Network fault processing method and computer device

Abstract

The application relates to the technical field of wireless communication, and discloses a network fault processing method and computer equipment, which comprises the following steps: acquiring actual index data of a network to be detected; using a target detection model to perform fault detection on the network to be detected based on the actual index data, so as to obtain a fault detection result; in the case that the fault detection result indicates that the network to be detected is in a target fault type, using a target twin network to perform parameter optimization processing on the network to be detected, so as to obtain a target parameter adjustment strategy; wherein the target twin network corresponds to the target fault type; performing a fault recovery operation based on the target parameter adjustment strategy, so as to complete network fault processing. The parameter optimization processing using the target twin network can safely and efficiently explore an optimal parameter adjustment scheme in a virtual simulation environment of a specific fault type, so that the automation of detection and the rapid repair of faults are realized, and the network operation and maintenance efficiency and the user communication experience are improved.

Description

Technical Field

[0001] This application relates to the field of wireless communication technology, and in particular to a network fault handling method and a computer device. Background Technology

[0002] Network failures are inevitable during the operation and maintenance of wireless communication networks. However, the heterogeneity and complexity of networks exacerbate the cost and difficulty of fault detection and repair, necessitating a new, highly autonomous network fault handling method. Summary of the Invention

[0003] This application aims to at least partially solve one of the technical problems in related technologies. To this end, this application proposes a network fault handling method and a computer device. The main technical solutions adopted in this application include: In a first aspect, this application provides a network fault handling method, which includes: acquiring actual indicator data of the network to be tested; using a target detection model to perform fault detection on the network to be tested based on the actual indicator data to obtain a fault detection result; if the fault detection result indicates that the network to be tested is in a target fault type, using a target twin network to perform parameter optimization processing on the network to be tested to obtain a target parameter tuning strategy; wherein, the target twin network corresponds to the target fault type; and performing a fault recovery operation based on the target parameter tuning strategy to complete the network fault handling.

[0004] By acquiring actual performance data of the network under test, the communication quality of the network can be monitored in real time, providing a reliable data foundation for fault detection. Using a target detection model based on actual performance data for fault detection can quickly and accurately identify the presence and specific type of fault in the network, providing a basis for subsequent targeted handling. Furthermore, in the event of a detected fault, parameter optimization is performed using a target twin network corresponding to the target fault type. This allows for the safe and efficient exploration of optimal parameter adjustment schemes in a virtual simulation environment, avoiding the risks of trial and error in a real network. Finally, fault recovery operations are performed based on the target parameter tuning strategy, applying the optimized parameter configuration to the actual network to achieve rapid network performance recovery. Thus, a fully collaborative network fault handling solution has been constructed, achieving automated detection and rapid fault repair, improving network operation and maintenance efficiency and user communication experience.

[0005] Secondly, this application also provides a computer device including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the above methods. Attached Figure Description

[0006] To more clearly illustrate the technical solutions in the specific embodiments of this application or the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0007] Figure 1a This is a flowchart of a network fault handling method according to an embodiment of this application; Figure 1b This is a schematic diagram of a network to be detected according to an embodiment of this application; Figure 2 This is a flowchart of a method for determining a target detection model according to an embodiment of this application; Figure 3 This is a flowchart of a method for obtaining a target parameter tuning strategy according to an embodiment of this application; Figure 4 This is an internal structural diagram of a computer device provided according to an embodiment of this application. Detailed Implementation

[0008] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0009] In fault detection, most fault management technologies rely on artificial intelligence, requiring the collection, storage, and analysis of key network performance indicators. However, these methods separate storage and computation, leading to frequent data transmissions. Furthermore, fault compensation methods require interaction with the actual network, increasing communication overhead. Extensive testing and debugging further drive up maintenance costs. Therefore, these technologies suffer from low fault management efficiency and high costs.

[0010] Furthermore, in wireless communication networks, since the network operates in normal mode most of the time, fault data is difficult to obtain, leading to reduced accuracy of fault detection and root cause analysis methods. Meanwhile, supervised machine learning requires sample labels for training, but these labels rely on manual identification by technicians, which is difficult to verify and whose accuracy is hard to guarantee. Therefore, related technologies suffer from low accuracy in fault detection and root cause analysis methods.

[0011] Furthermore, network fault recovery and compensation technologies rely heavily on deep network enhancement to predict optimal parameter adjustments for wireless networks, requiring accurate and complete real-time channel state information. However, for dynamic networks with massive access volumes, such information is difficult to obtain. Therefore, these technologies suffer from an over-reliance on real-time channel state information.

[0012] In summary, the existing technologies have shortcomings in terms of fault management efficiency, detection accuracy, and reliance on real-time information. To address these limitations, digital twin technology, as a key technology for mapping between the virtual and real worlds, can create virtual representations of physical objects, providing a more intelligent network management solution.

[0013] Based on this, according to the embodiments of this application, a network fault handling method embodiment is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0014] This embodiment provides a network fault handling method, such as Figure 1a As shown, the method includes the following steps: S110. Obtain the actual performance data of the network to be tested.

[0015] The network to be detected can refer to a wireless communication network composed of multiple base stations and multiple user terminals, enabling communication between these base stations and terminals. For example, the detection network can adopt a millimeter-wave wireless network architecture. Taking the downlink as an example, the operating frequency can be 80 GHz, the coverage radius of each base station in the network can be 25m, and each base station is equipped with three uniform rectangular arrays (URAs). Each URA spans 120°, uniformly dividing the horizontal plane. The size N of each URA... t The number N of antennas in the horizontal plane can be determined. az The number of antennas N in the elevation plane el Decision, for example, can take N t =N az ×N el =8×16. The horizontal half-power beamwidth of the antenna can be 65°, the vertical half-power beamwidth can be 35°, the horizontal sidelobe level can be 25dB, and the vertical sidelobe level can be 20dB.

[0016] Furthermore, we can assume that the base station is located at a height of 1.5m, and the user terminals are evenly distributed within the hexagonal coverage area of ​​the base station at a height of 8m. For simplified analysis, please refer to... Figure 1bThe network under test may include 7 base stations and 1000 user terminals, each equipped with a single antenna. Next, for each user terminal, a distance relationship table can be established between it and all base stations, with the corresponding base stations arranged from closest to furthest. Further, it is assumed that each user terminal is served by the base station ranked first in its distance relationship table (the closest base station) and is subject to interference from the base stations ranked second and third in the distance relationship table (the second and third closest base stations). Additionally, it can be assumed that the standard deviation of shadow fading during network transmission is 4 dB to quantify its impact.

[0017] Actual performance data can refer to performance metrics that can quantify the quality of network communication under test, i.e., key performance indicators (KPIs). For example, these may include the reference signal received power (RSRP) and signal-to-interference plus noise ratio (SINR) of the user terminal.

[0018] Specifically, the actual performance data of the network under test can be obtained by using a minimized drive test report.

[0019] The Minimization Drive Test (MDT) report is a standardized report that user terminals automatically measure and report during normal communication. It contains various network signal measurement data obtained by user terminals at specific times or locations, so that base stations can extract key performance indicators after collecting these reported data.

[0020] For example, user terminals within the network under test can automatically measure the currently received network signal data according to a preset period, and package the measured RSRP and SINR data, along with their own identifier, measurement time, and serving base station identifier, into an MDT report, which is then automatically reported to the serving base station. The base station, after summarizing the MDT reports from all user terminals, transmits them to the network backend so that the RSRP and SINR values ​​reported by each user terminal can be extracted and organized into key performance indicators for the corresponding batches.

[0021] Furthermore, to facilitate subsequent model processing, the RSRP and SINR of multiple user terminals can be defined as a set of V wireless network KPIs vectors, where v∈{1,2,…,V} KPIs vectors can be defined as: Where v is the current count; V represents the total number of data collections; m refers to the current user terminal; and M represents the total number of user terminals participating in data reporting within the network during a single data collection process. It is the SINR of the m-th user terminal during the v-th collection; It is the RSRP of the m-th user terminal during the v-th collection.

[0022] For example, during the first data collection, the RSRP and SINR of M user terminals within the network are collected to construct the first KPIs vector. During the second data collection, the RSRP and SINR of another M user terminals within the network are collected to construct the second KPIs vector. This process continues until the Vth collection is completed, resulting in a total of V KPIs vectors. These vectors serve as actual indicator data, providing a sufficient data foundation for subsequent fault detection.

[0023] S120. Using the target detection model, fault detection is performed on the network to be detected based on actual index data to obtain fault detection results.

[0024] It is important to understand that during actual operation, the network under test may experience various types of faults due to factors such as equipment aging, environmental changes, and configuration errors. These faults can lead to decreased network performance and affect user communication quality, such as insufficient signal coverage, increased interference, and reduced transmission rates. Optionally, the target fault types in the network under test may include antenna tilt angle too small, antenna tilt angle too large, and power too low.

