratio, and neighboring cell interference power; and the environmental data includes altitude, wind speed, and building density. Step A4: calibrate the 5G signal parameters and environmental data using a wavelet transform filtering algorithm to remove outliers and generate a standardized feature vector; Step A5: constructing a Transformer-based coverage assessment model based on the standardized feature vector and the three-dimensional digital map information to identify low-altitude coverage blind spots and interference types; Step A6: If the low-altitude coverage blind spot or abnormal interference is detected, the UAV flight path is dynamically adjusted using the A-star algorithm, the target area is tested, and a coverage optimization plan is generated based on the test results and pushed to the operation and maintenance platform.

2. A 5G network signal low-altitude coverage test method according to claim 1, characterized in that: Construct a 3D digital map and divide the dynamic test grid. The specific steps are as follows: Step B1, obtaining DEM elevation data, building three-dimensional model data and base station longitude and latitude coordinates; Step B2: Based on the DEM elevation data, the three-dimensional building model data and the latitude and longitude coordinates of the base station, the surface features such as the terrain undulation and the building height are integrated to construct a three-dimensional digital map with the terrain features; Step B3: Based on the three-dimensional digital map, combined with the base station transmission power, working frequency band and 3GPPUMi model ,in, is the path loss in the line-of-sight scenario; The three-dimensional distance (Unit: meter, m) The path loss increases by 22 dB for every 10-fold increase in distance. is a fixed constant term; The carrier frequency (Unit: GHz) For each 10-fold increase in frequency, the path loss increases by 20 dB. Calculate the theoretical coverage radius and use 1 / 3 to 1 / 2 of the coverage radius as the grid side length to initially divide the target area into grids. Step B4: For areas where signals are easily blocked, such as densely built areas and areas with complex terrain, local encryption division is performed based on the building density and terrain undulation in the area, according to 1 / 2 to 1 / 4 of the initial grid side length, to generate a dynamic test grid that adapts to different scene complexities.

3. A 5G network signal low-altitude coverage test method according to claim 1, characterized in that: The 5G signal parameters and environmental data are calibrated using a wavelet transform filtering algorithm. The specific steps are as follows: Step C1: performing wavelet decomposition on the collected 5G signal parameters and environmental data to obtain sub-signals at different frequency levels; Step C2: Adopting an adaptive threshold algorithm to remove noise for the sub-signals at different frequency levels. The formula of the adaptive threshold algorithm is: ,in, is the sub-signal standard deviation, N is the signal length, is the natural logarithm, is the threshold; Step C3: Perform wavelet reconstruction on the denoised sub-signal to obtain calibrated 5G signal parameters and environmental data, and generate a standardized feature vector based on the calibrated 5G signal parameters and environmental data.

4. The 5G network signal low-altitude coverage test method according to claim 1, wherein the 5G signal parameters and environmental data of each grid point are collected at a preset frequency based on the multi-source sensor group carried by the low-altitude test drone and the initial drone flight path. The specific steps are as follows: Step D1, constructing a high-precision clock model; Step D2: Using the GNSS positioning module of the drone, combined with the high-precision clock model ,in, is the time value of the local clock at time t; The standard time value at time t provided by the GNSS module; is the initial clock bias; is the frequency deviation coefficient; Get real-time location and time information for the initial reference moment; Step D3: Based on the position and time information and the initial flight path, add time and space stamps to the signal parameters and environmental data collected by each sensor, and perform time alignment and spatial matching on the data of different sensors based on the time and space stamps.

5. A 5G network signal low-altitude coverage test method according to claim 1, characterized in that: If the low-altitude coverage blind spot or abnormal interference is detected, the A-star algorithm dynamically adjusts the drone's flight path to test the target area. The specific steps are as follows: Step E1: Construct a weighted directed graph G = (V, E), where the node V is the test grid point, the edge E is the feasible path of the drone, and the edge weight is the weighted sum of the flight time and the signal mutation risk. The weight formula is: ,in, is the flight time of edge e, is the signal mutation risk value corresponding to edge e, and is the normalized weight coefficient ( ), and is the global historical maximum of flight time and risk; Step E2: Based on historical test data, a random forest classifier is used to construct a node risk assessment model, perform model training, and output the probability of a coverage problem occurring at node v. The node risk assessment model training process includes: extracting feature vectors such as signal strength fluctuation, signal-to-noise ratio change rate, terrain relief, and base station density; optimizing hyperparameters using 5-fold cross-validation; and evaluating model performance using the AUC-ROC curve. Step E3: Based on the probability of the coverage problem, dynamically adjust the edge weight and plan the path based on the A-star algorithm. The edge weight adjustment formula is: ,in, is the adjusted edge weight; is the original edge weight; is the adjustment coefficient; is the probability of coverage problem at the terminal node; at the same time, the heuristic function is designed: ,in, is the comprehensive evaluation function value of node v; is the actual cost from the starting point to node v; is the heuristic function; is the probability that node v has a coverage problem; For additional risk weights, paths in high-risk areas are prioritized for testing, and multi-objective optimality is ensured through real-time weight updates.

