Active learning-based 5G wireless communication network drive test system

Abstract

The invention discloses a 5G wireless communication network drive test system based on active learning, which belongs to the technical field of mobile communication network test and optimization and comprises the following steps: acquiring a terminal QoE index set and mapping the terminal QoE index set into a three-dimensional space matrix; generating a dynamic drive test path according to the space matrix QoE gradient change; evaluating the drive test path measurement data to obtain a joint confidence map; screening an active labeling sample set according to the confidence map; and constructing an incremental diagnosis model output optimization decision instruction set based on the sample set. According to the invention, an active learning framework is adopted, and QoE space gradient analysis, two-dimensional confidence evaluation and incremental diagnosis modeling technologies are fused, so that dynamic intelligent planning of a 5G network drive test path, efficient screening of test data and accurate diagnosis of network problems can be realized; and the efficiency of network drive test, the accuracy of problem diagnosis and the response speed of network optimization are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of mobile communication network testing and optimization, and more particularly to a road test system for 5G wireless communication networks based on active learning. Background Art

[0002] The rapid deployment and application of 5G (the fifth generation of mobile communication technology) networks carry growing diverse service demands, such as ultra-high-definition video, virtual reality, and vehicle-to-everything. To ensure high-quality user experience (QoE) for these services and continuously optimize network performance, it is crucial to accurately and efficiently test and evaluate the actual deployed 5G networks. Traditional network road testing is a common means to obtain wireless signal coverage, interference, and service performance data in the real environment.

[0003] However, existing 5G network road testing methods often have some limitations. Traditional road testing usually adopts the method of presetting fixed routes or grid-based comprehensive coverage, which may lead to low testing efficiency, generating a large amount of redundant data in areas with good network quality, while insufficient testing investment in areas with frequent problems or dynamic changes. In addition, the data volume generated by large-scale road testing is huge, the subsequent analysis and processing are complex, and the cycle for feedback of test results to the network optimization link is long, making it difficult to quickly respond to real-time changes in network status and meet the requirements of refined operation and maintenance. Summary of the Invention

[0004] To solve the above problems, the present invention provides a road test system for 5G wireless communication networks based on active learning. By adopting an active learning framework and integrating QoE space gradient analysis, two-dimensional confidence evaluation, and incremental diagnostic modeling techniques, it can achieve dynamic intelligent planning of 5G network road test paths, efficient screening of test data, and accurate diagnosis of network problems, significantly improving the efficiency of network road testing, the accuracy of problem diagnosis, and the response speed of network optimization.

[0005] The above objectives can be achieved through the following solutions:

[0006] 5G wireless communication network road test system based on active learning, including a data acquisition and mapping module, which is used to acquire the QoE (Quality of Experience) metric set of terminal services in a preset area and map the QoE metric set into a three-dimensional rasterized space matrix; a path generation module, which is used to generate a dynamic road test path according to the QoE attenuation gradient change of the three-dimensional rasterized space matrix; a data evaluation module, which is used to perform two-dimensional confidence evaluation on the measurement data in the dynamic road test path to obtain a space-time joint confidence distribution map; a sample screening module, which is used to screen an active annotation sample set according to the fuzzy interval distribution of the joint confidence distribution map; a model construction and decision module, which is used to construct an incremental diagnosis model based on the active annotation sample set and generate and output a network optimization decision instruction set according to the incremental diagnosis model.

[0007] Optionally, the data acquisition and mapping module includes: a data extraction unit, which is used to extract end-to-end delay probe data from a preset base station interface; a data parsing unit, which is used to parse the video buffer event duration distribution in the delay probe data; a metric generation unit, which is used to generate a QoE metric set based on the video buffer event duration distribution and associate the handover request message density in a preset user movement trajectory; a space matrix mapping unit, which is used to map the QoE metric set into a three-dimensional rasterized space matrix.

[0008] Optionally, the path generation module includes: a coefficient of variation calculation unit, which is used to calculate the coefficient of variation of the signal strength per unit area in the QoE attenuation gradient change; a path node generation unit, which is used to generate multi-level optimization path nodes when the difference in the coefficient of variation of the signal strength between adjacent grids exceeds a preset threshold; a path construction unit, which is used to construct a topology structure including spatial weights according to the multi-level optimization path nodes to form a dynamic road test path.

[0009] Optionally, generating the multi-level optimization path nodes includes: analyzing the building penetration loss pattern by using a preset coverage hole prediction algorithm; marking potential signal shielding areas in the three-dimensional rasterized space matrix according to the loss pattern; using the boundary points of the signal shielding areas as the forced constraint conditions for generating priority path nodes.

[0010] Optionally, the data evaluation module includes: a spatial dimension analysis unit, which is used to analyze the measurement data in the dynamic road test path in the spatial dimension and count the signal quality fluctuation parameters reflecting the state of the micro base station overlapping coverage area; a time dimension analysis unit, which is used to analyze the measurement data in the dynamic road test path in the time dimension and extract the burst frequency change rate of the tracking area update (TAU) event; a confidence fusion unit, which is used to perform weighted fusion on the signal quality fluctuation parameters and the burst frequency change rate to obtain a joint confidence distribution map.

[0011] Optionally, the signal quality fluctuation parameter reflecting the state of the overlapping coverage area of the micro base stations includes: obtaining a reference signal received quality (RSRQ) measurement report set within a preset time period; establishing an RSRQ coefficient of variation matrix for the multi-cell joint coverage area based on the RSRQ measurement report set; identifying an interference-sensitive area set that meets the preset co-coverage condition in the RSRQ coefficient of variation matrix, and using the interference-sensitive area set as the signal quality fluctuation parameter.

[0012] Optionally, the model construction and decision-making module includes: a feature extraction unit for extracting PRACH preamble detection failure sequence features from the active annotation sample set; a training sample generation unit for associating the PRACH preamble detection failure sequence features with a preset TCP retransmission response rate time series map to generate a composite training sample set; a model construction unit for using the network state features as input and the network performance indicators as output, and establishing and training a neural network model using the composite training sample set to obtain an incremental diagnosis model; a model update unit for using a preset network optimization objective function as a constraint condition, and updating the incremental diagnosis model using a preset incremental learning algorithm and the composite training sample set; a decision generation and output unit for analyzing the network state based on the incremental diagnosis model and generating a network optimization decision instruction set for output.