[0025] Excessive Uptilt (EU) fault refers to a fault type in which the actual downtilt angle of the base station antenna is less than the preset normal reference downtilt angle. This EU fault manifests as the antenna signal main lobe spreading to distant areas, resulting in insufficient signal strength in the near-site coverage area of ​​the base station and causing strong interference to distant base stations.

[0026] Excessive Downtilt (ED) fault refers to a fault type in which the actual downtilt angle of the base station antenna is greater than the preset normal reference downtilt angle. This ED fault manifests as the antenna signal main lobe being excessively concentrated near the base station, resulting in an excessively contracted signal coverage area. Users in peripheral areas cannot obtain effective service, and blind spots appear in cell coverage.

[0027] Excessive Reduced Power (ERP) fault, also known as insufficient transmit power fault, refers to a situation where the actual transmit power of a base station is far lower than the preset normal reference transmit power, which will result in insufficient signal strength received by user terminals covering the entire range of the base station and a general decline in communication quality.

[0028] Limiting the target fault types to the three mentioned above covers the vast majority of network performance anomalies that affect user communication experience. At the same time, it avoids the increased detection complexity and decreased detection accuracy caused by having too many fault types, thus balancing the coverage and execution efficiency of fault detection.

[0029] Furthermore, given the existence of the aforementioned fault types and the different impact characteristics of different fault types on network performance, failure to accurately determine the fault type during network operation and management will directly lead to poor fault recovery and may even exacerbate network performance anomalies. Therefore, this solution pre-trains a target detection model with fault classification capabilities to achieve automated detection and type determination of network faults.

[0030] The target detection model can determine whether the network under test is in a normal state or has experienced a certain type of fault based on the actual input index data, and determine the specific fault type when a fault occurs.

[0031] Specifically, the object detection model can contain multiple independent classifiers, each responsible for determining a network state (i.e., a normal state or a pre-defined type of fault state). When input data is fed into the model, each classifier calculates the confidence score of the input data belonging to its corresponding network state. Subsequently, the model combines pre-defined cost weights to perform a weighted evaluation of each classifier's output, calculating the total cost required to classify it as various state types, and selecting the state type with the lowest total cost as the fault detection result. This fault detection result can refer to the judgment of the fault state of the network under test, indicating whether the network under test has a fault, and if so, what specific fault type it belongs to.

[0032] For example, assuming the network state includes a normal state and three preset types of fault states (antenna tilt angle too small fault, antenna tilt angle too large fault, and power too low fault), the target detection model can be a classification model with four independent classifiers built in, and each classifier can be a binary classifier. The first classifier can identify the normal state as a positive example and the other three preset types of fault states as negative examples, and is responsible for determining whether the network to be detected is in a normal state; the second classifier can identify the antenna tilt angle too small fault as a positive example and the other three states as negative examples, and is responsible for determining whether the network to be detected has an antenna tilt angle too small fault; the third classifier can identify the antenna tilt angle too large fault as a positive example and the other three states as negative examples, and is responsible for determining whether the network to be detected has an antenna tilt angle too large fault; the fourth classifier can identify the power too low fault as a positive example and the other three states as negative examples, and is responsible for determining whether the network to be detected has a power too low fault.

[0033] It's important to note that directly selecting the category with the highest score from the four classifiers as the fault detection result implies that all misclassifications have the same consequence. However, in actual operation and maintenance, the loss from missing a fault (classifying it as normal) far outweighs the loss from falsely reporting a normal situation as a fault. For example, a network may actually experience a low power fault, but the fourth classifier outputs an incorrect result. Suppose that after detection, the first classifier outputs a score of +60; the second classifier outputs a score of +55; the third classifier outputs a score of -80; and the fourth classifier outputs a score of -70. If only the highest score is used as the criterion, due to the incorrect output of the fourth classifier, the model will directly classify it as normal (the category corresponding to the highest score of +60), ultimately leading to serious missed fault detection. This results in a continuous deterioration of network performance, and the maintenance personnel will be unaware of the cause of the problem.

[0034] To avoid the above situation, a cost-sensitive decision-making method can be introduced. Instead of using the score output by the classifier as the judgment criterion, it uses the "total misclassification cost that the classifier must bear when classifying the network to be detected as a certain category" as the judgment criterion, and finally selects the network state with the minimum total cost as the final fault detection result. For example, the preset cost weights Cost of the four classifiers can be set to [cost1, cost2, cost3, cost4], where cost1 is the cost coefficient of misclassifying a faulty network as a normal network, that is, the loss in the fault missed detection scenario; cost2 is the cost coefficient of misclassifying a fault other than excessive antenna tilt angle as an excessive antenna tilt angle fault, that is, the loss in the fault false alarm scenario of this type; cost3 is the cost coefficient of misclassifying a fault other than excessive antenna tilt angle as an excessive antenna tilt angle fault, that is, the loss in the fault false alarm scenario of this type; and cost4 is the cost coefficient of misclassifying a fault other than excessive power as an excessive power fault, that is, the loss in the fault false alarm scenario of this type.

[0035] It should be noted that the cost coefficients must satisfy cost1>cost2>cost3, cost3 and cost4 can be of similar value, and cost1 and cost2 are much larger than cost3 and cost4. For example, cost1 can be set to 2000, cost2 can be set to 1000, and cost3 and cost4 can be set to 1, in order to reduce the probability of missed fault detection.

[0036] Next, after inputting the actual indicator data into the target detection model, the four classifiers will each output a signed score. A positive score with a larger absolute value indicates that the classifier is more certain that the network to be detected belongs to the network state it is responsible for; a negative score with a larger absolute value indicates that the classifier is more certain that the network to be detected does not belong to the network state it is responsible for. Subsequently, the model combines the preset cost weights and the scores to calculate the total cost required to classify the network to be detected into each category, thereby amplifying the potential loss of missed fault detection through high weights. In this way, even if the score of the normal state classifier is higher, its corresponding total cost will be much higher than the total cost of the fault category. Finally, the model selects the state type with the lowest total cost as the final fault detection result output. As shown in the example above, although the first classifier outputs the highest score of +60, after calculation with the high cost coefficient of 2000, the cost of its corresponding missed fault detection is much higher than that of the other three categories. The cost coefficient for a low power fault is only 1. Even if the score of the fourth classifier is lower, the cost is far lower than the risk loss of classifying it as a normal state. Therefore, the model will be more inclined to choose the fault category with the smaller total cost (i.e., classifying it as a low power fault) so as not to miss the real fault.

[0037] S130. If the fault detection results indicate that the network under test is in the target fault type, the target twin network is used to perform parameter optimization processing on the network under test to obtain the target parameter tuning strategy.

[0038] The target fault type can refer to the specific fault type of the network under test as determined by the fault detection results, namely, any one of the following: antenna tilt angle too small fault, antenna tilt angle too large fault, or power too low fault.

[0039] It should be noted that the characteristics of network performance changes and the impact of base station parameter adjustments on communication quality differ significantly depending on the type of fault the network under test is in. Therefore, to achieve accurate and effective fault recovery, it is necessary to optimize the parameters for the target fault type of the network under test in order to find the optimal base station parameter configuration that can restore network communication quality to a normal level.

[0040] Understandably, the parameter optimization process requires repeated verification of network performance changes corresponding to different parameter adjustments, which necessitates a large amount of data reflecting network state changes under the target fault type. However, in actual network operation, obtaining data for specific fault types is extremely difficult. If trial-and-error parameter tuning is carried out directly in the actual network under test, it will not only continuously interfere with user terminals that are communicating normally within the network, but also require waiting for multiple rounds of performance data reports from user terminals, resulting in extremely low optimization efficiency and failing to meet the need for rapid network fault recovery.

[0041] To overcome the above difficulties, this solution introduces digital twin technology. For each target fault type, a corresponding target twin network is pre-constructed to accurately simulate the communication state of the network under test when it is in the target fault type, thereby providing a virtual simulation environment that allows for a large number of parameter adjustment attempts for parameter optimization.

[0042] The target twin network can refer to a virtual simulation network built based on digital twin technology, capable of simulating network communication states under specific fault types. In other words, the target twin network corresponds to the target fault type and can accurately simulate the operating state of the real network under the target fault type and the changes in key network performance indicators after base station parameter adjustments.

[0043] Specifically, after obtaining the fault detection result, if the result indicates that the network under test has experienced a certain type of target fault, the first step is to determine the target twin network corresponding to that target fault type. This target twin network can generate twin samples consistent with the distribution of real network data under the target fault type, thereby constructing a virtual simulation environment. In this virtual simulation environment, various parameter adjustment operations can be safely attempted without affecting the operation of the real network. Subsequently, based on the network configuration parameters of the current network under test, the initial state and action space are determined. Then, using the target deep Q-network, parameter optimization is performed in the virtual simulation environment constructed by the target twin network. That is, an action is selected from the action space according to the current state, the action is executed in the virtual simulation environment and a reward feedback is obtained, and the parameter selection strategy is updated according to the reward value. Then, the iterative process of action selection, simulation, and reward calculation is repeated until the preset iteration termination condition is reached. Finally, the base station parameter configuration scheme with the optimal reward result is output as the target parameter tuning strategy.