6. A 5G network signal low-altitude coverage test method according to claim 1, characterized in that: Based on the standardized feature vector and the three-dimensional digital map information, a Transformer-based coverage assessment model is constructed. The specific steps are as follows: Step F1: Encode the terrain, buildings, and signal feature vectors of the three-dimensional digital map to generate a token sequence with spatial location information; Step F2: Based on the token sequence, the multi-head attention mechanism of the Transformer model is used to capture the signal propagation dependency between different grid points, and the spatial position of the grid points is represented by position encoding. The formula is: ,where Q, K, V are query, key, and value matrices; is the dimension of the key vector; Step F3: Based on the spatial position of the grid points, feature fusion is performed on the attention output through a feedforward neural network, and the model is trained using a gradient normalization method to output the coverage blind spot probability distribution and interference type classification results.

7. A 5G network signal low-altitude coverage test method according to claim 6, characterized in that: The method for training the coverage assessment model using the gradient clipping method in step F3 includes: Step F301: Collect historical test data including three-dimensional digital map data (topography, building distribution), sensor signal parameters (RSRP, SINR, RSSI, etc.), and manually annotated coverage blind spot locations and interference type labels to construct a training sample set; Step F302: Design a Transformer-based shared encoder to extract signal propagation dependencies between grid points, combined with a task-specific decoder: The coverage quality prediction branch uses a "fully connected layer + BatchNorm + ReLU activation" structure, and the output layer uses Softmax to classify coverage levels or directly regress signal strength. The interference type classification branch uses a multi-head attention layer to capture interference features, and the output layer uses a Sigmoid activation function to achieve multi-label classification (each interference type is independently predicted). The underlying feature extraction is shared, and specific prediction results are output for different tasks, balancing model parameter efficiency and task adaptability. Step F303: Initialize model parameters using the Xavier / Glorot method. Training samples are batched into the model. After feature extraction via a shared encoder, each task branch independently performs forward propagation and calculates losses. Backward propagation uses the AdamW optimizer to update parameters. Optimization strategies include: calculating the weighted sum of all task losses during backpropagation; applying gradient clipping to global gradients to limit the gradient norm to a threshold (e.g., 1.0) to prevent gradient explosion; dynamically adjusting the learning rate using cosine annealing learning rate decay; using an early stopping mechanism based on validation set performance to prevent overfitting; and improving model ensemble techniques (e.g., voting) to enhance prediction stability in complex scenarios and robustness in low-altitude coverage testing.

8. A 5G network signal low-altitude coverage test method according to claim 1, characterized in that: Generate a coverage optimization plan based on the test results and push it to the operation and maintenance platform. The specific steps are as follows: Step G1: Based on the blind spot location and interference type output by the coverage assessment model, combined with base station parameters (transmit power, antenna tilt) and terrain characteristics, a multi-dimensional optimization parameter set is generated; Step G2: Use particle swarm optimization algorithm , in, is the velocity of particle i in generation t; is the inertia weight; , is the learning factor (acceleration constant); , is a random number uniformly distributed in the interval (0,1); is the best historical position of particle i; is the best historical position of the entire particle swarm; For the position of particle i in the tth generation, perform global optimization on the multi-dimensional optimization parameter set to obtain an optimization solution; Step G3: Convert the optimization plan into base station parameter configuration instructions, push them to the operation and maintenance platform in real time through the edge computing node, and perform remote one-click parameter distribution.

9. A 5G network signal low-altitude coverage test system, characterized in that: include: An acquisition module is used to obtain geographic information of the target area and base station layout information; A division module is used to construct a three-dimensional digital map based on the geographic information of the target area and the base station layout, divide the dynamic test grid, and determine the initial UAV flight path; An acquisition module is configured to collect 5G signal parameters and environmental data at each grid point at a preset frequency based on the multi-source sensor group carried by the low-altitude test drone and the initial drone flight path, wherein the signal parameters include signal strength, signal-to-noise ratio, and neighboring cell interference power; and the environmental data includes altitude, wind speed, and building density; a calibration module, configured to calibrate the 5G signal parameters and environmental data using a wavelet transform filtering algorithm, remove outliers, and generate a standardized feature vector; An identification module is used to construct a Transformer-based coverage assessment model based on the standardized feature vector and the three-dimensional digital map information to identify low-altitude coverage blind spots and interference types; The generation module is used to dynamically adjust the UAV flight path through the A-star algorithm if the low-altitude coverage blind spot or abnormal interference is detected, test the target area, generate a coverage optimization plan based on the test results, and push it to the operation and maintenance platform.

Images