[0013] Optionally, the extraction of the PRACH preamble detection failure sequence features from the active annotation sample set includes: constructing a time-domain and frequency-domain joint distribution template based on a preset preamble sequence; comparing the matching degree distribution of the cyclic prefix correlation values of the measurement data in the dynamic road test path with the time-domain and frequency-domain joint distribution template; generating a network topology correlation degree parameter for the preamble detection failure event according to the attenuation slope of the matching degree distribution, and using the network topology correlation degree parameter as the PRACH preamble detection failure sequence feature.

[0014] Optionally, the updating of the incremental diagnosis model using a preset incremental learning algorithm and the composite training sample set includes: monitoring the feature weight correlation coefficient matrix of adjacent diagnosis cycles; when the eigenvalue divergence of the correlation coefficient matrix exceeds a preset range, processing the measurement data of the current cycle, identifying the fuzzy interval measurement data corresponding to a preset confidence fuzzy interval, and preferentially updating the incremental diagnosis model using the fuzzy interval measurement data.

[0015] Optionally, the system further includes a feedback and adaptive control module, including: a feedback processing unit, configured to generate model correction parameters according to the execution feedback of the network optimization decision instruction set; an adjustment instruction unit, configured to generate adjustment instructions based on the model correction parameters, and dynamically adjust the sampling density gradient parameters of the three-dimensional rasterized space matrix by using the adjustment instructions; a reconstruction trigger unit, configured to reconstruct the dynamic road test path based on the corrected three-dimensional rasterized space matrix.

[0016] Compared with the prior art, the present invention has the following advantages:

[0017] 1. Improve the road test efficiency and pertinence: By constructing a QoE space matrix and dynamically generating a road test path according to its attenuation gradient, the test resources can be accurately focused on the areas with network quality degradation or potential problems, avoiding the redundancy and inefficiency caused by traditional fixed paths or blind full-coverage tests, and significantly improving the road test efficiency and resource utilization rate.

[0018] 2. Optimize the data value and analysis load: Use the two-dimensional confidence evaluation to measure the data value, and combine the active learning idea to select the most informative key samples for analysis and modeling, which can effectively compress the data scale to be processed, greatly reduce the storage, transmission and analysis complexity of massive road test data, and improve the data utilization efficiency.

[0019] 3. Enhance the diagnostic accuracy and adaptive ability: Build and update the diagnostic model incrementally based on the actively selected high-value samples, and introduce execution feedback for adaptive correction and adjustment of model parameters, which can diagnose dynamic network problems more accurately and quickly, improve the diagnostic accuracy and the intelligent adaptive level of the system, and shorten the closed-loop cycle of "test - analysis - optimization".

[0020] Other features and advantages of the present invention will be described in the following specification, and part of them will become obvious from the specification, or be understood by implementing the present invention. The objectives and other advantages of the present invention can be achieved and obtained through the structures pointed out in the specification, claims and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0022] Figure 1 It is a framework diagram of a 5G wireless communication network road test system based on active learning according to an embodiment of the present invention.

[0023] Figure 2 It is a schematic structural diagram of the 5G wireless communication network road test system based on active learning according to an embodiment of the present invention.

[0024] Figure 3 It is a schematic diagram of the spatial distribution of QoE indicators according to an embodiment of the present invention.

[0025] Figure 4 It is a comparative curve graph of the active learning effect according to an embodiment of the present invention.

[0026] Figure 5 It is a schematic diagram of the dynamic road test path and nodes according to an embodiment of the present invention. Detailed implementation manners

[0027] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0028] Refer to Figure 1 According to an embodiment of the present invention, a 5G wireless communication network road test system based on active learning is proposed. By adopting an active learning framework and integrating QoE spatial gradient analysis, two-dimensional confidence evaluation, and incremental diagnosis modeling technologies, it can realize dynamic intelligent planning of the 5G network road test path, efficient screening of test data, and accurate diagnosis of network problems, significantly improving the efficiency of network road testing, the accuracy of problem diagnosis, and the response speed of network optimization.

[0029] As Figure 2 shown, the system of this embodiment specifically includes:

[0030] A data acquisition and mapping module, configured to acquire a set of QoE indicators of the terminal service quality experience in a preset area, and map the set of QoE indicators into a three-dimensional rasterized space matrix;

[0031] Specifically, the data acquisition and mapping module is responsible for collecting QoE indicator data reflecting the actual feelings of users in a preset area to form an initial set of indicators. To facilitate spatial analysis and processing, this module further converts and stores these possibly geographically discretely distributed indicator data into a structured three-dimensional rasterized space matrix through mapping or grid technology. This matrix can generally characterize the basic distribution of QoE within a specified spatial range, such as Figure 3As shown, this figure intuitively demonstrates the spatial distribution of QoE. The areas with darker colors in the figure are potential problem areas with lower QoE, providing an intuitive basis for subsequent dynamic path planning.

[0032] A path generation module for generating a dynamic road test path according to the QoE attenuation gradient change of the three-dimensional rasterized space matrix;

[0033] Specifically, the path generation module receives and analyzes the three-dimensional space matrix generated by the data acquisition and mapping module. It calculates the change rate of the QoE index in different spatial directions in the matrix, that is, the attenuation gradient, to identify the areas where the QoE index value drops rapidly or is lower than the preset quality level. These areas are judged as key locations that may have network coverage problems or performance bottlenecks. Based on this judgment, this module plans and generates a dynamic and non-fixed road test path, aiming to guide subsequent road test equipment to preferentially or focus on accessing these identified key areas to improve the pertinence and efficiency of the test process.

[0034] A data evaluation module for performing a two-dimensional confidence evaluation on the measurement data in the dynamic road test path to obtain a spatial-temporal joint confidence distribution map;

[0035] Specifically, the data evaluation module receives the measurement data actually collected during the road test. Since data collection may be affected by various factors such as the environment and equipment status, resulting in differences in data quality, this module needs to evaluate the credibility of these data. It comprehensively considers from two levels: the spatial dimension and the time dimension, and quantitatively evaluates the confidence of the measurement data. The results of the evaluation are finally organized into a spatial-temporal joint confidence distribution map, which can characterize the reliability or uncertainty of the measurement data obtained at different geographical locations and time points, providing a key basis for screening valuable samples.