[0044] The target parameter tuning strategy can refer to a parameter adjustment scheme for restoring normal network performance, i.e., a set of specific adjustment instructions for the adjustable configuration parameters of the base station. For example, taking the base station transmit power and antenna tilt angle as adjustable configuration parameters, the target parameter tuning strategy can include the first configuration values ​​of the transmit power and antenna tilt angle of the faulty base station to directly correct the parameters of the faulty base station and solve the signal anomaly problem.

[0045] Optionally, to further improve the communication quality of all user terminals across the network, in addition to repairing the faulty base station itself, the parameters of adjacent base stations can be adjusted collaboratively to compensate for coverage blind spots and suppress interference, thereby achieving global optimization. Therefore, the target parameter tuning strategy can also include second configuration values ​​for the transmit power and antenna tilt angle of adjacent base stations around the faulty base station. These values ​​guide the adjacent base stations to collaboratively adjust their transmit power and antenna tilt angle; for example, appropriately increasing the transmit power of adjacent base stations to compensate for the coverage blind spots of the faulty base station, or fine-tuning the antenna tilt angle of adjacent base stations to suppress their interference to edge users. By correcting the parameters of the faulty base station itself, combined with the collaborative adjustment of the parameters of adjacent base stations, the global optimal recovery of the entire network's communication performance can be achieved.

[0046] S140. Perform fault recovery operations based on the target parameter tuning strategy to complete network fault handling.

[0047] Specifically, after determining the target parameter tuning strategy, the configuration values ​​contained in the strategy can be encapsulated according to a configuration protocol format recognizable by the base station to generate corresponding parameter configuration instructions, which are then sent to the corresponding target base station through the network management interface. Upon receiving the parameter configuration instructions, the target base station can update its own adjustable configuration parameters according to the values ​​in the instructions, restoring the performance of the network under test to normal levels, thereby completing the network fault recovery operation.

[0048] The target base station is the base station that needs to perform parameter configuration updates, which may include base stations that have experienced failures. After determining the target parameter tuning strategy, the faulty base station can make corresponding configuration adjustments based on the first configuration value to eliminate the fault at its root.

[0049] Optionally, the target base station may also include neighboring base stations around the faulty base station. While the faulty base station is being adjusted, the second configuration value in its corresponding target parameter tuning strategy is used to adjust the parameters of the neighboring base stations for collaborative repair, to compensate for the coverage blind spots of the faulty base station, and to suppress interference spread, thereby achieving network-wide collaborative fault recovery.

[0050] It should be noted that the embodiments in this specification are described using a single fault type as an example, but in actual networks, multiple faults may coexist. In this case, the fault detection and recovery steps of this method can be repeated. After detecting and recovering one fault, network indicators are collected again and the detection and recovery process is repeated until all networks return to normal.

[0051] In the above implementation, by acquiring the actual indicator data of the network under test, the communication quality status of the network can be monitored in real time, providing a reliable data foundation for fault detection. Using a target detection model based on actual indicator data for fault detection can quickly and accurately identify whether a fault exists in the network and its specific type, providing a basis for subsequent targeted processing. Furthermore, when a fault is detected, parameter optimization is performed using a target twin network corresponding to the target fault type. This allows for the safe and efficient exploration of the optimal parameter adjustment scheme in a virtual simulation environment, avoiding the risks of trial and error in a real network. Finally, fault recovery operations are performed based on the target parameter tuning strategy, applying the optimized parameter configuration to the actual network to achieve rapid network performance recovery. Thus, a fully collaborative network fault handling solution is constructed, achieving automated detection and rapid fault repair, improving network operation and maintenance efficiency and user communication experience.

[0052] In some implementation methods, please refer to the appendix. Figure 2 The target detection model is obtained through the following method: S210. Determine the normal performance data and fault performance data of the sample network.

[0053] The sample network can refer to a wireless network used to train and test the target detection model. Its network architecture can be the same as the network to be detected, containing multiple base stations and multiple user terminals, and can reflect the operating characteristics of the real network.

[0054] Normal performance data refers to key performance indicator data collected by the sample network under normal operating conditions, which can be used to reflect the network communication quality of the wireless network under normal circumstances. Fault performance data refers to key performance indicator data generated when the wireless network experiences a preset type of fault. It is generated based on the normal performance data of the sample network and can be used to reflect the network communication quality of the wireless network under fault conditions.

[0055] It should be noted that, for the convenience of subsequent model processing, status identifiers can be used. l Differentiate the different operating states of the sample network. For example, consider antenna tilt angle too small fault, antenna tilt angle too large fault, and power too low fault, where l=0 indicates the sample network is in normal operation, l=1 indicates an antenna tilt angle too small fault, l=2 indicates an antenna tilt angle too large fault, and l=3 indicates insufficient transmit power fault. Furthermore, this is only an illustrative example; in other implementations, l∈{4,…,L} can represent other possible fault states of the sample network.

[0056] Specifically, the normal performance data and fault performance data of the sample network can be determined using the following method. First, normal performance data is determined based on the minimum drive test report of the wireless network under normal communication conditions. Then, pseudo-fault calculations are performed based on the normal performance data and fault distance data to obtain fault performance data.

[0057] Specifically, a method similar to obtaining actual indicator data in S110 above can be adopted. That is, when the sample network is in normal working condition (l=0), the RSRP and SINR values ​​of each user terminal are extracted from the V groups of MDT reports reported by user terminals in the network according to a preset period. These values ​​are then organized into the v-th wireless network key performance indicator vector ∈ {1, 2, ..., V} according to the collection batch. Vectors, multiple batches of KPI vectors form complete normal performance data. .

[0058] After obtaining normal performance data, pseudo-fault calculations can be performed based on the normal performance data and fault distance data to obtain fault performance data.

[0059] Among them, fault distance data can refer to the spatial distance information between the user terminal and the base station, which can quantify the path loss during signal propagation and thus simulate the signal change pattern when a fault occurs.

[0060] It should be noted that the effective signal received by a user terminal in a wireless network is the result of the combined effect of the useful signal sent by the serving base station and the interference signals sent by neighboring base stations. Therefore, when simulating key performance indicators under fault conditions, it is necessary to consider both the impact of the faulty base station on the signal strength of the user terminal and the interference impact of neighboring base stations on the user terminal in order to ensure that the generated pseudo-fault data truly reflects the network state under fault scenarios.

[0061] Specifically, the fault distance data may include first distance data between the user terminal and the preset faulty base station and at least one second distance data between the user terminal and each of the preset faulty base stations adjacent to the base station.

[0062] The preset faulty base station can be a pre-selected base station used to simulate a fault. The first distance data can refer to the straight-line distance between the user terminal and the preset faulty base station, and the second distance data can refer to the straight-line distance between the same user terminal and a neighboring base station in the vicinity of the preset faulty base station. A neighboring base station can be a base station that is in contact with the boundary of the service coverage area of ​​the preset faulty base station. For example, please continue to refer to... Figure 1b The neighboring base stations of base station BS1 are BS2, BS3 and BS4.

[0063] For example, if there are N base stations and M user terminals in the sample network, and the cell radius of each base station (i.e., the coverage area of ​​each base station) is R, then the distance from the user terminal m to the preset faulty base station is... The distance is the first distance data. The user terminal m is connected to the faulty base station. The distance between the adjacent base stations n is the second distance data. Specifically, the aforementioned distance can be calculated by combining the location information reported by the user terminal with the deployment coordinates of the base station.

[0064] Furthermore, after obtaining the fault distance data, pseudo-fault calculation can be performed in the following way: using the first distance data, normal performance data, and second distance data to determine the baseline fault data and fault adjustment amount, and based on the fault adjustment amount and the baseline fault data, the fault performance data can be determined.

[0065] Among them, the baseline fault data can refer to the change in the signal received by the user terminal under fault conditions when only considering the influence of the preset faulty base station itself.

[0066] Specifically, based on normal samples from the collected normal performance data, pseudo RSRP and pseudo SINR caused by abnormal parameters of the pre-faulted base station to the user terminal can be generated using corresponding calculation formulas for different fault types, without considering interference from neighboring cells, as baseline fault data.

[0067] Fault adjustment refers to the change in interference from adjacent base stations to user terminals, reflecting the fluctuation of signal strength between adjacent base stations under fault conditions. It should be noted that fault adjustment is only used to correct the calculation of the signal-to-interference-plus-noise ratio and does not affect the value of the reference signal received power.

[0068] It should be noted that, due to the different physical mechanisms of different faults, the degree of impact of different types of faults on user terminal signals varies significantly.

[0069] Specifically, for antenna tilt-related faults, when a base station experiences an antenna tilt angle that is too small, its main lobe will shift to a distant location. Conversely, an antenna tilt angle that is too large will cause the main lobe to sink excessively. In such cases, different user terminals located within different coverage areas of the base station will be affected to varying degrees. The degree of impact can be determined based on the relationship between the distance of each user terminal to the faulty base station and the coverage area of ​​the faulty base station.

