Network optimization method and device, and storage medium

Abstract

Disclosed in the present application are a network optimization method and device, and a storage medium. The method comprises the steps of: acquiring a user intent, and using an intent engine to process the user intent to obtain a first intent configuration, wherein the first intent configuration comprises an intent optimization strategy corresponding to a newly added intent type, and the intent optimization strategy is obtained by means of prediction based on initial network indicators; and optimizing the current network by means of the first intent configuration, so as to allow target network indicators of the current network to reach an intent objective corresponding to the user intent.

Description

Network optimization method, device and storage medium

[0001] Related applications

[0002] This application claims priority to Chinese patent application No. 202311728262.5 filed on December 14, 2023, the entire contents of which are incorporated by reference into this application. Technical Field

[0003] The present application relates to the technical field of wireless communication network security, and in particular to a network optimization method, device, and storage medium. Background Art

[0004] In recent years, with the application of intent-based networks in communication networks, when a communication network encounters an emergency, resulting in a sudden increase in specific business traffic and deterioration of terminal network indicators, specific intent targets can be issued to the intent-based network to optimize the network, greatly reducing the workload of operation and maintenance personnel.

[0005] Currently, the types of intent supported by intent-based networks are very limited. Different networks have local differences in the problems and solutions they face. Consequently, insufficient achievement of existing intent targets is a common problem. Support for these new issues requires a comprehensive process from intent collection to solution design and R&D delivery. As a result, the expansion of intent types cannot be separated from the productivity of the development team, slowing down the expansion of intent types and leading to inefficient communication network optimization.

[0006] Summary of the Invention

[0007] The main purpose of this application is to provide a network optimization method, device and storage medium.

[0008] To achieve the above-mentioned purpose, an embodiment of the present application provides a network optimization method, which includes: obtaining user intent, and using an intent engine to process the user intent to obtain a first intent configuration, wherein the first intent configuration includes an intent optimization strategy corresponding to the newly added intent type, and the intent optimization strategy is predicted based on the initial network indicators; optimizing the current network through the first intent configuration so that the target network indicators of the current network reach the intention target corresponding to the user intent.

[0009] The present application also provides a network optimization device, which is a physical node device. The network optimization device includes: a memory, a processor, and a program of the network optimization method stored in the memory and capable of running on the processor. When the program of the network optimization method is executed by the processor, the steps of the network optimization method described above can be implemented.

[0010] To achieve the above-mentioned purpose, a storage medium is further provided, on which a network optimization program is stored. When the network optimization program is executed by a processor, the steps of any of the above-mentioned network optimization methods are implemented. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] FIG1 is a flow chart of a first embodiment of the network optimization method of the present application;

[0012] FIG2 is a schematic diagram of the system architecture of the network optimization method of the present application;

[0013] FIG3 is a schematic diagram of the device structure of the hardware operating environment involved in the embodiment of the present application;

[0014] FIG4 is a schematic diagram of the execution flow involved in the network optimization method of the present application;

[0015] FIG5 is a schematic diagram of the device structure involved in the network optimization method of the present application;

[0016] FIG6 is a schematic diagram of a detailed flow chart of S10 in the first embodiment of the network optimization method of the present application;

[0017] FIG7 is a schematic diagram of a detailed flow chart of S11 in the first embodiment of the network optimization method of the present application;

[0018] FIG8 is a schematic diagram of a detailed flow chart of S13 in the first embodiment of the network optimization method of the present application;

[0019] FIG9 is a flow chart of a second embodiment of the network optimization method of the present application;

[0020] FIG10 is a flow chart of a third embodiment of the network optimization method of the present application;

[0021] FIG11 is a schematic diagram of a detailed flow chart of S110 in the third embodiment of the network optimization method of the present application;

[0022] FIG12 is a flow chart of a fourth embodiment of the network optimization method of the present application;

[0023] FIG13 is a schematic diagram of the model training process involved in the fourth embodiment of the network optimization method of the present application. DETAILED DESCRIPTION

[0024] The specific embodiments described herein are merely used to explain the present application and are not intended to limit the present application.

[0025] The present invention provides a network optimization method. FIG1 is a flow chart of a first embodiment of the network optimization method of the present invention. Referring to FIG1 , the method includes the following steps:

[0026] Step S10, obtain user intent, and use the intent engine to process the user intent to obtain a first intent configuration, wherein the first intent configuration includes an intent optimization strategy corresponding to the newly added intent type, and the intent optimization strategy is predicted based on initial network indicators.

[0027] The network optimization method can be applied to a network optimization device, which belongs to a network optimization system, and the network optimization system belongs to a network optimization device.

[0028] The executor of this method is a network optimization system. In this system, by receiving user intentions, identifying user intentions, collecting samples, and inputting network indicators related to user intentions into the model for processing, the required intention configuration is obtained, and then the current network is optimized with the obtained intention configuration. Among them, when it is determined that the user intention is a stored intention type, the intention configuration stored in the system can be directly applied to optimize the network. When it is determined that the user intention is a new intention type, the network indicators related to the user intention are predicted through the model, and combined with operation and maintenance experts and the simulation environment to quickly obtain the optimal intention configuration corresponding to the new intention type.

[0029] User intent may be a network indicator that a user hopes to achieve. For example, user intent may be: UE average delay < 100ms, UE average delay < 80ms, etc., which may be changed according to user needs and is not specifically limited.

[0030] The intent engine processes user intent and outputs intent configurations for optimizing the current network. It carries the general capabilities of intent, connects to the intent input of the business system, drives the intent model on the AI ​​platform to identify intent, generate configurations, perceive the real-time status of the network management system, and records and feeds back to operations and maintenance experts. Combined with the configuration experience of operations and maintenance experts, the intent implementation model is automatically optimized regularly.

[0031] The first intention configuration is an optimization strategy for optimizing the current network obtained after the intent engine processes the user intent. For example, when the user intention is that the UE average delay is <100ms, the generated first intention configuration can be "increase the single batch message processing volume by 10" or "reduce the data message processing interval by 10ms". The first intention configuration can be changed according to the current status of the indicators, and the first intention configuration includes multiple output optimization strategies. Among these optimization strategies, the optimal value of the solution is selected to optimize the current network.