[0036] A sample screening module for screening an active annotation sample set according to the fuzzy interval distribution of the joint confidence distribution map;

[0037] Specifically, the sample screening module receives the spatial-temporal joint confidence distribution map from the data evaluation module. It aims to apply the strategy of Active Learning (AL) to optimize sample selection. It analyzes this distribution map, focusing on identifying the areas with lower confidence or greater uncertainty, that is, the fuzzy intervals. Selecting the data samples in these fuzzy intervals for priority processing can usually more effectively improve the model performance or discover hidden problems than random sampling, as Figure 4 As shown, this performance comparison chart shows that the active learning of the present invention can achieve a higher diagnostic accuracy rate with fewer training samples compared to random sampling, thus improving the model learning efficiency.

[0038] The model construction and decision-making module is used to construct an incremental diagnosis model based on the active annotation sample set, and generate and output a network optimization decision instruction set according to the incremental diagnosis model.

[0039] Specifically, the model construction and decision-making module receives the active annotation sample set with high information value screened by the sample screening module. The core task of this module is to use these key samples to train or update a diagnosis model that can diagnose network performance problems. This model has the ability of incremental learning, which means it can continuously optimize its own performance and adapt to the changes of the network environment by using new samples, without the need for complete retraining every time. After using the model to complete the diagnosis of the current network state or problems, this module will combine the preset network optimization objectives or policy rules, and convert the diagnosis results into specific and executable network optimization decision instruction sets, such as generating instructions to adjust the base station transmission power, modify the handover threshold or suggesting to enhance the coverage of specific areas, for network operation and maintenance personnel to refer to or directly input into the automated execution system.

[0040] Through the active learning framework, integrating QoE space gradient analysis, two-dimensional confidence evaluation and incremental diagnosis modeling technology, it is possible to realize the dynamic intelligent planning of 5G network road test paths, the efficient screening of test data and the accurate diagnosis of network problems, significantly improving the efficiency of network road tests, the accuracy of problem diagnosis and the response speed of network optimization.

[0041] Optionally, the data acquisition and mapping module includes:

[0042] The data extraction unit is used to extract end-to-end delay probe data from a preset base station interface;

[0043] Specifically, as an important way for the data acquisition and mapping module to obtain raw data, the core task of the data extraction unit is to obtain probe measurement data that can reflect the end-to-end transmission delay performance of services from specific interfaces of the mobile communication network. The data extraction unit can actively pull or passively receive end-to-end delay probe data from these interfaces through configured protocols and authentication methods, using mechanisms such as timed polling, event subscription notifications or reading log files in specific formats. Such probe data is usually generated by active detection programs deployed in the network or on user terminals. The obtained raw delay probe data provides a basic measurement basis for the subsequent data analysis unit to analyze specific service events.

[0044] The data analysis unit is used to analyze the duration distribution of video buffering events in the delay probe data;

[0045] Specifically, the data parsing unit receives the original latency probe data from the data extraction unit. The main task of this unit is to deeply interpret these latency-related data to identify and quantify the key events that affect the user's video viewing experience, namely video buffering events. After identifying the occurrence time and end time of each video buffering event, the data parsing unit calculates the duration of this buffering event. By statistically summarizing the duration data of all buffering events that occur within a certain time window or within a specific geographical area.

[0046] The metric generation unit is used to generate a QoE metric set based on the duration distribution of the video buffering events and in association with the handover request message density in the preset user movement trajectory;

[0047] Specifically, the metric generation unit integrates the video buffering distribution characteristics from the data parsing unit and introduces context information from other dimensions for comprehensive evaluation to generate a metric set that can more comprehensively reflect QoE. The context information that this unit needs to obtain mainly includes the preset user movement trajectory and the handover request message density. User movement trajectory data can usually be obtained by processing the Global Positioning System (GPS) logs recorded by the drive test equipment itself, which reflects the movement path of the test terminal in the geographical space. The handover request message density usually needs to be obtained from the network side, for example, by analyzing the signaling logs of the base station or the core network to extract the frequency or density of handover request events that occur within a specific geographical area or time period. The key of the metric generation unit lies in performing correlation analysis, that is, spatio-temporal matching and correlation calculation of the video buffering situation experienced by the user at specific locations or time periods on the movement trajectory and the handover request density encountered at this location or adjacent areas. This correlation aims to identify those scenarios that may lead to the deterioration of the QoE of video services due to network mobility management problems. For example, the conditional probability of long buffering events occurring in areas where the handover density is higher than the preset threshold can be calculated. Based on the correlation analysis results and the statistical characteristics of the buffering event distribution itself, such as the average buffering duration and the probability of long buffering events, combined with the handover request message density, this unit finally generates a quantified QoE metric set. This set can be represented as a comprehensive QoE score for each geographical grid unit or drive test trajectory segment , and this score can be calculated through the following exemplary weighted model:

[0048] ,

[0049] In the formula: represents the preset full score value of QoE, , , are the weight coefficients preset for the influence of the average buffering duration, the influence of long buffering events, and the influence of handover density respectively, represents the average buffering duration calculated according to the distribution of video buffering event durations, represents the occurrence probability of long buffering events where the buffering duration calculated according to this distribution exceeds a preset threshold, represents the density of handover request messages associated with the current location, represents a normalization function. This model intuitively reflects that the longer the average buffering, the higher the long buffering probability, and the greater the handover density, the greater the deduction item, and the lower the final comprehensive QoE score.

[0050] A spatial matrix mapping unit for mapping the QoE metric set into a three-dimensional rasterized spatial matrix.

[0051] Specifically, the spatial matrix mapping unit receives the QoE metric set output by the metric generation unit, which contains the quantization scores corresponding to each geographical location or road test segment. The core task of this unit is to accurately organize and fill these QoE score data, which may still carry discrete geographical coordinates or regional identifiers, into the corresponding grid cells of a preset three-dimensional rasterized spatial matrix covering the target test area. This mapping process ensures that QoE metric values from different sources and forms are unified onto a standardized spatial grid, forming structured QoE spatial distribution data, providing regular input for all subsequent modules that rely on this spatial matrix for analysis.

[0052] Optionally, the path generation module includes:

[0053] A coefficient of variation calculation unit for calculating the coefficient of variation of the signal strength per unit area in the QoE attenuation gradient change;

[0054] Specifically, the coefficient of variation calculation unit receives and processes the QoE attenuation gradient change information analyzed by the path generation module, which is derived from the three-dimensional rasterized spatial matrix generated by the data acquisition and mapping module. The QoE attenuation gradient reflects the speed and direction of QoE change in space. The task of this unit is to further quantify the instability or severity of this gradient change. It uses the basic grid cells or a preset neighborhood range in the three-dimensional spatial matrix as the unit area and calculates the coefficient of variation (Coefficient of Variation, CV) of the QoE attenuation gradient values within this unit area. The coefficient of variation is a standardized statistic, usually defined as the ratio of the standard deviation to the mean, which can measure the dispersion of gradient data relative to its average level and is not affected by the size of the mean itself. By calculating the gradient coefficient of variation for each unit area, this unit can identify areas where the QoE change is not only fast but also extremely unstable in its change pattern. The calculated result of the signal strength coefficient of variation will be used as a key input basis for the subsequent path node generation unit to determine whether to set optimized path nodes.