[0070] For example, the coverage area of ​​a base station can be roughly divided into edge locations and near-center locations. Taking two-thirds of the coverage area as the dividing threshold, if the distance from the user terminal to the faulty base station is less than the coverage area of ​​the faulty base station but greater than two-thirds of its coverage area, then the user terminal is determined to be located at the edge of the faulty base station; if the distance from the user terminal to the faulty base station is less than two-thirds of the coverage area, then the user terminal is determined to be located at the near-center of the faulty base station.

[0071] Furthermore, when a user terminal is located at the edge of a faulty base station, it will also be affected by adjacent base stations. Similarly, the degree of this impact can be determined based on the relationship between the distance of each user terminal to an adjacent base station and the coverage area of ​​that adjacent base station.

[0072] For example, taking an impact threshold of five-thirds of the coverage area of ​​an adjacent base station as an example, if the distance from the user terminal to the adjacent base station of the faulty base station is greater than the coverage area of ​​the adjacent base station but less than five-thirds of its coverage area, then the user terminal is considered to be within the influence range of the adjacent base station, and it is determined that the user terminal will be affected by the adjacent base station; if the distance from the user terminal to the adjacent base station of the faulty base station is less than the coverage area of ​​the adjacent base station, then the user terminal is considered to be outside the influence range of the adjacent base station, and it is determined that the user terminal will not be affected by the adjacent base station.

[0073] For power-related faults, since these faults manifest as abnormal base station transmit power (such as too low, too high, or unstable power), all user terminals within its coverage area will be affected, unlike antenna tilt faults where spatial main lobe offset leads to varying degrees of impact on user terminals located in different coverage areas. In other words, when a base station experiences a power-related problem, all user terminals within its coverage area can be considered to be affected to the same extent, while user terminals outside the coverage area are unaffected.

[0074] It is important to emphasize that the aforementioned boundary thresholds (such as two-thirds or five-thirds of the coverage area) are merely example values ​​set based on typical engineering experience. Their specific values ​​can be flexibly adjusted according to the actual network deployment environment. That is, in practical applications, those skilled in the art can customize distance ranges for different fault types based on the physical characteristics of the fault. For example, for a fault with an excessively small antenna tilt angle, if the user terminal is located at the edge of the coverage area of ​​a preset faulty base station (such as...) And simultaneously located within the influence range of adjacent base stations (such as...) Therefore, the baseline fault data under the fault of too small antenna tilt angle can be determined by the following formula: In the formula, This represents the pseudo RSRP of the m-th user terminal under EU failure during the v-th data acquisition. This represents the RSRP of the m-th user terminal under normal operating conditions during the v-th data acquisition. This represents the pseudo SINR of the m-th user terminal under EU fault during the v-th data acquisition. h1 represents the SINR of the m-th user terminal under normal working conditions during the v-th data acquisition; h1 is the first quantization parameter, which can be 3.

[0075] If the user terminal is located near the midpoint of a pre-set faulty base station (e.g.) Therefore, the baseline fault data under the fault of too small antenna tilt angle can be determined by the following formula: In the formula, h2 is the second quantization parameter, which can be 2.

[0076] Furthermore, regarding the fault of excessive antenna tilt angle, if the user terminal is located at the edge of the coverage of a pre-set faulty base station (e.g., And simultaneously located within the influence range of adjacent base stations (such as...) Therefore, the baseline fault data under an excessive antenna tilt fault can be determined by the following formula: In the formula, This represents the pseudo RSRP of the m-th user terminal under ED failure during the v-th data acquisition. h3 represents the pseudo SINR of the m-th user terminal under ED fault during the v-th data acquisition; h3 is the third quantization parameter, which can be 11.

[0077] Furthermore, regarding excessive power reduction faults, if the user terminal is located within the coverage area of ​​a pre-defined faulty base station (e.g., Therefore, the baseline fault data under excessive power reduction fault conditions can be determined by the following formula: In the formula, This represents the pseudo RSRP of the m-th user terminal under ERP failure during the v-th data collection. h4 represents the pseudo SINR of the m-th user terminal under ERP failure during the v-th data acquisition; h4 is the fourth quantization parameter, which can be 10.

[0078] It should be noted that the specific values ​​of all the above-mentioned quantitative parameters are empirical values ​​derived from simulation experiments or field test data in typical network scenarios. In practical applications, these parameters can be flexibly adjusted and optimized according to the network deployment environment and operation and maintenance requirements. The values ​​listed in this embodiment are merely examples.

[0079] After obtaining the baseline fault data, the fault adjustment amount for each adjacent base station can be further determined.

[0080] Specifically, for each adjacent base station, the same distance interval judgment rules and quantization parameters used in calculating the baseline fault data are adopted. Based on the second distance data from the user terminal to the adjacent base station, the change in the signal of the adjacent base station relative to the normal operating state is judged, and used as the fault adjustment amount of the adjacent base station.

[0081] For example, for an antenna tilt angle too small fault, if the user terminal is located at the edge of the coverage of an adjacent base station (e.g., And at the same time, it is located outside the coverage area of ​​the preset faulty base station (such as...). Therefore, the fault adjustment amount under the fault of too small antenna tilt angle can be determined by the following formula: In the formula, represents the fault adjustment amount of the m-th user terminal under EU fault during the v-th data acquisition; h1 is the first quantization parameter, which can be 3.

[0082] If the user terminal is located near the midpoint of an adjacent base station (e.g.) Therefore, the fault adjustment amount under the fault of too small antenna tilt angle can be determined by the following formula: In the formula, h1 represents the fault adjustment amount of the m-th user terminal under EU fault during the v-th data acquisition; h2 is the second quantization parameter, which can be 2.

[0083] For antenna tilt angle too large faults, if the user terminal is located at the edge of the coverage of an adjacent base station (e.g., And at the same time, it is located outside the coverage area of ​​the preset faulty base station (such as...). Therefore, the fault adjustment amount under the fault of excessive antenna tilt angle can be determined by the following formula: In the formula, The table shows the fault adjustment amount of the m-th user terminal under ED fault during the v-th data collection; h3 is the third quantization parameter, which can be 11.

[0084] For excessive power reduction faults, if the user terminal is located at the coverage edge of an adjacent base station (e.g., Therefore, the fault adjustment amount under excessive power reduction fault can be determined by the following formula: In the formula, h4 represents the fault adjustment amount of the m-th user terminal under ERP fault during the v-th data collection; h4 is the fourth quantization parameter, which can be 10.

[0085] Furthermore, fault performance data can be obtained by superimposing and calculating fault adjustment amounts and baseline fault data.

[0086] Specifically, after obtaining the fault adjustment amounts for each adjacent base station, they can be processed separately according to signal type. First, since the interference changes of adjacent base stations do not affect RSRP, the RSRP value in the final fault performance data can directly adopt the pseudo RSRP value in the baseline fault data under the corresponding fault condition.

[0087] The SINR value in the fault performance data can be calculated by superimposing it with the SINR values ​​generated by all adjacent base stations under normal operating conditions.

[0088] Specifically, firstly, the baseline pseudo-SINR value of the user terminal is extracted from the baseline fault data. Then, all adjacent base stations of the preset fault base station are traversed, and the fault adjustment amount of each adjacent base station corresponding to the user terminal is sequentially added to the baseline pseudo-SINR value to obtain the final pseudo-SINR value of the user terminal.

[0089] For example, taking the antenna tilt angle too small (EU) fault scenario as an example, assume the preset faulty base station is base station 4, and its neighboring base stations are base stations 1, 2, 3, 5, 6, and 7, a total of 6 base stations. The baseline pseudo SINR value of a certain user terminal m is 22dB. After traversing the neighboring base stations, the corresponding fault adjustment amounts for each base station are as follows: base station 1 corresponds to adjustment amount +3dB, base station 2 corresponds to adjustment amount -2dB, base station 3 corresponds to adjustment amount 0dB, base station 5 corresponds to adjustment amount +3dB, base station 6 corresponds to adjustment amount -2dB, and base station 7 corresponds to adjustment amount 0dB. Then, using 22dB as the initial value, by sequentially adding each adjustment amount, the final pseudo SINR value of the user terminal is calculated to be 22 + 3 - 2 + 0 + 3 - 2 + 0 = 24dB.

[0090] It should be noted that the above calculation process is based on a user terminal m and a preset faulty base station. This is accomplished by one of its adjacent base stations, n. Since the sample network contains M user terminals and multiple sets of fault performance data need to be generated, the above calculation process needs to be repeated for each user terminal.

[0091] Specifically, for each set of wireless network parameter configurations, a counter is first initialized, and all M user terminals are iterated. For each user terminal, its baseline fault data is first determined, then all adjacent base stations are iterated, and the fault adjustment amount of each adjacent base station is calculated one by one. The above superposition calculation process is repeated for each user terminal until the pseudo-KPIs calculation for all M user terminals in the v-th data acquisition is completed. Subsequently, the pseudo RSRP and pseudo SINR of the M user terminals are combined into a pseudo-KPIs vector. The above process is repeated until all V data acquisitions are completed, finally obtaining V pseudo-KPIs vectors under a fault category, which constitute the fault performance data corresponding to that fault type. That is, the above process is continuously repeated to obtain the fault performance data under EU faults. Fault performance data under ED failure Fault performance data under EPR fault conditions .