[0032] The initial network indicators are the indicators to be adjusted associated with the newly added intent type. For example, when the user intention is UE average delay <100ms, the network indicators that need to be adjusted are "single batch message processing volume" and "data message processing interval". These two indicators are the initial network indicators, and the initial network indicators are predicted through the preset configuration model.

[0033] In the existing intent engine, intent configuration can only be generated for fixed user intent, while the first intent configuration involved in this application is the intent configuration output for the newly added intent type, wherein the newly added intent type is the new indicator requirement that needs to be achieved by the current network indicator. By predicting the user intent / newly added intent type through the preset configuration model, the intent optimization strategy is obtained, and the optimization strategy required for the new intent can be obtained more accurately. If the simulation result obtained after simulating the obtained optimization strategy does not achieve the required network goal, the optimization strategy is analyzed and adjusted, and the remaining relevant indicators are input into the model for processing until the required network optimization strategy is obtained.

[0034] Figure 2 is a schematic diagram of the system architecture of the network optimization method of this application. As shown in Figure 2, the system includes: an interface where the operation and maintenance expert is located, an interface where the intention user is located, an intention engine, a network management system, and an AI platform. The functions of each module in the system mainly include: Intent guarantee UI: the main interface for the intention user operation, which is mainly responsible for intention input, intention recognition confirmation, recommended configuration confirmation, intention list presentation, and intention achievement status viewing. Among them, the intention user is generally a low-threshold operator personnel without advanced operation and maintenance experience, who is responsible for issuing system operation and maintenance intentions in natural language to ensure specific business scenarios.

[0035] The intent engine includes modules such as expert UI, work order management, intent management, sample collection, intent model management, intent recognition, environmental perception, and configuration generation. The main functions of each module are as follows: Expert UI: An interface for operation and maintenance experts, where all intent tasks, configurations for intent modifications, the status of intent achievement, and related KPI indicators, alarms and other real-time environmental information can be viewed. The expert UI provides expert users with a configuration tuning entry for intent generation, and can provide suggestions for configuration modifications for intent implementation, as well as the addition and deletion of related indicator monitoring items. Among them, operation and maintenance experts are operation and maintenance personnel with advanced operation and maintenance knowledge, who are familiar with the functions of each module on the UME and are responsible for system status monitoring, alarm processing, network tuning, etc.

[0036] Work Order Management: Responsible for the closed-loop tracking of work orders for adding new intents and optimizing intent implementations. When the intent fulfillment rate is insufficient, a related alarm occurs, or the intent user enters an unsupported intent type, the intent engine automatically dispatches a work order to the operations and maintenance expert. The operations and maintenance expert will review the information of each module based on the work order information, troubleshoot the problem, and provide a configuration modification strategy. After the expert feedback on the work order is completed, the work order management system will push the configuration to the intent user. After the intent user confirms it, the intent configuration generation module will send it to the network management for configuration management. The work order system then continues to monitor the intent fulfillment rate until the intent is achieved. The modified configuration will be passed to the sample collection module and marked as a valid modification. If the intent is still not achieved after the modification is sent, the work order will continue to be dispatched to the expert for further tuning and passed to the sample collection module, marking the modification as invalid. The work order will be closed until the intent is achieved, or the user actively deletes or modifies the intent.

[0037] The work order module is embedded in the network management system for model optimization, allowing users to unconsciously feed back data to the model for model optimization. Unlike conventional work order modules, the work order process of this application will not only promote solution formulation and effect simulation, but also track the closed loop of solution execution effects, while promoting sample recording.

[0038] Sample Collection: This module is responsible for collecting automatically generated configurations, expert modification strategies, KPI indicators before and after configuration delivery, and the achievement of intent. Unlike existing sample collection methods (manual sample collection), this module automatically collects sample information in a standard format. By automatically collecting the configuration samples, indicator samples, and intent samples required for training, users can seamlessly provide sample data for model optimization.

[0039] Intent model management: This is responsible for managing the training and reasoning of intent-related models on the AI ​​platform. This mainly includes text intent recognition models, configuring pre-trained LMs, and configuring generative models. The functions of each model are used to:

[0040] 1. Configure pre-trained LM: Multiple configurations can be generated based on KPIs and partial configuration information.

[0041] 2. Preset configuration model: The model generated by pre-trained LM after supervised fine-tuning can automatically generate multiple configurations and recommended rankings based on KPIs and partial configuration information.

[0042] 3. Text intent recognition model / preset intent recognition model: Recognizes the user's natural language into several key elements of intent, including scope, KPI target, time, and operation type.

[0043] 4. Intent scenario clustering model / preset clustering model: Cluster intent optimization scenarios based on the differences before and after the KPI snapshot.

[0044] This module is different from the existing intent model management (which only includes intent recognition models). This module includes text recognition, configuration generation, and intent type recognition. In addition to driving model inference, it also drives model training. Through GPU, the online training speed is increased by more than 10 times, allowing the system to achieve day-granular model optimization. This allows the intent engine, which originally only included specific intent recognition capabilities, to become a general intent capability core with general intent recognition, implementation, and optimization.

[0045] Intent Recognition: This module receives natural language input from the user and sends it to the Intent Model Management module. This module then drives the AI ​​platform's model inference, identifies the intent type and key elements, and provides feedback to the user for confirmation. If the identified type isn't a supported type, a new intent ticket is automatically issued after user confirmation.

[0046] Environmental Perception: Responsible for collecting network status, providing configuration snapshots, alarm monitoring, KPI monitoring, and network element status monitoring. This module differs from existing intent systems (which directly feed back information from network elements to the intent assurance module for processing by the business module). This module triggers work orders, which are automatically processed by experts delegated by the intent engine without the need for business module intervention. It provides information for experts to use in decision-making and provides environmental data for the sample collection module.

[0047] Configuration generation: Responsible for identifying the network elements and cells affected by the intent, collecting environmental information, configuring inference, pushing configuration modifications, and distributing the configuration. Based on the intent recognition results, the environment perception module is queried for environmental information on the relevant network elements and cells. The KPI targets in the intent recognition results, the current configuration status of the relevant configurations provided by the environment perception module, and the current status of the KPI indicators are then sent to the AI ​​platform through intent model management for configuration generation. The generated configuration results are then recommended to the intent user for confirmation. After the user confirms, the configuration is automatically distributed to the network management system for configuration management.