[0055] A path node generation unit, configured to generate multi-level optimized path nodes when a difference in the coefficient of variation of the signal strength between adjacent grids exceeds a preset threshold;

[0056] Specifically, the path node generation unit receives the coefficient of variation values of the signal strength corresponding to each unit area in the three-dimensional rasterized space matrix output by the coefficient of variation calculation unit. The core function of this unit is to identify geographical boundaries or regions where the instability of the network QoE changes significantly based on the spatial distribution differences of these coefficients of variation, and generate key nodes for path planning at these locations. It traverses the grid cells i in the spatial matrix and compares the coefficient of variation values with those of each adjacent grid cell j. To quantify this difference and determine whether to generate a node, an index representing the boundary strength between adjacent grids can be calculated :

[0057] ,

[0058] where: represents the boundary strength index between adjacent grids , , are respectively the coefficient of variation values of the signal strength of grid and grid , is a preset threshold for determining whether the difference in the coefficient of variation is significant. When the calculated boundary strength is greater than zero, it indicates that this is a boundary where the QoE gradient volatility changes significantly and has a high detection value. The path node generation unit will generate an optimized path node here. Moreover, to achieve multi-level optimization, the generated path nodes can be assigned different priorities or attributes, and the level of this priority or attribute can be determined according to the specific numerical value of the calculated boundary strength according to a preset classification rule. Usually, the larger the boundary strength value, the higher the optimization priority of the corresponding node. Finally, this unit outputs a set of path nodes containing this priority or attribute information for use by the subsequent path construction unit.

[0059] A path construction unit, configured to construct a topological structure with spatial weights based on the multi-level optimized path nodes to form a dynamic road test path.

[0060] Specifically, the path construction unit receives a set of multi-level optimized path nodes with different priorities or attribute information output by the path node generation unit. These nodes represent key locations on the map that need to be paid attention to. The primary task of this unit is to organize these discrete nodes and construct a topological structure that can reflect the spatial relationship and connection possibility between them, which is usually expressed as a weighted graph. In this graph, the path nodes are the vertices of the graph, and the potential feasible path segments connecting two nodes constitute the edges of the graph. The key is to assign spatial weights to these edges. This weight can be a comprehensive value, which can not only include travel cost factors such as the physical distance between nodes and the estimated travel time, but also incorporate information related to the test value, such as the average QoE gradient value and coefficient of variation value of the area traversed by the path segment, or directly use the priority information of the node. After constructing this topological structure containing spatial weights, the path construction unit will use one or more path planning and optimization algorithms, such as heuristic search algorithms such as A* search (A-Star Search), or customized algorithms for specific goals, to find an optimal or suboptimal path sequence on this weighted topological graph. Finally, the dynamic drive test path output by the unit is a specific, ordered sequence of instructions, such as a series of geographic coordinate points or road segment identifiers, which can be directly used by the drive test equipment to navigate and perform test tasks, such as Figure 5 As shown in the figure, this diagram illustrates how the dynamically generated drive test path connects key test nodes in series. The nodes of different shapes in the figure represent different priorities. For example, high-priority nodes represent the areas that need the most attention, providing clear guidance for the intelligent and efficient allocation of test resources.

[0061] Optionally, generating a multi-level optimization path node includes:

[0062] Analyze building penetration loss patterns using a preset coverage hole prediction algorithm;

[0063] Specifically, in order to accurately identify and quantify the obstruction effect of buildings on 5G signal propagation and predict potential coverage holes, a preset coverage hole prediction algorithm is used to carefully analyze the building penetration loss pattern in the target area. The prediction algorithm can be based on a ray tracing model based on the principles of computational electromagnetics, which requires the input of detailed three-dimensional building geometry model information and electromagnetic parameters of building materials. Typical three-dimensional model information can come from building information models or high-precision geographic information system data, while material electromagnetic parameters include dielectric constants and conductivity of walls or glass of different materials. More computationally efficient engineering empirical models can also be used, such as the logarithmic distance path loss model combined with additional loss factors for specific scenarios, in which building penetration loss is a key consideration. Penetration loss describes the energy attenuation of the signal when it penetrates an obstacle. For example, a simplified multi-wall penetration loss model It can be expressed as:

[0064] ,

[0065] where: represents the total predicted penetration loss of the signal through the building, represents the base loss value through the first exterior wall, represents the loss value averaged through each interior wall, is the number of equivalent interior walls or floors traversed on the signal path, represents other additional losses. By applying such an algorithm, a predicted penetration loss value can be calculated for each grid in the three-dimensional space matrix affected by the building, forming a detailed distribution pattern of building penetration loss.

[0066] Mark potential signal shielding areas in the three-dimensional rasterized space matrix according to the loss pattern;

[0067] Specifically, based on the predicted building penetration loss pattern, those areas in the three-dimensional rasterized space matrix where the signal may be too weak to meet the basic service requirements, i.e., potential signal shielding areas, are clearly marked. The marking is based on comparing the predicted signal reception strength with the minimum signal strength threshold required to maintain a specific service QoE. The predicted signal reception strength can be estimated by subtracting the total loss including path loss and penetration loss from the transmission power at the transmitter. If the predicted reception strength of a certain grid is lower than the preset service availability signal threshold, then that grid is determined to be within the potential signal shielding area. The marking process can be achieved through an indicator function as follows:

[0068] ,

[0069] where: is the shielding mark for grid , is the signal transmission power, is the total path loss to reach grid , is the preset minimum available signal strength threshold. By performing this judgment and marking for all relevant grids, the spatial range and shape of the potential signal shielding area are clearly defined on the three-dimensional space matrix.

[0070] Take the boundary points of the signal shielding area as the mandatory constraint conditions for generating priority path nodes.