[0092] In this way, the baseline fault data provides the direct impact of abnormal parameters of the faulty base station on the useful signal, while the fault adjustment quantity quantifies the dynamic changes in the interference level of adjacent base stations. The fault performance data generated by integrating the two can comprehensively reflect the actual composition of the signal received by the user terminal under the fault scenario, and provide accurate and diverse fault sample data for the subsequent training of the target detection model.

[0093] In the above implementation, normal performance data is first determined using MDT reports from normal communication scenarios, ensuring that the subsequently generated fault samples are highly consistent with the basic characteristics of the real network. Then, by utilizing the physical correlation between distance and signal strength, and through preset adjustment rules, the variation patterns of signal strength under different fault types are simulated. This enables the batch generation of fault performance data covering multiple fault types without the need for real fault samples, solving the problem of obtaining real fault samples. This provides sufficient training samples with accurate state labels for subsequent support vector machine training, ensuring the training effect of the target detection model.

[0094] S220. Support vector machine training is performed based on normal performance data and fault performance data to obtain the target detection model.

[0095] Understandably, Support Vector Machines (SVMs) are supervised learning models used for classification. Their core idea is to find an optimal hyperplane in the feature space that maximizes the margin between samples of different classes, thereby enabling the classification of unknown samples.

[0096] Specifically, the target detection model can be obtained by training a support vector machine based on normal performance data and fault performance data. This can be achieved as follows: First, a digital twin network is constructed based on the normal performance data and fault performance data; then, the digital twin network is used to generate data to obtain target training samples; finally, the basic detection model is trained with a support vector machine based on the target training samples and target state labels to obtain the target detection model.

[0097] In this context, a digital twin network can refer to a data generation network optimized through generative adversarial training, used to simulate the distribution of key performance indicators under different network states.

[0098] Specifically, the digital twin network can be composed of multiple independent sub-twin networks, each sub-twin network corresponding to a network state, which can be a normal state or various preset fault states, and can generate twin samples that are highly consistent with the real data distribution in that state based on random input.

[0099] It should be noted that generative adversarial training is a training method for iteratively optimizing generative adversarial networks (GANs).

[0100] As is understandable, a Generative Adversarial Network (GAN) consists of two adversarial neural networks: a generator network and a discriminator network. The generator network is responsible for generating simulated data based on random vectors, while the discriminator network is responsible for determining whether the input data is a real sample or a fake sample generated by the generator network. Through iterative game-like interaction and alternating optimization, the two networks eventually enable the generator network to produce fake samples sufficient to deceive the discriminator network.

[0101] Specifically, the objective function of the generative adversarial network can be defined as: In the formula, G represents the generator network. D represents the trainable parameters of the generative network; D represents the discriminative network. To identify the trainable parameters of the network; Z is an input matrix composed of random vectors, following a pre-defined random distribution. X represents the input real data, i.e., normal performance data, which follows the distribution of real data. D(X) is the classification output of the discriminator network for the real data X, representing the probability that the discriminator network judges the sample as real data; G(Z) is the simulated data generated by the generator network based on the random vector Z, i.e. the target training sample; D(G(z)) is the classification output of the discriminator network for the generated simulated data.

[0102] Specifically, by constructing digital twin networks based on normal performance data and fault performance data, corresponding generative adversarial networks can be built for each operating state of the wireless network, forming a complete set of digital twin networks.

[0103] For example, taking four network states—normal operation, antenna tilt angle too small fault, antenna tilt angle too large fault, and power too low fault—as examples, four sets of sub-twin networks with completely identical structures can be constructed accordingly. Specifically, for the normal operation state (l=0), using normal performance data... Substituting X into the objective function of the generative adversarial network described above, a sub-Siamese network is trained to obtain the sub-Siamese network. For EU faults (l=1), the fault performance data under EU faults is used. As input X, the sub-Siamese network is trained. For ED faults (l=2), the fault performance data of ED faults are used. As input X, the sub-Siamese network is trained. For ERP failures (l=3), the failure performance data under ERP failure conditions is used. As input X, the sub-Siamese network is trained. Each sub-twin network is trained independently until the discrimination network's accuracy in distinguishing between generated and real samples stabilizes at around 50%. This indicates that the generative network can generate samples that are highly consistent with the distribution of real data. At this point, the training of the generative adversarial network can be considered complete, thus obtaining a complete digital twin network.

[0104] Furthermore, after obtaining the digital twin network, data can be generated to obtain the target training samples.

[0105] The target training sample can refer to the key performance indicators of the simulated wireless network generated by the digital twin network. It has a corresponding target state label, which can be an identifier used to identify the network state to which the target training sample belongs, and can reflect the network state corresponding to the target training sample, such as a normal state or a certain preset fault state.

[0106] Specifically, each sub-twin network in a digital twin network can generate twin performance data under the corresponding network state, and the generated samples are automatically matched with the corresponding target state labels.

[0107] For example, for the normal operating state (l=0), a random vector Z is input into the sub-Siamese network. This generates N twin performance data sets, forming a sample set. Where N is the number of generated samples (e.g., 100), these Siamese performance data are the target training samples under normal conditions, and the corresponding target state labels are all 0. For EU faults (l=1), the random vector Z is input into the sub-Siamese network. Similarly, N twin performance data points are generated to form a sample set. The corresponding target state label is 1. Similarly, for ED faults (l=2), N twin performance data are generated to form a sample set. The corresponding target state label is 2. For ERP failures (l=3), N twin performance data are generated to form a sample set. The corresponding target state label is 3. Merging all twin performance data and their corresponding target state labels constitutes a complete target training sample. Furthermore, each training sample X carries a clear target state label.

[0108] Furthermore, after obtaining the target training samples, the basic detection model can be trained using a support vector machine based on the target training samples and the target state labels to obtain the target detection model.

[0109] The basic detection model has the same structure as the object detection model, both containing multiple independent classifiers, each responsible for determining a network state. However, its classification accuracy is not high, requiring support vector machine training using target training samples and target state labels.

[0110] Specifically, for the first Network state ( The training can be performed using a one-to-one complementary strategy. That is, for a certain classifier, the twin performance data of a certain network state corresponding to it is used as a positive example (classification label is +1), and all the twin performance data of the other network states are used as negative examples (classification label is -1), so as to simplify the complex multi-class problem into a binary classification problem for training.

[0111] For example, if the hyperplane normal vector of the e-th classifier is defined as... The intercept is The optimization objective of the binary classification problem can be expressed as: In the formula, Represents the magnitude of the normal vector; This is the training sample for the nth target. For the input category label, if If it belongs to class e, then ,otherwise ; These are slack variables; Represents the nth target training sample The distance to the boundary of class e is: ; is the soft margin penalty parameter for the classifier of class e, used to balance the margin size and misclassification loss, and the optimal value can be determined through cross-validation; N is the total number of target training samples.

[0112] Furthermore, the above optimization problem can be transformed into a dual problem by introducing Lagrange multipliers and utilizing the Karush-Kuhn-Tucker conditions, ultimately yielding the optimal solution. and .

[0113] Meanwhile, to reduce the risk of missed fault diagnosis, as mentioned above, a preset cost weight (Cost) can be introduced and integrated into the decision logic of the detection model to assist in the complete training of the classifier. When the preset training conditions are met (e.g., the number of training iterations is reached), the training is considered complete, and a target detection model with fault classification capabilities is finally obtained.

[0114] It is understandable that, in the application phase, the actual performance data of the network to be detected is used in conjunction with the object detection model. When making decisions, each classifier in the object detection model outputs a decision value (i.e., ), to indicate the input The signed distance to the classification hyperplane can be used to express the decision-making process of the object detection model: In the formula, This represents the network state output by the target detection model, i.e., the fault detection result of the network to be detected (0 represents normal working state, 1 represents antenna tilt angle too small fault, 2 represents antenna tilt angle too large fault, 3 represents power too low fault). This represents the decision value of the classifier for class e, with positive values ​​indicating... A value indicates that a class is defined as belonging to it; a negative value indicates that a class is not belonging to it; the larger the absolute value, the higher the confidence level. The preset cost weight represents the e-th class, which is used to amplify or reduce the risk of classifying the sample as the e-th class.

[0115] As can be seen from the above formula, the decision rule is essentially to select the category with the lowest weighted cost when classifying a sample as that category among all categories, thereby ensuring overall accuracy while prioritizing the avoidance of high-cost misclassifications (such as misdiagnosing a fault as normal).