[0048] This module differs from existing built-in rule-based configuration generation. Operations and maintenance experts do not need to directly formulate configuration data; they only need to enter related KPIs as input prompts for configuring LM. This module automatically feeds KPI targets, related KPI indicators, and existing configuration snapshots into the AI ​​model to generate multiple complete configurations, identify configuration modification points, and prioritize the three (or five, with no specific restrictions) configurations with the fewest modification points. After simulation in the digital twin environment, the module sorts and selects the configuration with the best KPI target achievement rate. The modification points are then pushed to the user, and only delivered to the live network after the user confirms them.

[0049] The functions of the network management are:

[0050] 1. Alarm management: responsible for network alarm management. When KPI is abnormal, alarms will be pushed to related systems.

[0051] 2. Configuration management: responsible for the management, distribution, and synchronization of all configuration information of network devices.

[0052] 3. Performance management: provide regular monitoring and storage of all KPI indicators of network devices.

[0053] 4. Network element management, responsible for network element information management, network element link establishment, model package management, etc.

[0054] The AI ​​platform’s functions are:

[0055] 1. Model training: Based on GPU, configure pre-trained LM, configure generation model, and train and optimize parameters of intent recognition model.

[0056] 2. Model inference: Based on the GPU, call the existing model and return the inference results based on the input.

[0057] 3. Digital twins use digital simulation technology to deploy network configurations in a simulation environment. Through channel simulation, geographical simulation, physical simulation, etc., 3D algorithms are used for real-time calculation to obtain the simulated impact of network configurations on the network.

[0058] 4. GPU driver, the underlying hardware driver module, sends algorithm training and reasoning to the GPU deployed on PAAS for parallel computing, accelerating the training and reasoning process.

[0059] The execution process of each module in the system is as follows:

[0060] 1. Intent user input intent: natural language intent input, mainly including scope, target KPI, time, and operation type.

[0061] 2. Intent type identification: Identify the type of intent. If it is an existing type, directly identify the five elements and after user confirmation and configuration generation, directly enter the effect evaluation stage. If it is not an existing type, the new intent process will be started after user confirmation.

[0062] 3. Dispatching work orders: If the environmental perception module finds that the intent is not achieved effectively, or an alarm occurs, or a new intent process is started, the work order management dispatches an intent work order, which includes the intent input and the trigger information provided by the environmental perception module, and the operation and maintenance expert accepts the order.

[0063] 4. View associated KPIs: Operations experts view intent-related KPIs and set these KPIs to intent-related KPIs. These KPIs are used to configure the input of the generation model and serve as the key KPIs for environmental awareness.

[0064] 5. Model generates multiple sets of recommended configurations: The KPIs added by experts are weighted and input into the configuration pre-training LM or configuration generation model together with the existing configuration snapshot and other KPIs. The model then generates multiple sets of recommended configurations.

[0065] 6. Simulation and optimization: Use digital twin technology to simulate the generated configurations, sort and select the best ones based on the degree of achievement. If none of them can be achieved, report to the experts for intervention, modify the recommended configuration as the new configuration, and feedback that the configuration modification strategy has been formulated.

[0066] 7. Configuration issuance: The configuration generation module pushes the configuration modification policy to the user for confirmation. After the user confirms, the configuration is issued and the policy sample is recorded, including the configuration and KPI snapshots.

[0067] 8. Intent fulfillment monitoring: The work order system continuously monitors the intent fulfillment rate. If the intent is not fulfilled, the work order is issued again and the policy sample is marked as an invalid modification sample. If the intent is fulfilled, it is marked as a valid modification sample and passed to the sample collection module.

[0068] 9. Work order closing: The work order will be closed until the intention is achieved.

[0069] Among them, Figure 6 is a detailed flow chart of S10 in the first embodiment. As shown in Figure 6, the step S10 of using the intention engine to process the user intention to obtain the first intention configuration includes: step S11, based on the intention engine, identifying the intention type corresponding to the user intention.

[0070] In the intent engine, a preset intent recognition model is used to identify the user's natural language into several key elements of intent, including scope, KPI target, time, and operation type. After obtaining these elements, the intent type corresponding to the user's intent can also be determined.

[0071] In step S12, the existing intention types are used to determine that the intention type corresponding to the user intention is a newly added intention type, and the initial network indicator associated with the user intention is input into a preset configuration generation component.

[0072] When it is determined that the existing intent type does not include the currently received user intent, and the intent type corresponding to the user intent is determined to be a new intent type, the initial network indicators associated with the user need to be input into the preset configuration generation component for processing to obtain the required intent configuration.

[0073] The initial network indicators include KPI indicators to be input into the preset configuration model. When operating the initial network indicators using the preset configuration model, it is also necessary to pay attention to the task configuration information required for creating the intent, such as network scope, environment information, and operation type.

[0074] The preset configuration generation component is a component that processes user intentions and outputs the required intention configuration. Among them, the preset configuration generation component mainly includes a preset clustering model, a preset configuration model, a preset intention recognition model, etc. The user intentions are clustered and predicted through each model to determine the intention configuration to be output.

[0075] Step S13: Determine a first intended configuration based on the preset configuration generation component and the initial network indicator.

[0076] Before inputting the initial network indicators into the preset configuration model, it is also necessary to determine the network indicator elements such as network range, time, operation type, etc. related to the network indicators. By inputting the initial network indicators into the preset configuration model for predictive processing, multiple sets of intention configurations and recommended rankings of these intention configurations can be automatically generated. Among them, the preset configuration model is a model based on deep reinforcement learning.

[0077] FIG. 7 is a schematic diagram of a detailed process of S11 in the first embodiment. As shown in FIG. 7 , step S11 of identifying the intent type corresponding to the user intent based on the intent engine includes:

[0078] Step S111 , clustering the user intention to obtain intention clustering data, and inputting the intention clustering data into a preset intention recognition model.

[0079] The user intention may be clustered by clustering the user intention with the existing intentions, and the user intention may be classified in advance to distinguish it from the existing intention types.