[0071] Specifically, after determining the spatial range of potential signal shielding areas, it is necessary to identify the boundary points of these areas and convert them into mandatory constraint conditions when generating path nodes. The identification of boundary points can apply edge detection algorithms in image processing to the two-dimensional or three-dimensional space of the marked result matrix. Common algorithms include, for example, the Canny operator or the Sobel operator, which are used to find the transition positions where the marked values change abruptly. The set of these identified boundary points has important test significance as they represent the critical transition zone of the network coverage quality. This boundary point information is then used as a mandatory constraint condition to affect the process of the path node generation unit generating multi-level optimized path nodes. This prediction-based physical shielding constraint makes the generation of path nodes more physically meaningful and practically testable.

[0072] Exemplarily, assume that a road test path is being planned for a city block with many high-rise buildings. First, use a coverage hole prediction algorithm based on ray tracing, combined with the precise three-dimensional building model of the block, to analyze and calculate the penetration loss pattern of the signal passing through a specific high-rise building C, and predict that there is a large area on its shaded side where the signal strength is lower than the available threshold. Then, according to this prediction result, mark the corresponding grids behind building C in the three-dimensional rasterized space matrix as potential signal shielding areas. Subsequently, identify the set of boundary points where this shielding area intersects with the surrounding non-shielding areas. Finally, when generating multi-level optimized path nodes, apply this boundary point information as a mandatory constraint condition. For example, it is required that high-priority path nodes must be generated within a specific range outside these boundary points, and at the same time, avoid generating any mandatory access nodes inside the marked shielding area. In this way, the set of nodes finally output by the path node generation unit can ensure that the subsequent generated dynamic road test path will focus on detecting the coverage edge of the shaded area of building C, rather than inefficiently penetrating deep into the coverage hole or lingering in areas with good signal.

[0073] Optionally, the data evaluation module includes:

[0074] A spatial dimension analysis unit, which is used to analyze the measurement data in the dynamic road test path in the spatial dimension and statistically calculate the signal quality fluctuation parameters reflecting the state of the micro base station overlapping coverage area;

[0075] Specifically, the spatial dimension analysis unit focuses on evaluating the reliability of measurement data from the perspective of geographical spatial distribution. It receives measurement data collected along the dynamic road test path, which usually contains rich wireless environment information, such as Received Signal Strength Indicator (RSSI), Reference Signal Received Power (RSRP), Reference Signal Received Quality (RSRQ), and the identifiers of the current serving cell and neighboring cells, etc. A key task of this unit is to identify the overlapping coverage areas of micro base stations, which can be achieved by analyzing the micro base station cell identifiers with multiple strong signals simultaneously existing in the measurement data, or by combining the geographical location and coverage range information of micro base stations recorded in the preset base station engineering parameter database for judgment. Within these identified overlapping coverage areas, due to possible problems such as pilot contamination, increased interference, or frequent handovers, the signal quality is often unstable. Therefore, this unit will further count the signal quality fluctuation parameters used to quantify this instability. For example, the standard deviation or coefficient of variation of the RSRP value of the primary serving cell within a short distance can be calculated, or the number of times or frequency of handovers of the serving cell can be counted. The calculated values of these fluctuation parameters will be used as one of the bases for evaluating the confidence level of data in the spatial dimension.

[0076] The time dimension analysis unit is used to analyze the measurement data in the dynamic road test path in the time dimension and extract the burst frequency change rate of the Tracking Area Update (TAU) event;

[0077] Specifically, the time dimension analysis unit focuses on the dynamic characteristics of measurement data changing over time to evaluate the timeliness and stability of the data. It also analyzes the received measurement data, especially when these data contain the third-layer signaling message records of the terminal. This unit will identify and extract specific mobility management events from them, such as the Tracking Area Update (TAU) event. TAU is a signaling process that the terminal must initiate to maintain its reachability in the network when it crosses the boundary of the tracking area (TA) pre-divided by the network during movement. This unit calculates the occurrence frequency of the TAU event by monitoring the number of TAU events occurring within consecutive time windows, and further analyzes the change rate of this frequency over time, such as calculating the first-order difference or short-term fluctuation amplitude of the frequency time series, so as to obtain the burst frequency change rate. A high burst frequency change rate may imply a sharp change in the user's movement state, network boundary configuration problems, or abnormalities in paging and location update signaling processes, all of which may affect the representativeness and stability of the measurement data near this time point and are important references for evaluating the confidence level of data in the time dimension.

[0078] The confidence fusion unit is used to perform weighted fusion on the signal quality fluctuation parameter and the burst frequency change rate to obtain a joint confidence distribution map.

[0079] Specifically, the confidence fusion unit brings together the evaluation results of data quality or stability from two dimensions, space and time. It receives the quantitative signal quality fluctuation parameters output by the spatial dimension analysis unit and the quantitative TAU event burst frequency change rate parameters output by the temporal dimension analysis unit. The goal of this unit is to comprehensively calculate a final joint confidence score based on the information of these two or more dimensions to represent the overall reliability of the measurement data obtained at a specific time and space point. The score can be designed so that the greater the fluctuation and the higher the rate of change, the lower the confidence. This can be achieved through a preset weighted fusion model, as shown below:

[0080] ,

[0081] Where: represents the calculated joint confidence score, is the preset upper limit of the confidence score, , are the preset weight coefficients assigned to the spatial dimension parameters and the temporal dimension parameters, respectively. is the input quantized signal quality fluctuation parameter value, is the input quantized TAU event burst frequency change rate value, Represents a normalized function. The more drastic the spatial fluctuation or temporal change, the larger the normalized value, resulting in more deductions from the full score and a lower final joint confidence score. By performing this fusion calculation on the data of all sampling points on the drive test path, a spatial-temporal joint confidence distribution map covering the entire test area and time range can be generated.

[0082] Exemplarily, when the road test vehicle travels to an area B where micro base stations are densely deployed and happens to be the junction of multiple tracking areas, the spatial dimension analysis unit analyzes the measurement data and finds that the service cell changes frequently and the standard deviation of the received signal power is large, and calculates a higher signal quality fluctuation parameter. At the same time, the time dimension analysis unit analyzes the signaling data of this period and detects that the terminal has initiated multiple tracking area update events, and the short-term change rate of the frequency of occurrence is also very high. After receiving these two parameters that both indicate high data instability, the confidence fusion unit calculates according to the preset weights and fusion formulas, and obtains a significantly low joint confidence score. This low score value is then filled into the space-time joint confidence distribution map corresponding to area B and the test time, clearly indicating that the data confidence of this time-space point is low and may need to be given priority by the subsequent sample screening module.