[0116] By introducing generative adversarial networks (GANs) to construct digital twin networks, the game-like interaction between the generator and discriminator networks allows the generator to learn the inherent patterns of real data distribution under normal and various fault states. This enables the generator to produce twin samples that are highly consistent with the original sample distribution, solving the problems of difficulty in obtaining and labeling fault samples. Furthermore, these twin samples can be used as training data for support vector machines (SVMs), allowing the SVMs to be trained on sufficient samples with accurate state labels, providing a data foundation for the subsequent precise classification of boundaries.

[0117] In the above implementation, normal performance data is first obtained through a minimized drive test report under normal communication scenarios, ensuring the reliability and authenticity of the data source. Then, based on the normal performance data and distance information, fault performance data covering multiple fault types is generated, constructing a complete training sample set including normal states and various fault states. Finally, these samples are used for training to obtain a target detection model capable of accurately distinguishing between normal states and various fault types, thus providing a reliable diagnostic basis for subsequent fault recovery.

[0118] In some implementation methods, please refer to the appendix. Figure 3 If the fault detection results indicate that the network under test is in the target fault type, parameter optimization processing is performed on the network under test to obtain the target parameter tuning strategy, including: S310. Determine the initial state and action space based on the network configuration parameters of the network to be detected.

[0119] Network configuration parameters refer to a set of indicators that reflect the operating characteristics of a wireless network base station, such as base station transmit power, antenna tilt angle, or antenna azimuth angle, which can reflect the current network coverage and interference level.

[0120] Specifically, network configuration parameters can include current configuration parameters and adjustable configuration parameters.

[0121] For example, taking base station transmit power and antenna tilt angle as network configuration parameters, the current configuration parameters may include the current transmit power and the current antenna tilt angle. Adjustable configuration parameters include the transmit power adjustment amount and the antenna tilt angle adjustment amount.

[0122] Assume the network to be detected contains The base station, the first The current transmit power of each base station is denoted as: The current antenna tilt angle is denoted as Network configuration parameters It can be represented as a set of all base station parameters: In the formula, Represents the current transmit power of N base stations; This represents the current antenna tilt angle of N base stations.

[0123] The adjustable range of the configuration parameters for each base station can be seen as follows: In the formula, This represents the transmit power of the nth base station under normal network conditions. This represents the antenna tilt angle of the nth base station under normal network conditions; This represents the maximum adjustable transmit power of the nth base station; This represents the minimum adjustable transmit power of the nth base station; This represents the maximum adjustable antenna tilt angle of the nth base station; This represents the minimum adjustable antenna tilt angle of the nth base station; This represents the adjustment amount of the transmit power of the nth base station; This represents the antenna tilt adjustment amount for the nth base station. It should be noted that the transmit power adjustment amount... Antenna tilt adjustment amount It can be set by the operator according to the specific network conditions.

[0124] For example, taking the network to be detected in S110 (N=7, M=1000) as an example, its network configuration parameters have a total of L=139972 combinations of configuration parameters, where the l-th combination of the current configuration parameters in {1, 2, ..., L} can be as follows: In the formula, This represents the current transmit power of the seven base stations; This represents the current antenna tilt angle of the 7 base stations. l=0 indicates the base station is in normal working condition, and l=1 indicates a pre-set faulty base station. An antenna tilt angle too small fault has occurred; l=2 indicates a preset faulty base station. An antenna tilt angle overshoot fault has occurred; l=3 indicates a pre-set faulty base station. An excessive power reduction fault has occurred.

[0125] Assume the pre-set faulty base station is the fourth base station (i.e. Taking the example of a network with a transmit power of 1 and an antenna tilt angle of 40° under normal network conditions, and considering the basic parameters of the network under test, the specific parameter combinations for each fault condition are as follows: It should be noted that in actual wireless networks, base station transmit power and antenna tilt angle are parameters that can be finely adjusted. They are usually adjusted within a preset continuous value range according to a fixed adjustment step size (e.g., power step size 0.1, tilt angle step size 1°).

[0126] For example, the adjustable range of the adjustable configuration parameters of the neighboring base stations around the preset faulty base station can be: The transmit power adjustment range of adjacent base stations is from a minimum value of 1 to a maximum value of 3. Assuming that 1 is the normal reference value, discrete adjustable values ​​can be taken within this range according to a preset power step size (e.g., 0.1). The antenna tilt angle adjustment range of adjacent base stations is from a minimum value of 30° to a maximum value of 50°. Assuming that 40° is the normal reference value, discrete adjustable values ​​can be taken within this range according to a preset tilt angle step size (e.g., 1°).

[0127] It is understood that the adjustment range, normal baseline value and adjustment step size of the above adjustable configuration parameters are only illustrative examples, and their specific values ​​can be flexibly adjusted according to the actual network deployment environment and operation and maintenance needs.

[0128] Furthermore, if a fault is found in the network under test after detection, a parameter optimization repair process can be performed, which can be done through the target depth Q network.

[0129] The target deep Q-network can refer to a reinforcement learning model built on a deep Q-network (DQN) to find the optimal parameter adjustment strategy in a preset action space. Its core is to fit the state-action value function through a deep neural network. It can predict the long-term benefits of different parameter adjustment actions based on the current network state, and thus select the optimal action that maximizes the benefits.

[0130] It should be noted that the core operating logic of the deep Q-network is as follows: In the current state, the neural network predicts the state-action Q value corresponding to each executable action, selects the action with the largest Q value to execute, and obtains the immediate reward corresponding to the action through interaction with the environment. After storing the experience data of state, action, reward, and next state, the neural network parameters are iteratively optimized through the experience replay mechanism, so that the neural network's prediction of action value becomes more and more accurate, and finally converges to the state that can output the optimal parameter tuning strategy.

[0131] Furthermore, when configuring the parameters of the target depth Q-network, the current configuration parameters (current transmit power and current antenna tilt angle) of the network under test at the time of the fault occurrence can be used as the initial state for optimization. The adjustable configuration parameters (transmit power adjustment and antenna tilt angle adjustment) allowed by the network under test are used as executable actions to construct the action space.

[0132] Specifically, action space a can be defined in the following way: In the formula, This represents the adjustment amount of the transmit power of the nth base station; This represents the antenna tilt adjustment amount for the nth base station.

[0133] State parameters It can be set to: In the formula, This represents the current transmit power of the 1st to Nth base stations at time t; This represents the current antenna tilt angle of the 1st to Nth base stations at time t.

[0134] For example, taking the network to be detected in S110 (N=7, M=1000) as an example, its action space can be defined as a set containing the following four types of actions: increasing the transmit power of any base station. Reduce the transmit power of any base station Increase the antenna tilt angle of any base station Reduce the antenna tilt angle of any base station That is, the action space 'a' can be set as: Its state parameters It can then be set as: In the formula, This represents the current transmit power of the nth base station (n=1,2,3,4,5,6,7) at time t; This represents the current antenna tilt angle of the nth base station (n=1,2,3,4,5,6,7) at time t. Its value range is as follows: It should be noted that the values ​​listed above are merely exemplary discrete values ​​set based on specific fault types (antenna tilt angle too small, antenna tilt angle too large, power too low), and do not imply that the value range of the faulty base station is limited to these three combinations. For example, assuming the normal baseline value for antenna tilt angle is 40°, although the above uses a current antenna tilt angle of 30° to represent a fault of antenna tilt angle too small, it is not strictly limited to 30°; it can also be 25° or 20°, etc. Similarly, assuming the normal baseline value for transmit power is 1, although the above uses a current transmit power of 0.1 to represent a fault of power too low, it is not strictly limited to 0.1; it can also be 0.2 or 0.3, etc. In practical applications, the discrete granularity and specific values ​​can be flexibly set according to network operation and maintenance requirements.

[0135] S320. Using the target depth Q-network and the target twin network, parameter optimization is performed in the action space to obtain the target parameter tuning strategy.

[0136] Specifically, parameter optimization in the action space using a target depth Q-network and a target Siamese network can include the following steps: First, the initial state is taken as the current state; then, for the current state, the target depth Q-network is used to determine the current action to be performed; next, the target Siamese network is used to determine the current reward result of the current action; then, the current state is updated to the next state; then, the next state is taken as the new current state, and the above steps of determining the action and reward result are repeated; finally, when a preset termination condition is met, the current state is taken as the target parameter tuning strategy.

[0137] Here, the current state can refer to the set of network configuration parameters corresponding to the network to be detected at a certain moment during the parameter optimization iteration process, that is, the state parameter vector at that moment, which can completely represent the current configuration state of the wireless network. The currently executed action can refer to the specific adjustment operation performed on the base station parameters by the target depth Q network based on the current state and selected from the action space during a single iteration, that is, the adjustment amount of transmit power or antenna tilt angle.

[0138] For example, the initial state at the time of the fault is used as the first current state for parameter optimization. Then, the current state can be calculated using the deep neural network built into the target depth Q-network. Each executable action The corresponding Q value is given by the following formula: In the formula, The Q value represents the current state; The neural network parameters represent the target depth Q-network; Represents the current state. Represents the currently executed action; This represents the next state corresponding to the next time step (t+1); This represents the next action to be executed at the next moment (t+1), which is the action with the largest Q value; This represents the current reward result; This represents the discount factor.