[0080] Intent clustering data can be data obtained after clustering user intents. Intent clustering data also includes newly added intent types, user intents of existing intent types, and outliers that do not belong to any cluster, thereby enhancing the differentiation of user intents.

[0081] Step S112: Based on the preset intention recognition model, the intention clustering data is identified to determine the intention type corresponding to the user intention.

[0082] The preset intent recognition model may be a text intent recognition model, which recognizes the natural language of the user's intent into several key elements of the intent, thereby determining the intent type to which the user's intent belongs.

[0083] The preset configuration generation component includes a preset configuration model; FIG8 is a schematic diagram of a detailed flow of S13 in the first embodiment. As shown in FIG8 , the step S13 of determining the first intended configuration based on the preset configuration generation component and the initial network indicator includes:

[0084] Step S131: input the initial network indicators into a preset configuration model.

[0085] Step S132: Based on the preset configuration model, the initial network indicators are predicted to obtain a first intended configuration.

[0086] When using the preset configuration model to predict and process the initial network indicators, it is also necessary to determine the intended optimization scenario, where the intended optimization scenario can be multiple network indicators that need to be optimized. For example, when the indicator to be optimized is network latency, the intended optimization scenario is scenario 1; when the indicator to be optimized is the number of data packets transmitted per unit time, the intended optimization scenario is scenario 2.

[0087] The prediction processing is the same as the prediction process of the input parameters by the deep learning-based model in the related technology. When the initial network indicators corresponding to the intention optimization scenario and the newly added intention type are determined, the preset configuration model predicts the initial network indicators corresponding to the intention optimization scenario and the newly added intention type to obtain the first intention configuration.

[0088] Step S20: Optimize the current network through the first intention configuration so that the target network indicator of the current network reaches the intention target corresponding to the user intention.

[0089] The target network indicator can be the network indicator achieved by the current network indicator. For example, the target network indicator can be a network delay less than 50ms. When the intended target is a network delay less than 50ms, the target network indicator is a network delay equal to 40ms. At this time, it can be determined that the target network indicator of the current network has achieved the intended target corresponding to the user intention.

[0090] Figure 4 is a schematic diagram of the execution process involved in this application. As shown in Figure 4, this application first parses the intention of the user input through semantic recognition to identify the intention type of the user's intention, where the user's intention also includes multiple intention target names and target values. The intention target is all KPIs supported by the system. The machine learning classification algorithm is used to identify the intent type and first determine whether it is an existing intent type. If the intention type input by the user is in the existing intent type, the configuration generation can be performed directly. If it is not in the currently supported existing intent type, the intent configuration corresponding to the new intent type is generated through the configuration generation model.

[0091] After determining that the user intent is a new intent, the user's input intent target and the current target value are sent to the intent engine. The intent engine dispatches a work order, and then through the corresponding network management operation, the indicator is checked according to the target. According to the KPI associated with the user intent, the initial network target is determined, and the initial network target is input into the previously trained configuration generation model for configuration generation. Then, through simulation selection, the optimal intent configuration is selected and transmitted to the client for confirmation. After receiving the issuance instruction from the user interface, it is finally issued to the existing network and feedback is fed back that the work order has been completed. Based on the achievement of the intent target after issuance, the sample data is marked as valid modification, and the invalid modification sample or the valid modification sample is used for model training.

[0092] Then, after collecting sample data, new samples are regularly summarized and the model is tuned. First, a clustering algorithm is used to cluster the differences before and after KPI optimization to identify whether the sample is a new category of sample data. Through outlier detection, samples that do not belong to any known aggregation are identified and these outliers are classified into a new category. Before the sample data is used for fine-tuning training, the proportion of high-quality samples is increased and the proportion of harmful samples is reduced through sample similarity calculation and fine-tuning data set optimization to obtain the optimized sample set. Finally, incremental supervised fine-tuning training is performed on the optimized sample set to perform incremental SFT (Supervised Fine-Tun-ing) to fine-tune the existing model weights so that it can generate more accurate recommendation results in the future, and the fine-tuned configuration generation model is used for the configuration generation of the next new intent type.

[0093] The next time the same type of intent is delivered, the intent engine will automatically obtain multiple sets of model recommendation values ​​and rankings, then evaluate and select the appropriate one to deliver directly, or modify and feed back the results to the intent engine. The intent engine records the changes and uses the output intent configuration as sample data to continue training.

[0094] The existing system has different requirements for network performance indicators when encountering different network conditions. When local residents have different network usage habits or encounter intent types that the system does not support, it is necessary to collect manual samples, identify intent, implement algorithm development, and then release versions to solve these differentiation problems.

[0095] The present application provides a network optimization method, device and storage medium. Compared with the related art, the use of intent network to optimize the communication network requires intention collection, solution design, and R&D delivery to support the newly added intent type, which slows down the speed of intent type expansion and leads to low efficiency of communication network optimization. In the present application, by obtaining the user intent and using the intent engine to process the user intent, a first intent configuration is obtained. Since the first intent configuration includes the intent optimization strategy corresponding to the newly added intent type, and the intent optimization strategy is based on the preset configuration model to predict the initial network indicators associated with the newly added intent type, the first intent configuration is more in line with the optimization strategy corresponding to the newly added intent type, and then, the corresponding intent configuration can be generated, so as to optimize the current network in a timely manner, thereby improving the network optimization efficiency.

[0096] Based on the first embodiment of the present application, another embodiment of the present application is provided. FIG9 is a flow chart of the second embodiment. As shown in FIG9 , in this embodiment, after the step S111 of clustering the user intent to obtain intent clustering data, the following steps are further included:

[0097] Step S1110 : sending the intent clustering data to the user terminal, and receiving newly added intent types obtained by the user terminal after naming the intent clustering data at every preset time period.

[0098] In addition to using the language big model for processing, the method of determining new intentions in this application can also be processed through the user side. The specific execution process is: after clustering the sample vector space, if the user sets it to automatic classification, there is no need for manual access, and the language big model can be used to automatically generate category names for it; if the user sets it to require user confirmation, the new categories generated in each clustering will be pushed to the user, and the user can manually name the new categories regularly.