[0083] Optionally, the signal quality fluctuation parameter that statistically reflects the overlapping coverage area status of the micro base station includes:

[0084] Obtain a reference signal received quality (RSRQ) measurement report set within a preset time period;

[0085] Specifically, in order to quantitatively evaluate the signal quality stability of the micro base station overlapping coverage area, it is first necessary to obtain basic data that can reflect the signal quality status of the area. The spatial dimension analysis unit will collect and merge the Reference Signal Received Quality (RSRQ) measurement reports reported by all terminals whose geographical locations fall into the identified micro base station overlapping coverage area within a preset time period from specified sources, such as detailed measurement logs recorded by road test terminal equipment, user measurement reports collected by the network management system, or data obtained through the mobility driven test mechanism. These reports constitute the RSRQ measurement report set for subsequent analysis. As a key indicator, RSRQ can comprehensively reflect the strength of the received signal and the level of co-channel interference and noise, and is a common basis for evaluating the signal quality of the cell edge and overlapping coverage areas.

[0086] Establishing an RSRQ coefficient of variation matrix for a multi-cell joint coverage area based on the RSRQ measurement report set;

[0087] Specifically, after collecting enough RSRQ measurement report samples, it is necessary to quantify the degree of fluctuation of signal quality based on these measurement report sets. For the multi-cell joint coverage area within the analysis range, calculate the coefficient of variation of all RSRQ sample values ​​in the area. The calculation process is to first calculate the arithmetic mean and standard deviation of all RSRQ samples in the area, and then obtain the coefficient of variation of RSRQ. By summarizing the RSRQ coefficient of variation values ​​calculated for each analysis area, an RSRQ coefficient of variation matrix or mapping diagram can be established. The values ​​in this matrix reflect the relative volatility of signal quality at different overlapping coverage locations. Higher values ​​indicate more severe fluctuations.

[0088] An interference-sensitive area set satisfying a preset cooperative coverage condition is identified in the RSRQ variation coefficient matrix, and the interference-sensitive area set is used as a signal quality fluctuation parameter.

[0089] Specifically, in the established RSRQ coefficient of variation matrix, specific regions that meet the preset co-coverage condition are identified to form a set of interference-sensitive regions. Here, the co-coverage condition usually refers to the criterion that the signal quality fluctuation exceeds the normal or acceptable range. For example, an RSRQ coefficient of variation threshold can be preset. When the RSRQ coefficient of variation value corresponding to a certain region in the matrix is greater than the RSRQ coefficient of variation threshold, it is determined that this region meets the condition and is classified as an interference-sensitive region. There may be significant co-channel interference, unreasonable handover parameter configuration, or other problems that cause drastic signal quality fluctuations in these regions. Finally, the location information or identifiers of all the identified interference-sensitive regions are collected to form a set. This set of interference-sensitive regions itself, or the features further extracted based on this set, is used as the signal quality fluctuation parameter reflecting the state of the overlapping coverage area output by the spatial dimension analysis unit.

[0090] Exemplarily, assume that the spatial dimension analysis unit is analyzing the known overlapping coverage area C of micro base stations BS1 and BS2. It first obtains the RSRQ measurement reports reported by all terminals that entered or were in area C in the past 5 minutes. Based on these reports, the variation matrix establishment step calculates that the overall RSRQ coefficient of variation within area C is 0.4. Subsequently, the area identification step compares this calculated value of 0.4 with the preset co-coverage condition. The co-coverage condition may be set such that an RSRQ coefficient of variation greater than 0.35 is regarded as interference-sensitive. Since the calculated value is greater than the threshold and meets the condition, area C is identified and marked as an interference-sensitive region. This identification result constitutes a part of the output signal quality fluctuation parameter, indicating that the spatial signal stability in this region is poor, and its information will be used by the subsequent confidence fusion unit.

[0091] Optionally, the model construction and decision module includes:

[0092] A feature extraction unit for extracting PRACH preamble detection failure sequence features from the active annotation sample set;

[0093] Specifically, the feature extraction unit processes the received active annotation sample set, focusing on analyzing the underlying information that reflects the situation of the terminal's initial access to the network. It parses the sample data, identifies the failure events in the physical random access channel PRACH process, and extracts sequence features that can quantify the difficulty of access, such as the frequency or pattern of failures, to provide input for subsequent correlation analysis.

[0094] A training sample generation unit for correlating the PRACH preamble detection failure sequence features with a preset TCP retransmission response rate time series map to generate a composite training sample set;

[0095] Specifically, the training sample generation unit aims to construct training data that reflects the association between underlying access and upper-layer transmission performance. It performs spatio-temporal correlation matching on the PRACH failure features output by the feature extraction unit and a preset time-series map that records the performance of the Transmission Control Protocol (TCP) layer of the network transmission. By combining the matched features at different levels, a composite training sample set containing cross-layer information is generated for training the diagnostic model.

[0096] The model construction unit is used to take network state features as input and network performance metrics as output, and establish and train a neural network model using the composite training sample set to obtain an incremental diagnostic model;

[0097] Specifically, the model construction unit is responsible for performing the first creation process of the incremental diagnostic model from scratch. It receives the composite training sample set provided by the training sample generation unit, uses the network state features therein as model inputs, such as the PRACH preamble detection failure sequence features, and uses the associated network performance metrics as the expected outputs or labels, such as the TCP retransmission response rate. By executing the standard supervised learning training process, an initial neural network diagnostic model is established and trained. This initial model forms the basis for subsequent incremental updates and provides a baseline model that can be optimized for the model update unit.

[0098] The model update unit is used to update the incremental diagnostic model by using a preset network optimization objective function as a constraint condition and a preset incremental learning algorithm and the composite training sample set;

[0099] Specifically, the model update unit uses the composite training sample set provided by the training sample generation unit to adjust and optimize the incremental diagnostic model by executing a preset incremental learning algorithm. This learning method enables the model to continuously adapt to network changes based on new samples. During the update process, the adjustment direction of the model is also restricted by the constraints imposed by the preset network optimization objective function to ensure that its evolution conforms to the overall network optimization strategy. This unit finally outputs an updated incremental diagnostic model with improved performance.

[0100] The decision generation and output unit is used to analyze the network state based on the incremental diagnostic model and generate a network optimization decision instruction set for output.

[0101] Specifically, the decision-making generation and output unit utilizes the latest incremental diagnostic model provided by the model update unit. Based on this model, this unit conducts diagnostic analysis on the current network state to identify potential network performance issues or points for optimization. After completing the diagnosis, it combines the preset network optimization strategies or expert rules to convert the diagnostic conclusions of the model into specific and executable network optimization decision instruction sets. These instruction sets are then formatted and output to guide subsequent network adjustment and optimization work.