[0139] It should be noted that during each iteration, the target depth Q-network calculates the Q-values ​​of all executable actions in the action space under the current state, and finally selects the action with the largest Q-value as the current action to be executed in this iteration. This ensures that the parameter adjustments at each step are made in the direction of maximizing the benefit. The above process can utilize... -greedy strategy (i.e., with) The probability of randomly selecting an action, in order to To achieve this, we choose the action with the highest Q value (with a probability of success), which means: After determining the current action, in order to evaluate the effect of the action on improving network communication quality and thus provide feedback for parameter optimization and action decision-making of the target depth Q network, a reward mechanism can be introduced to provide feedback on the quality of the action. The data that can quantify the effect of the action is the current reward result corresponding to the current action.

[0140] It is understandable that the current reward result can refer to the degree of improvement in network communication quality of the network under test after performing the current action. For example, the current reward result can be characterized by the number of user terminals in the network that meet the preset reward conditions.

[0141] Specifically, the current reward result of the current action can be determined by the following method: First, using the target Siamese network, Siamese performance data is generated based on the current action; then, using the Siamese performance data and preset reward conditions, the current reward result of the current action is determined.

[0142] Among them, twin performance data can refer to key performance indicators that can simulate the current execution action of the network under the target fault type at a certain moment. It is generated by the target twin network and can be used to reflect the network communication quality of the wireless network under the target fault type.

[0143] For example, if the fault detection result indicates that the target fault type of the network to be detected is a certain fault, then the corresponding target twin network is first determined in the digital twin network using the fault type. Then, after determining the current action to be executed, the state update formula is used... Calculate the network configuration parameters after executing the action. Then, input the updated transmit power and antenna tilt parameters of each base station into the target twin network. The target twin network can then simulate the changes in network signal under the target fault type based on the input network configuration parameters and the distance information between the user terminal and each base station, generate pseudo RSRP and pseudo SINR for each user terminal one by one, and finally integrate them to obtain the twin performance data corresponding to the currently executed action.

[0144] It should be noted that the process of generating twin performance data here is similar to the method of determining fault performance data in S210.

[0145] For example, assuming the base station cell radius is R, and the distance from the user terminal m to the preset faulty base station... The distance is denoted as the first distance data. The user terminal m is connected to the faulty base station. The distance between adjacent base stations n is denoted as the second distance data. .

[0146] If the current transmit power is greater than the transmit power in normal operation, the service range of the pre-set faulty base station ( The pseudo RSRP and pseudo SINR of user terminal m are: In the formula, This represents the fourth quantization parameter, which can be 10.

[0147] Not within the service area of ​​the faulty base station ( The pseudo RSRP and pseudo SINR of user terminal m are: In the formula, This represents the second quantization parameter, which can be 2.

[0148] If the current transmit power in the current state is less than or equal to the transmit power in the normal operating state, the pseudo RSRP and pseudo SINR of user terminal m within the service range of the faulty base station are preset as follows: In the formula, This represents the fifth quantization parameter, which can be 6.

[0149] The pseudo RSRP and pseudo SINR of user terminal m, which is not within the service range of the faulty base station, are: In the formula, This represents the sixth quantization parameter, which can be 1.

[0150] If the current antenna tilt angle is greater than the antenna tilt angle in normal operation, the pseudo RSRP and pseudo SINR of user terminal m within the service range of the faulty base station are preset as follows: In the formula, This represents the second quantization parameter, which can be 2.

[0151] The pseudo RSRP and pseudo SINR of user terminal m, which is not within the service range of the faulty base station, are: In the formula, This represents the seventh quantization parameter, which can be 4.

[0152] If the current antenna tilt angle is less than or equal to the antenna tilt angle in normal operation, the pseudo RSRP and pseudo SINR of user terminal m within the service range of the faulty base station are preset as follows: In the formula, This represents the eighth quantization parameter, which can be 8.

[0153] The pseudo RSRP and pseudo SINR of user terminal m, which is not within the service range of the faulty base station, are: In the formula, This represents the second quantization parameter, which can be 2.

[0154] To reiterate, the specific values ​​of all the above-mentioned quantitative parameters can be flexibly adjusted and optimized according to the network deployment environment and operation and maintenance requirements. The values ​​listed in this embodiment are only examples.

[0155] Subsequently, based on the above formula, all M user terminals in the twin network are traversed. After calculating the pseudo RSRP and pseudo SINR for each user terminal, the pseudo RSRP and pseudo SINR of all user terminals are concatenated in order to obtain the complete twin performance data corresponding to the currently executed action.

[0156] Furthermore, after determining the twin performance data, the current reward result of the currently executed action can be determined by combining it with preset reward conditions.

[0157] The preset reward condition can refer to a quantitative judgment rule used to measure whether the communication quality of the user terminal meets the standard. For example, the preset reward condition can be set to the user terminal's RSRP value being greater than a preset signal strength threshold. It can be set according to the service quality requirements of network operation and maintenance. For example, it can be set to -110dBm, that is, when the RSRP value of the user terminal is greater than -110dBm, it can be determined that the communication quality of the user terminal meets the preset reward conditions.

[0158] Specifically, the current reward result of the currently executed action can be calculated as follows: First, iterate through the pseudo RSRP values ​​of all M user terminals in the twin performance data and count the number of user terminals that meet the preset reward conditions; then, calculate the current reward result corresponding to the currently executed action using the reward function, the formula of which is as follows: In the formula, The current reward result corresponds to the currently executed action; M is the total number of user terminals in the network to be detected; This represents the number of user terminals that meet the preset reward conditions, i.e., the [number of terminals]. In the twin performance data, RSRP is greater than the signal strength threshold. The number of user terminals.

[0159] It is understandable that the reward result is positively correlated with the number of user terminals that meet the communication quality standards within the network. After the current parameter adjustment action is executed, the more users that meet the standards, the larger the reward result will be, which means that the action has a better recovery effect on the target fault. This guides the target deep Q network to iteratively seek optimization in the direction of the optimal global network performance.

[0160] By using the percentage of user terminals that meet the preset reward conditions as the current reward result, the global optimization goal of wireless network fault recovery can be transformed into a quantitative feedback signal that can be identified and iterated by the target depth Q network. This ensures that the parameter optimization process always focuses on improving the communication experience of all network users, avoiding the problem of only repairing faulty base stations while ignoring global interference, and significantly improving the global optimality of the fault recovery solution.

[0161] After determining the current reward result, the current state, current action, current reward result, and updated next state in this iteration are packaged into a single empirical data point and stored in a pre-defined replay memory. The current state is then updated to the next state; subsequently, the next state is used as the new current state, and the steps of determining the action, generating twin performance data, and calculating the reward result are repeated. Finally, when a pre-defined termination condition is met, the current state is used as the target parameter tuning strategy.

[0162] Among them, the preset termination condition can refer to the preset judgment rule used to end the parameter optimization iteration process, such as the number of iterations reaching the preset maximum number of iterations, or the current reward result not improving for several consecutive times (i.e., convergence), or the ideal configuration that makes all user terminals meet the standard has been found during the iteration process.

[0163] For example, the complete execution flow of this parameter optimization process can be shown below: Step 1: First, perform global initialization.

[0164] Specifically, the network weights can be initialized using a random normal distribution. For example, the deep neural network parameters of the target depth Q-network are first initialized. The counter for the initial training rounds Maximum number of training rounds Initialize the counter for the global cumulative iteration step size. Set the interval step size for parameter updates within a single round. The maximum time step in a single round of training Set the number of samples per session for experience replay. Simultaneously, the trainable parameters θ of the deep neural network in the target depth Q-network are initialized, and the network weights can be initialized using a random normal distribution.

[0165] Step 2: Perform the initialization operation for a single round of training.

[0166] Specifically, at the beginning of each training round, the time step counter for that round is initialized first. Set the network configuration parameters of the network under test to the initial state when the fault occurs. This represents the current state of this round of training. Simultaneously, based on the target twin network corresponding to the identified target fault type, the twin environment is initialized to ensure that the twin environment can accurately simulate the network performance change patterns under the current fault scenario.

[0167] Step 3: Execute single-step action decisions and environmental interaction operations, which may include the following steps: (1) Based on the formula Select the current state The current action with the largest Q value .

[0168] (2) Set the currently executing action The input is fed into the target Siamese network to simulate the network performance changes after the action is performed, and the current action is calculated using its Siamese performance data. Corresponding current reward result .

[0169] (3) Through the state update formula The next state of the network at the next time step is calculated. .

[0170] Step 4: Perform experience storage and counter update operations.

[0171] Specifically, the empirical data generated in this single-step iteration can be... Store it in the preset playback memory, and simultaneously update the single-round time step counter and the global cumulative iteration step size, that is, let , .

[0172] Step 5: Perform network parameter update judgment and operation.