[0099] In this embodiment, by sending the intent clustering data to the user end, the user is allowed to name the intent clustering data, so that the newly added intent type is more in line with user needs, and the processing on the user end can also make it easier for the model to process the input intent data.

[0100] Based on the first and second embodiments of the present application, another embodiment of the present application is provided. FIG10 is a flow chart of the third embodiment. As shown in FIG10 , in this embodiment, the first intent configuration further includes multiple parameter intent configurations; after the step S10 of using the intent engine to process the user intent to obtain the first intent configuration, the following steps are further included:

[0101] Step S110: Based on the first intention configuration, perform real-time calculations on the current network to obtain multiple network simulation results.

[0102] Parameter intention configuration can be a parameter optimization strategy of the network. For example, parameter intention configuration can be to adjust the A parameter so that the current network can achieve the intended target.

[0103] After obtaining the first intention configuration, the current network is deployed in a simulation environment, and the current network is calculated in real time using the first intention configuration, thereby obtaining a variety of network simulation results, wherein one network simulation result corresponds to a result obtained after implementation using one parameter intention configuration.

[0104] Step S120, sending the plurality of network simulation results to the user end, and upon receiving a confirmation instruction from the user end based on the feedback of the network simulation results, determining to send the parameter intention corresponding to the optimal value among the plurality of network simulation results to the current network.

[0105] After obtaining the network simulation results, they need to be sent to the user end for confirmation. When the user confirms that the network simulation results have achieved the expected effect, the user will send a confirmation instruction through the user end, thereby selecting the parameter intention corresponding to the optimal value of the network simulation result and configuring it to the current network, thereby optimizing the current network.

[0106] Step S130, when receiving the refusal instruction from the user terminal based on the feedback of the network simulation result, determine to feed back the first intention configuration to the operation and maintenance interface where the operation and maintenance expert is located, so that the operation and maintenance expert can modify the first intention configuration based on the network simulation result.

[0107] When the user end receives a refusal to issue an instruction based on the feedback of the network simulation results, it means that the network simulation results have not achieved the expected effect or the network indicators have not achieved the intended goals. At this time, the first intention configuration needs to be fed back to the operation and maintenance interface where the operation and maintenance expert is located, and after the operation and maintenance expert modifies / improves the first intention configuration, the adjusted intention configuration sent by the operation and maintenance expert is received.

[0108] FIG11 is a schematic diagram of a detailed process of S110 in the third embodiment. As shown in FIG11 , the step S110 of performing real-time calculations on the current network based on the first intention configuration to obtain multiple network simulation results includes:

[0109] Step S1101: determining the number of parameter modifications of each of the intended parameter configurations, wherein the intended parameter configurations are arranged in order from least to greatest based on the number of parameter modifications.

[0110] The number of parameter modifications corresponds to the number of parameter pairs that need to be modified for each parameter intention configuration. The number of parameter modifications can be 1, 2, 3, etc., and there is no specific limit. The fewer parameters that need to be modified for the parameter intention configuration, the higher the ranking.

[0111] Step S1102: Based on the number of parameter modifications, select a preset number of parameter intention configurations in the first intention configuration before sorting.

[0112] In the process of obtaining the network simulation results, the parameter intention configuration with the highest first intention configuration is selected, and the calculation is performed according to the recommended ranking of the parameter intention configuration. The parameter intention configuration with the lowest ranking is considered to have a poor calculation effect and is not considered. The preset number can be 3, 5, 7, etc., and there is no specific limitation.

[0113] Step S1103 , performing real-time calculations on the current network one by one with the intended parameter configuration to obtain a plurality of network simulation results.

[0114] The current network is calculated in real time according to the order of parameter intention configuration to obtain various network simulation results.

[0115] In this embodiment, before the first intention configuration is sent to the current network and put into use, the first intention configuration is calculated in real time on the network arranged in the simulation environment, so as to select the optimal intention configuration output by the preset configuration model according to the network simulation results.

[0116] Based on the first, second, and third embodiments of the present application, another embodiment of the present application is provided. FIG12 is a flow chart of the fourth embodiment. As shown in FIG12 , in this embodiment, after step S20 of optimizing the current network through the first intent configuration so that the target network indicator of the current network reaches the intent target corresponding to the user intent, the following steps are further included:

[0117] Step S30: If the target network indicator does not reach the intention target corresponding to the user intention, the first intention configuration is marked as an invalid modification sample.

[0118] When the first intent configuration is applied to the current network, if the target network indicator does not reach the intent target corresponding to the user intent, it means that the currently generated first intent configuration is invalid and is marked as an invalid modification sample.

[0119] Step S40: If the target network indicator reaches the intention target corresponding to the user intention, the first intention configuration is marked as a valid modification sample.

[0120] When the first intent configuration is applied to the current network, and the target network indicator reaches the intention target corresponding to the user intent, it means that the currently generated first intent configuration is valid and is marked as a valid modification sample.

[0121] Step S50: input the invalid modification sample or the valid modification sample into a preset configuration generation component for iterative training of the preset configuration model.

[0122] After determining invalid modification samples or valid modification samples, these two types of samples are input into the preset configuration generation component to fine-tune and iteratively train the preset clustering model and the preset configuration model.

[0123] FIG13 is a schematic diagram of the model training process. As shown in FIG13 , after step S50 of inputting the invalid modified sample or the valid modified sample into the preset configuration generation component, the process further includes:

[0124] Step S510: taking the invalid modified sample or the valid modified sample as newly added sample data, clustering the newly added sample data and the existing sample data to obtain clustered data;

[0125] Invalid modified samples or valid modified samples are used as sample data, and the sample data is used as the input of the model. At the same time, the newly added sample data and the existing sample data are clustered. After the cluster data is determined, the cluster data is used to fine-tune the preset configuration model to obtain a fine-tuned clustering model.