[0102] Exemplarily, low-confidence samples from region C are screened out. The feature extraction unit identifies the feature of a high frequency of PRACH access failures from them. The training sample generation unit discovers that the preset TCP map corresponding to this region during the corresponding period also shows a high retransmission rate, so it associates the access failure feature with the high retransmission rate to generate composite training samples. The model update unit uses these samples and updates the diagnostic model under the constraint of the network stability objective through an incremental learning algorithm. The updated model can better identify the TCP performance degradation phenomenon caused by such access problems and guide subsequent decision-making.

[0103] Optionally, extracting the PRACH preamble detection failure sequence feature from the active annotation sample set includes:

[0104] Constructing a time-domain and frequency-domain joint distribution template based on a preset preamble sequence;

[0105] Specifically, in order to deeply analyze the cause of access failure of the physical random access channel from the signal processing level, it is necessary to construct a reference template based on a preset preamble sequence. This preset preamble sequence usually refers to a specific waveform sequence defined in 5G or other cellular network standards for random access. The constructed time-domain and frequency-domain joint distribution template describes the energy or phase distribution characteristics of this ideal preamble sequence in two dimensions of time and frequency, serving as a reference for subsequent comparison.

[0106] Comparing the matching degree distribution of the cyclic prefix correlation values of the measurement data in the dynamic road test path with the time-domain and frequency-domain joint distribution template;

[0107] Specifically, the measurement data in the dynamic road test path corresponding to the active annotation sample set is used to compare the actually collected signal with this ideal template, and the cyclic prefix (CP) correlation values in these actual measurement data are calculated. The cyclic prefix is a technique used in OFDM systems to combat multipath effects. Calculating its correlation value helps to evaluate the synchronization state and channel quality of the signal. By comparing the similarity between the cyclic prefix correlation values of the measured data and the ideal template, a matching degree distribution can be obtained, which reflects the extent to which the received signal conforms to the expected preamble sequence characteristics.

[0108] Generate a network topology correlation parameter for the preamble detection failure event according to the attenuation slope of the matching degree distribution, and use the network topology correlation parameter as the PRACH preamble detection failure sequence feature.

[0109] Specifically, generate a quantified feature parameter according to the matching degree distribution. Pay attention to the attenuation slope of the matching degree changing with certain factors. If the matching degree decays rapidly, it usually indicates poor received signal quality, difficult synchronization, or strong interference, which is likely to be the reason for the preamble detection failure. By calculating and quantifying this attenuation slope, a network topology correlation parameter for the preamble detection failure event can be generated. This parameter aims to characterize the association strength between this access failure and a specific network topology location or channel state. Finally, the generated network topology correlation parameter is used as the PRACH preamble detection failure sequence feature output by the feature extraction unit for subsequent analysis and model training.

[0110] Optionally, the updating of the incremental diagnosis model by using the preset incremental learning algorithm and the composite training sample set includes:

[0111] Monitor the feature weight correlation coefficient matrix of adjacent diagnosis cycles;

[0112] Specifically, in order to ensure the stability and adaptability of the incremental diagnosis model during the continuous learning process, the changes in the internal state of the model will be monitored. Calculate and compare the feature weight correlation coefficient matrices of the models in adjacent diagnosis cycles. This matrix can reflect the degree of mutual correlation that the model currently believes exists between the input features. By comparing this matrix between consecutive update cycles, it can be determined whether there have been drastic changes in the internal structure of the model or its understanding of the relationships between features.

[0113] When the eigenvalue divergence of the correlation coefficient matrix exceeds a preset range, process the measurement data of the current cycle, identify the fuzzy interval measurement data corresponding to the preset confidence fuzzy interval, and use the fuzzy interval measurement data to preferentially update the incremental diagnosis model.

[0114] Specifically, calculate the degree of difference between adjacent cycle feature weight correlation matrices, for example, by calculating the eigenvalue divergence or other distance metrics between them. Compare this quantified difference value with a preset range or threshold. If the calculated divergence exceeds this preset range, it indicates that the model may have experienced significant concept drift or learned an unstable state. According to the preset confidence fuzzy interval standard, identify the data points in the current measurement data that fall into these high-uncertainty intervals, that is, fuzzy interval measurement data. These data are considered to have high learning value because they contain information that the model is not yet certain about. Insert the identified fuzzy interval measurement data into a priority queue associated with a preset historical training set. This means that these data will be preferentially selected in subsequent model updates to help the model learn faster and improve its prediction or diagnostic ability in fuzzy and uncertain regions.

[0115] Optionally, the system further includes a feedback and adaptive control module, including:

[0116] A feedback processing unit for generating model correction parameters based on the execution feedback of the network optimization decision instruction set;

[0117] Specifically, the feedback processing unit collects the actual execution feedback information of the network optimization instructions. By analyzing the deviation between this feedback information and the expected effect, this unit generates model correction parameters for guiding the improvement of the diagnostic model.

[0118] An adjustment instruction unit for generating adjustment instructions based on the model correction parameters and dynamically adjusting the sampling density gradient parameters of the three-dimensional rasterized space matrix;

[0119] Specifically, the adjustment instruction unit receives the model correction parameters and determines whether the construction parameters of the three-dimensional space matrix need to be adjusted accordingly. If so, it generates corresponding adjustment instructions to dynamically update settings such as the sampling density gradient parameters of the space matrix.

[0120] A reconstruction trigger unit for reconstructing the dynamic road test path based on the corrected three-dimensional rasterized space matrix.

[0121] Specifically, when the three-dimensional rasterized space matrix is updated to the corrected state according to the adjustment instruction, the reconstruction trigger unit monitors this update and triggers the path generation module. This trigger causes the path generation module to re-plan based on the corrected space matrix and generate a new dynamic road test path.

[0122] Exemplarily, after the decision instruction is executed, the feedback processing unit analyzes the effect feedback and generates model correction parameters indicating that the monitoring of a certain area needs to be strengthened. The adjustment instruction unit generates an instruction accordingly, increasing the sampling density parameter of this area in the three-dimensional space matrix. After the matrix is updated, the reconstruction trigger unit immediately triggers the path generation module, and a dynamic road test path is regenerated based on the corrected matrix. The new path naturally increases the coverage of this key area.

[0123] It should be noted that the electrical connections between the above-mentioned various units do not necessarily mean direct connection of the circuits. Indirect connection methods can be applied to the embodiments of the present invention as long as the purpose of the present invention is achieved. The above are only exemplary embodiments of the present invention, and the scope of the present invention cannot be limited thereby.