[0173] First, determine if the network parameter update condition is met: when the global cumulative iteration step size is greater than the preset counting threshold (e.g., ...). This is used to ensure that a sufficient amount of valid empirical data has accumulated in the replay memory, avoiding parameter update bias caused by insufficient samples in the early stages of training. The parameter update operation is triggered when the parameter update condition is met (e.g., mod(step, 50) = 0, meaning a parameter update is performed every 50 iterations). The parameter update is then randomly selected from the replay memory. We take empirical data, substitute it back into the optimization function of the target depth Q network, and update the neural network parameters θ of the target depth Q network using the gradient descent algorithm.

[0174] Step 6: Perform a single-round inner loop judgment operation.

[0175] If the current time step has not reached the interval step size for network parameter updates in a single round (i.e.) , Then return to the initialization of the second round of training, until the time step within the single round reaches the interval step size (i.e., The loop iteration ends if the current time step reaches the interval step size for network parameter updates in a single round, but has not reached the maximum time step size (i.e., ...). If the time step reaches the maximum time step, then repeat the entire process from step three to step five until the time step in a single round reaches the maximum time step, thus completing the single round of iteration. Step 7: Execute the training termination judgment.

[0176] After a single round of the inner loop ends, determine the current training round. Is it the maximum number of training rounds? If it is not achieved, then... Return to step two and begin a new round of training iterations. If the target has been reached, stop training and set the final time step for that round. Corresponding current state Output as the optimal target parameter tuning strategy.

[0177] By iteratively updating the state, repeatedly executing actions, and determining rewards, the deep Q-network can continuously try and optimize its action selection strategy in a virtual environment, gradually approaching the optimal action sequence that maximizes the cumulative reward. Finally, when the preset termination condition is reached, the target parameter tuning strategy is output, realizing end-to-end decision-making from the fault state to the optimal recovery configuration, which significantly improves the intelligence level and processing efficiency of fault recovery.

[0178] In the above implementation, the initial state is determined based on the current transmit power and current antenna tilt angle of the network under test, providing a clear problem framework for iterative optimization. By defining the transmit power adjustment and antenna tilt angle adjustment to construct the action space, the continuous parameter adjustment is discretized into a finite set of actions, reducing the complexity of the decision-making problem and facilitating optimization exploration. Furthermore, parameter optimization processing is performed in this environment, enabling the system to learn the mapping relationship from the network state to the optimal adjustment action, achieving automated and intelligent fault recovery decision-making, avoiding the inefficiency and risk of manual trial and error, and improving the timeliness and accuracy of network fault handling.

[0179] It should be understood that although the steps in the flowchart above are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise explicitly stated in this document, there is no strict order requirement for the execution of these steps, and they can be executed in other orders.

[0180] Please see Figure 4 , Figure 4 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application, such as... Figure 4As shown, the computer device includes one or more processors 10, a memory 20, and interfaces for connecting the various components. The components are interconnected via different buses and can be mounted on a common motherboard or otherwise installed as needed. The processors can process instructions executed within the computer device, including instructions stored in or on the memory to display graphical information of a GUI on an external input / output device. The processor 10 can be a central processing unit, a network processor, or a combination thereof. The processor 10 may further include a hardware chip. The hardware chip can be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The programmable logic device can be a complex programmable logic device (CLP), a field-programmable gate array (FPGA), a general-purpose array logic (GPRS), or any combination thereof. The memory 20 stores instructions executable by at least one processor 10 to cause the at least one processor 10 to perform the methods shown in the above embodiments. The memory 20 may include a program storage area and a data storage area. The program storage area may store the operating system and application programs required for at least one function; the data storage area may store data created based on the use of the computer device, etc. In addition, memory 20 may include high-speed random access memory, and may also include non-transient memory, such as at least one disk storage device, flash memory device, or other non-transient solid-state storage device. Memory 20 may include volatile memory, such as random access memory; memory may also include non-volatile memory, such as flash memory, hard disk or solid-state drive combination. The computer device also includes input device 30 and output device 40. Processor 10, memory 20, input device 30 and output device 40 may be connected via a bus or other means. Input device 30 may receive input digital or character information, and generate key signal input related to user settings and function control of the computer device, such as touch screen, keypad, mouse, trackpad, touchpad, joystick, one or more mouse buttons, trackball, joystick, etc. Output device 40 may include display device, auxiliary lighting device (e.g., LED) and haptic feedback device (e.g., vibration motor), etc.

[0181] It is understood that before using the technical solutions disclosed in the various embodiments of this disclosure, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in this disclosure in an appropriate manner in accordance with relevant laws and regulations, and user authorization should be obtained.

[0182] The systems, modules, or units described in the above embodiments can be implemented by computer chips or physical entities, or by products with certain functions. A typical implementation device is a computer. For ease of description, the above devices are described separately by function as various units. Of course, in implementing this application, the functions of each unit can be implemented in one or more software and / or hardware.

[0183] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus, and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions.

[0184] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, such that the instructions, which execute on the computer or other programmable apparatus, provide steps for implementing the functions specified in one or more flowcharts and / or one or more block diagrams.

[0185] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0186] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the device embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0187] The above description is merely an embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of this application should be included within the scope of the claims of this application.

[0188] Although embodiments of this application have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of this application, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A network fault handling method, characterized in that, The method includes: Obtain the actual performance data of the network under test; Using a target detection model, fault detection is performed on the network to be detected based on the actual index data to obtain fault detection results; If the fault detection results indicate that the network under test is in the target fault type, the target twin network is used to perform parameter optimization on the network under test to obtain the target parameter tuning strategy; wherein, the target twin network corresponds to the target fault type. Based on the target parameter tuning strategy, a fault recovery operation is performed to complete the network fault handling.

2. The method according to claim 1, characterized in that, The target detection model is obtained using the following method: Determine the normal performance data and fault performance data of the sample network; The normal performance data and fault performance data are used to train a support vector machine to obtain a target detection model.

3. The method according to claim 2, characterized in that, The determination of normal performance data and fault performance data of the sample network includes: The normal performance data is determined based on the minimized drive test report of the wireless network under normal communication conditions; Based on the normal performance data and fault distance data, a pseudo-fault calculation is performed to obtain fault performance data; wherein, the fault distance data includes first distance data between the user terminal and the preset fault base station and at least one second distance data between the user terminal and each of the adjacent base stations of the preset fault base station.

4. The method according to claim 3, characterized in that, The step of performing pseudo-fault calculations based on the normal performance data and fault distance data to obtain fault performance data includes: The baseline fault data and fault adjustment amount are determined using the first distance data, the normal performance data, and the second distance data; wherein the fault adjustment amount is used to reflect the interference changes of the adjacent base stations; The fault performance data is determined based on the fault adjustment amount and the baseline fault data.

5. The method according to claim 2, characterized in that, The step of training a support vector machine based on the normal performance data and the fault performance data to obtain the target detection model includes: A digital twin network is constructed based on the normal performance data and the fault performance data; wherein, the digital twin network is a data generation network optimized through generative adversarial training; Data is generated using the digital twin network to obtain target training samples; wherein, the target training samples correspond to target state labels; the target state labels are used to reflect the network state corresponding to the target training samples; Based on the target training samples and the target state labels, the basic detection model is trained using a support vector machine to obtain the target detection model.

6. The method according to claim 1, characterized in that, When the fault detection result indicates that the network under test is in the target fault type, the target twin network is used to perform parameter optimization processing on the network under test to obtain the target parameter tuning strategy, including: The initial state and action space are determined based on the network configuration parameters of the network to be detected; wherein, the network configuration parameters include current configuration parameters and adjustable configuration parameters; the current configuration parameters include current transmit power and current antenna tilt angle; the adjustable configuration parameters include transmit power adjustment amount and antenna tilt angle adjustment amount; The target depth Q-network and the target Siamese network are used to perform parameter optimization in the action space to obtain the target parameter tuning strategy.

7. The method according to claim 6, characterized in that, The step of using the target depth Q-network and the target Siamese network to perform parameter optimization processing in the action space to obtain the target parameter tuning strategy includes: The initial state is taken as the current state; For the current state, the target depth Q-network is used to determine the current action to be performed; Using the target twin network, determine the current reward result of the currently executed action; Update the current state to the next state; Take the next state as the new current state and repeat the above steps for determining the action and reward result. If the preset termination condition is met, the current state will be used as the target parameter tuning strategy.

8. The method according to claim 7, characterized in that, Determining the current reward result of the currently executed action using the target Siamese network includes: The target twin network is used to generate twin performance data based on the currently executed action; wherein the twin performance data is used to reflect the network communication quality of the wireless network under the target fault type; Using the twin performance data and preset reward conditions, the current reward result of the currently executed action is determined; wherein, the current reward result refers to the number of user terminals that meet the preset reward conditions.

9. The method according to claim 1, characterized in that, The target fault types include antenna tilt angle too small, antenna tilt angle too large, and power too low.

10. A computer device, characterized in that, include: A memory and a processor, the memory and the processor being communicatively connected to each other, the memory storing computer instructions, the processor executing the computer instructions to perform the method of any one of claims 1 to 9.

Images