[0126] The methods for pre-training and fine-tuning the preset configuration model can be: 1. Obtain a large amount of configuration and KPI snapshot information: Obtain a large amount of exported XML configuration information and current KPI value snapshots from the field or laboratory for unsupervised training, where the KPI can be used as an XML node and strung to the end of the configuration XML; 2. Generate a configuration pre-training LM: Based on the GPU, perform unsupervised training on the configuration and KPI snapshots after random masking; 3. Sample data contains the masked configuration matrix and KPI matrix as the neural network input. The middle layer can use Transformer or LSTM, and the output is the configuration matrix; 4 , allowing the model to generate complete configuration parameters through the input of partial configuration and KPI information; 5. Regularly summarize new samples: regularly collect samples from the sample collection module for fine-tuning training; 6. Sample aggregation: extract the differences between the new samples and existing samples before and after the KPI snapshots, and cluster them; 7. Identify new categories: through outlier detection, identify samples that do not belong to any known aggregation, and classify these outliers into a new category. If the system is set to automatic addition mode, use LLM (large language model, used to generate and understand natural language) to summarize and name them, otherwise push them to experts for naming or adjustment of classification on a regular basis.

[0127] Step S520 , performing vectorization conversion on the cluster data and the old sample data to obtain cluster sample vectors and old sample vectors.

[0128] After the cluster data and old sample data are vectorized using doc2vec (an algorithm that converts document samples into fixed-length vectors), cluster sample vectors and old sample vectors are obtained. The purpose of vectorization is to calculate the cosine distance between each vector and thus calculate the similarity between the two data.

[0129] Step S530 : Calculate the similarity between the clustered sample vector and the old sample vector, and determine an optimized sample data set based on the similarity.

[0130] The method for determining the optimized sample dataset is to traverse the category of each newly added sample, delete the sample with the closest cosine distance to the invalid sample, and directly add the valid samples to the fine-tuning dataset to increase the proportion of high-quality samples and reduce the proportion of harmful samples.

[0131] Step S540 : fine-tuning the preset configuration model using the optimized sample data set to obtain a fine-tuned preset configuration model.

[0132] Incremental SFT (Supervised Fine-Tuning) is performed on the optimized sample data set, which uses labeled sample data to train and adjust model parameters, and backs up and synchronizes the changed sample set. This is used to fine-tune the existing preset configuration model weights so that it can subsequently generate more accurate recommendation results.

[0133] The model training frequency can be once a day or once half a day, with no specific limit.

[0134] In this embodiment, the intent configuration output by the preset configuration model is sampled and marked as invalid modification samples and valid modification samples respectively. Then, the preset configuration model is iteratively trained with these sample data, so that the model can output the optimal intent configuration that is more in line with the user's intention.

[0135] The scenarios used in this application are as follows:

[0136] 1. Differences in site network requirements: For example, a concert network focuses more on improving user capacity, while a conference room network focuses more on user communication quality.

[0137] 2. Differences in user usage habits: Users in region A prefer to visit text-based websites, while users in region B prefer to watch short videos. The two have different requirements for network bandwidth.

[0138] 3. Unsupported intent type: The user used a KPI target created by the site, which is not within the scope of the system's built-in intent types.

[0139] If these scenarios are encountered, the marketing team needs to submit requirements to the R&D center, and then the planning team will determine the site specificity and intent new version release strategy. The requirements will then drive the R&D team to collect intent sample data, intent type definition, intent recognition algorithm development, intent implementation algorithm development, test delivery, version release, field verification, etc. from the bureau. The entire process involves the entire process and all roles of system delivery, with a general cycle of two to three months, which consumes huge manpower. In addition, many intent types may be applicable to fewer sites, and the R&D cost is not proportional to the frequency of use.

[0140] In scenario 1, the original solution might be to use a differentiated version strategy, resulting in a very complex version strategy, or to provide a complete set of requirements for all sites for the site to choose from, resulting in very complex functions that are difficult to maintain.

[0141] In scenario 2, it may be necessary to collect user habits in different regions and then develop different intent realization strategies based on different user habits. Regional differences among users are very common and their number is huge, so it is not feasible to rely entirely on R&D to collect and realize them.

[0142] In scenario 3, only the new demand R&D process can be followed, which may result in the R&D process not keeping up with the addition of new intention requirements.

[0143] Through the above intent addition process and model training process, when the above scenarios are discovered, the field intent system automatic optimization process can be automatically triggered. Through system work order operation, intervention of field operation and maintenance experts, simulation, sample collection, and model training, the entire intent addition and optimization process can be completed, reducing the R&D cycle of several months to one or two weeks, and the optimization strategy is more suitable for local network conditions and user habits.

[0144] Examples of application scenarios involved in this application are as follows:

[0145] Application example 1:

[0146] The user wants to reduce the average UE latency at the Shanghai New International Expo Center. Therefore, he tells the system: "Shanghai New International Expo Center, average UE latency < 100ms."

[0147] The intent recognition module identifies the KPI name as "UE Average Latency," with a target value of less than 100ms, and the scope is "Shanghai New International Expo Center." Based on the current KPI value of 110ms and the configured value, the module classifies the intent and finds that the intent does not match existing intent types. The system prompts the user, "This intent type is not currently supported. Do you want to enable the new intent type mode?"

[0148] If the user selects "Yes", the system automatically dispatches the work order to the operation and maintenance expert. At the same time, the intention is issued and the status is displayed as dispatched work order.

[0149] Based on the user's description and experience, the expert reviews the associated KPIs and marks them as KPIs of key focus for this type of intent. Based on the expert's markings, the system automatically generates multiple sets of configuration parameters using a large model. The first set recommends "increasing the single-batch message processing volume by 10," and the second set recommends "reducing the data message processing interval by 10 ms," and presents them to the expert.

[0150] Through simulation, it was found that the second set of configurations had better simulation effects, so the "data message processing interval is reduced by 10ms" in the configuration MOC was configured and feedback was given to complete the configuration strategy formulation.

[0151] The second set of recommended configurations is delivered to the network elements within the intended specified range.

[0152] After 5 minutes, the intention is achieved, the work order is automatically closed, and the recommended configuration is marked as a valid modification. The snapshots before and after the configuration modification and the snapshots before and after the KPI modification are recorded as valid modification samples.

[0153] That evening, the system automatically aggregated samples based on all the samples collected that day and found that "Shanghai New International Expo Center, UE average delay <100ms" was a new type. It was then pushed to the operation and maintenance experts for naming. After confirmation by the operation and maintenance experts, it was named "UE average delay optimization."