[0124] That is, any equivalent changes and modifications made in accordance with the teachings of the present invention still fall within the scope covered by the present invention. After considering the specification and the disclosure of the practical truth, those skilled in the art will easily think of other implementation schemes of the present invention. This application aims to cover any variations, uses, or adaptive changes of the present invention, which follow the general principles of the present invention and include common general knowledge or conventional technical means in the technical field not recorded in the present invention.

Claims

1. A road test system for a 5G wireless communication network based on active learning, characterized in that, The system includes: A data acquisition and mapping module, configured to acquire a set of terminal Quality of Experience (QoE) metrics for a preset area, and map the QoE metric set into a three-dimensional rasterized space matrix; A path generation module, configured to generate a dynamic road test path according to the QoE attenuation gradient change of the three-dimensional rasterized space matrix; A data evaluation module, configured to perform a two-dimensional confidence evaluation on the measurement data in the dynamic road test path to obtain a space-time joint confidence distribution map; A sample screening module, configured to screen an active annotation sample set according to the fuzzy interval distribution of the joint confidence distribution map; A model construction and decision module, configured to construct an incremental diagnosis model based on the active annotation sample set, and generate and output a network optimization decision instruction set according to the incremental diagnosis model.

2. The 5G wireless communication network road test system based on active learning according to claim 1, characterized in that, The data acquisition and mapping module includes: A data extraction unit, configured to extract end-to-end delay probe data from a preset base station interface; A data parsing unit, configured to parse the duration distribution of video buffering events in the delay probe data; An index generation unit, configured to generate a QoE metric set based on the duration distribution of the video buffering events and in association with the handover request message density in a preset user movement trajectory; A space matrix mapping unit, configured to map the QoE metric set into a three-dimensional rasterized space matrix.

3. The road test system for 5G wireless communication network based on active learning according to claim 1, wherein The path generation module includes: A coefficient of variation calculation unit, configured to calculate the coefficient of variation of the signal strength per unit area in the QoE attenuation gradient change; A path node generation unit, configured to generate multi-level optimized path nodes when the difference in the coefficient of variation of the signal strength between adjacent grids exceeds a preset threshold; A path construction unit, configured to construct a topological structure including spatial weights according to the multi-level optimized path nodes to form a dynamic road test path.

4. The road test system for 5G wireless communication network based on active learning according to claim 3, wherein The generation of the multi-level optimized path nodes includes: Analyzing the building penetration loss pattern by using a preset coverage hole prediction algorithm; Marking potential signal shielding areas in the three-dimensional rasterized space matrix according to the loss pattern; Taking the boundary points of the signal shielding areas as the forced constraint conditions for generating priority path nodes.

5. The 5G wireless communication network road test system based on active learning according to claim 1, characterized in that, The data evaluation module includes: A spatial dimension analysis unit, configured to analyze the measurement data in the dynamic road test path in the spatial dimension and statistically calculate signal quality fluctuation parameters reflecting the state of the micro base station overlapping coverage area; A time dimension analysis unit, configured to analyze the measurement data in the dynamic road test path in the time dimension and extract the burst frequency change rate of Tracking Area Update (TAU) events; A confidence fusion unit, configured to perform weighted fusion on the signal quality fluctuation parameters and the burst frequency change rate to obtain a joint confidence distribution map.

6. The road test system for 5G wireless communication network based on active learning according to claim 5, characterized in that, The signal quality fluctuation parameters statistically reflecting the state of the micro base station overlapping coverage area include: Obtaining a set of Reference Signal Received Quality (RSRQ) measurement reports within a preset time period; Establishing an RSRQ coefficient of variation matrix for a multi-cell joint coverage area based on the RSRQ measurement report set; Identifying a set of interference-sensitive areas satisfying preset co-coverage conditions in the RSRQ coefficient of variation matrix, and taking the set of interference-sensitive areas as the signal quality fluctuation parameters.

7. The road test system for a 5G wireless communication network based on active learning according to claim 1, wherein, The model construction and decision-making module includes: A feature extraction unit, configured to extract PRACH preamble detection failure sequence features from the active annotation sample set; A training sample generation unit, configured to associate the PRACH preamble detection failure sequence features with a preset TCP retransmission response rate time series map to generate a composite training sample set; A model construction unit, configured to use the network state features as inputs and the network performance metrics as outputs, and use the composite training sample set to establish and train a neural network model to obtain an incremental diagnosis model; A model update unit, configured to use a preset network optimization objective function as a constraint condition, and use a preset incremental learning algorithm and the composite training sample set to update the incremental diagnosis model; A decision generation and output unit, configured to analyze the network state according to the incremental diagnosis model, and generate a network optimization decision instruction set for output.

8. The road test system for 5G wireless communication network based on active learning according to claim 7, wherein The extracting the PRACH preamble detection failure sequence features from the active annotation sample set includes: Constructing a time-domain frequency-domain joint distribution template based on a preset preamble sequence; Comparing the matching degree distribution of the cyclic prefix correlation value of the measurement data in the dynamic road test path with the time-domain frequency-domain joint distribution template; Generating a network topology correlation parameter of the preamble detection failure event according to the attenuation slope of the matching degree distribution, and using the network topology correlation parameter as the PRACH preamble detection failure sequence feature.

9. The road test system for 5G wireless communication network based on active learning according to claim 7, characterized in that, The updating the incremental diagnosis model by using a preset incremental learning algorithm and the composite training sample set includes: Monitoring the feature weight correlation coefficient matrix of adjacent diagnosis cycles; When the eigenvalue divergence of the correlation coefficient matrix exceeds a preset range, processing the measurement data of the current cycle, identifying the fuzzy interval measurement data corresponding to a preset confidence fuzzy interval, and preferentially updating the incremental diagnosis model by using the fuzzy interval measurement data.

10. The road test system for 5G wireless communication network based on active learning according to claim 1, wherein, The system further includes a feedback and adaptive control module, including: A feedback processing unit, configured to generate a model correction parameter according to the execution feedback of the network optimization decision instruction set; An adjustment instruction unit, configured to generate an adjustment instruction based on the model correction parameter, and use the adjustment instruction to dynamically adjust the sampling density gradient parameter of the three-dimensional rasterized space matrix; A reconstruction trigger unit, configured to reconstruct the dynamic road test path based on the corrected three-dimensional rasterized space matrix.

Images