[0154] The system implements the results according to the intention, adds the second set of recommended configurations to the valid sample pool, performs fine-tuning training, and saves the fine-tuned configuration generation model.

[0155] The next day, the user input "Zhangjiang Hi-Tech Subway Station, UE average latency <80ms". The system classified the intent as "UE average latency optimization" based on clustering and determined the current KPI value to be 110ms. It then generated a set of parameters that was closer to the original second configuration plan. The "Data Message Processing Interval" was reduced by 30ms and was issued after user confirmation.

[0156] After 5 minutes, the intention is achieved and the system automatically marks the configuration as a valid modification and records the sample.

[0157] Application Example 2:

[0158] In step (11) of Application Example 1, the intention achievement rate is still poor after 5 minutes.

[0159] The system automatically marked the configuration as an invalid modification sample and issued a work order, which was then received by an expert. After investigation, it was found that the high latency within the network elements around Zhangjiang Hi-Tech Park Subway Station was not caused by the "data packet processing interval", but by the KPI "downlink utilization" being too high. Therefore, the weight of the associated KPI "downlink utilization" was increased for this intent type, and it was input into the model for reconfiguration. In the newly generated configuration, the third configuration plan was to increase the "cell bandwidth" within the network element.

[0160] After simulation results, the third solution was selected and feedback was given to complete the formulation of the modified strategy.

[0161] After confirmation by the user, the configuration is delivered.

[0162] After 5 minutes, the intention was achieved and the sample was recorded as validly modified.

[0163] Application Example 3:

[0164] In step (5) of Application Example 2, the expert felt that none of the solutions met the requirements, so he manually created a new solution, "Increase the number of message caches by 100", and reported that the modified strategy was completed.

[0165] After user confirmation, the configuration is delivered; after 5 minutes, the intention is achieved and recorded as a valid modification sample.

[0166] Application Example 4:

[0167] In the second application scenario, during clustering that evening, the system discovered that "Zhangjiang Hi-Tech Subway Station, UE average latency < 80ms." Combined with the current KPI snapshot, the system determined that the clustering intent, "UE average latency optimization," was inappropriate and should instead be assigned to the "bandwidth utilization optimization" intent.

[0168] Experts modify sample attribution and optimize clustering models.

[0169] The next time downlink utilization is insufficient, it will be more likely to be identified as "bandwidth utilization optimization" and pay attention to the KPI associated with this type.

[0170] Refer to Figure 3, which is a schematic diagram of the device structure of the hardware operating environment involved in the embodiment of the present application.

[0171] As shown in FIG3 , the network optimization device may include: a processor 1001 , a memory 1005 , and a communication bus 1002 . The communication bus 1002 is used to implement connection and communication between the processor 1001 and the memory 1005 .

[0172] In one embodiment, the network optimization device may further include a user interface, a network interface, a camera, an RF (Radio Frequency) circuit, a sensor, a WiFi module, and the like. The user interface may include a display screen and an input submodule such as a keyboard. The user interface may also include a standard wired interface and a wireless interface. The network interface may include a standard wired interface and a wireless interface (such as a WiFi interface).

[0173] Those skilled in the art will understand that the network optimization device structure shown in FIG3 does not constitute a limitation on the network optimization device, and may include more or fewer components than shown in the figure, or a combination of certain components, or a different arrangement of components.

[0174] As shown in Figure 3, memory 1005, a storage medium, may include an operating system, a network communication module, and a network optimization program. The operating system is a program that manages and controls the hardware and software resources of the network optimization device and supports the operation of the network optimization program and other software and / or programs. The network communication module is used to enable communication between the various components within memory 1005, as well as communication with other hardware and software in the network optimization system.

[0175] In the network optimization device shown in FIG3 , the processor 1001 is configured to execute the network optimization program stored in the memory 1005 to implement the steps of any one of the above-mentioned network optimization methods.

[0176] The specific implementation of the network optimization device of the present application is basically the same as the embodiments of the above-mentioned network optimization method, and will not be repeated here.

[0177] The present application also provides a network optimization device. Referring to Figure 5, the network optimization device includes: an acquisition module, used to obtain user intent, and use an intent engine to process the user intent to obtain a first intent configuration, wherein the first intent configuration includes an intent optimization strategy corresponding to the newly added intent type, and the intent optimization strategy is predicted based on the initial network indicators; an optimization module, used to optimize the current network through the first intent configuration so that the target network indicators of the current network reach the intention target corresponding to the user intent.

[0178] In a possible implementation of the present application, the acquisition module includes: an identification unit, used to identify the intent type corresponding to the user intent based on the intent engine; a first determination unit, used to use the existing intent type to determine that the intent type corresponding to the user intent is a new intent type, and then input the initial network indicator associated with the user intention into a preset configuration generation component; a processing unit, used to determine the first intention configuration based on the preset configuration generation component and the initial network indicator.

[0179] In a possible implementation of the present application, the processing unit includes: an input subunit, used to input the initial network indicators into a preset configuration model; and a processing subunit, used to perform predictive processing on the initial network indicators based on the preset configuration model to obtain a first intended configuration.

[0180] In a possible implementation of the present application, the device also includes: a calculation module, which is used to perform real-time calculations on the current network based on the first intention configuration to obtain multiple network simulation results; a sending module, which is used to send the multiple network simulation results to the user end, and upon receiving a confirmation sending instruction from the user end based on the feedback of the network simulation results, determine to send the parameter intention configuration corresponding to the optimal value of the multiple network simulation results to the current network; a receiving module, which is used to determine to feed back the first intention configuration to the operation and maintenance interface where the operation and maintenance expert is located, upon receiving a refusal to send instruction from the user end based on the feedback of the network simulation results, so that the operation and maintenance expert can modify the first intention configuration based on the network simulation results.

[0181] In a possible implementation of the present application, the calculation module includes: a second determination unit, used to determine the number of parameter modifications of each of the parameter intention configurations, wherein the parameter intention configurations are arranged in sequence from small to large based on the number of parameter modifications; a selection unit, used to select a preset number of parameter intention configurations in the first intention configuration before sorting based on the number of parameter modifications; and a real-time calculation unit, used to perform real-time calculation on the current network one by one with the parameter intention configuration to obtain multiple network simulation results.

[0182] In a possible embodiment of the present application, the identification unit includes: a third processing sub-unit, used to cluster the user intention, obtain intention clustering data, and input the intention clustering data into a preset intention recognition model; an identification sub-unit, used to identify the intention clustering data based on the preset intention recognition model, and determine the intention type corresponding to the user intention.

[0183] In a possible embodiment of the present application, the device also includes: a first marking module, which is used to mark the first intention configuration as an invalid modification sample if the target network indicator does not reach the intention target corresponding to the user intention; a second marking module, which is used to mark the first intention configuration as a valid modification sample if the target network indicator reaches the intention target corresponding to the user intention; and an input module, which is used to input the invalid modification sample or the valid modification sample into a preset configuration generation component for iterative training of the preset configuration model.

[0184] In a possible embodiment of the present application, the device also includes: a training module, which is used to take the invalid modified sample or the valid modified sample as the new sample data, cluster the new sample data with the existing sample data, and obtain cluster data; a conversion module, which is used to vectorize the cluster data and the old sample data to obtain a cluster sample vector and an old sample vector; a determination module, which is used to calculate the similarity between the cluster sample vector and the old sample vector, and determine the optimized sample data set based on the similarity; a fine-tuning module, which is used to fine-tune the preset configuration model through the optimized sample data set to obtain a fine-tuned preset configuration model.

[0185] As used herein, the terms "comprises," "comprising," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or system that includes a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or system. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not preclude the presence of additional identical elements in the process, method, article, or system that includes the element.

[0186] The serial numbers of the above embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0187] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus the necessary general hardware platform, and of course can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes a number of instructions for enabling a terminal device (which can be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in each embodiment of the present application.

[0188] The above are merely optional embodiments of the present application and do not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A network optimization method, wherein: The method comprises: Obtaining a user intent, and processing the user intent using an intent engine to obtain a first intent configuration, wherein the first intent configuration includes an intent optimization strategy corresponding to a newly added intent type, and the intent optimization strategy is predicted based on initial network indicators; The current network is optimized through the first intention configuration so that the target network indicator of the current network reaches the intention target corresponding to the user intention.

2. The network optimization method according to claim 1, wherein: The step of using the intent engine to process the user intent to obtain a first intent configuration includes: Based on the intent engine, identifying the intent type corresponding to the user intent; Determine using the existing intent types that the intent type corresponding to the user intent is a newly added intent type, and then input the initial network indicator associated with the user intent into a preset configuration generation component; Based on the preset configuration generation component and the initial network indicator, a first intended configuration is determined.

3. The network optimization method according to claim 2, wherein: The preset configuration generation component includes a preset configuration model; the step of determining the first intended configuration based on the preset configuration generation component and the initial network indicator includes: Inputting the initial network indicators into a preset configuration model; Based on the preset configuration model, the initial network indicators are predicted and processed to obtain a first intended configuration.

4. The network optimization method according to claim 1, wherein: The first intent configuration includes multiple parameter intent configurations; after the step of using the intent engine to process the user intent to obtain the first intent configuration, the step further includes: Based on the first intention configuration, the current network is calculated in real time to obtain multiple network simulation results; Sending the plurality of network simulation results to a user terminal, and upon receiving a confirmation instruction from the user terminal based on the feedback of the network simulation results, determining to send a parameter configuration corresponding to an optimal value among the plurality of network simulation results to the current network; When receiving a refusal to issue instruction from the user terminal based on the feedback of the network simulation result, it is determined that the first intention configuration is fed back to the operation and maintenance interface where the operation and maintenance expert is located, so that the operation and maintenance expert can modify the first intention configuration based on the network simulation result.

5. The network optimization method according to claim 4, wherein: The step of performing real-time calculation on the current network based on the first intention configuration to obtain multiple network simulation results includes: Determining the number of parameter modifications of each of the intended parameter configurations, wherein the intended parameter configurations are arranged in order from small to large based on the number of parameter modifications; Based on the number of parameter modifications, selecting a preset number of parameter intention configurations before sorting in the first intention configuration; The current networks are calculated one by one in real time according to the parameter intention configuration to obtain multiple network simulation results.

6. The network optimization method according to claim 2, wherein: The step of identifying the intent type corresponding to the user intent based on the intent engine includes: Clustering the user intentions to obtain intention clustering data, and inputting the intention clustering data into a preset intention recognition model; Based on the preset intention recognition model, the intention clustering data is identified to determine the intention type corresponding to the user intention; or, After the step of clustering the user intentions to obtain the intention clustering data, the method further includes: The intention clustering data is sent to a user terminal, and newly added intention types are received after the user terminal names the intention clustering data every preset time period.

7. The network optimization method according to claim 1, wherein: After the step of optimizing the current network by the first intention configuration so that the target network indicator of the current network reaches the intention target corresponding to the user intention, the method further includes: If the target network indicator does not reach the intention target corresponding to the user intention, marking the first intention configuration as an invalid modification sample; If the target network indicator reaches the intention target corresponding to the user intention, marking the first intention configuration as a valid modification sample; The invalid modification sample or the valid modification sample is input into a preset configuration generation component for iterative training of the preset configuration model.

8. The network optimization method according to claim 7, wherein: After the step of inputting the invalid modification sample or the valid modification sample into the preset configuration generation component, the method further includes: Taking the invalid modified sample or the valid modified sample as newly added sample data, clustering the newly added sample data and the existing sample data to obtain clustered data; Performing vectorization conversion on the cluster data and the old sample data to obtain a cluster sample vector and an old sample vector; Calculating the similarity between the clustered sample vector and the old sample vector, and determining an optimized sample data set based on the similarity; The preset configuration model is fine-tuned using the optimized sample data set to obtain a fine-tuned preset configuration model.

9. A network optimization device, wherein: The device comprises: a memory, a processor, and a network optimization program stored in the memory and executable on the processor, wherein the network optimization program is configured to implement the steps of the network optimization method according to any one of claims 1 to 8.

10. A computer storage medium, wherein: The computer storage medium stores a network optimization program, and when the network optimization program is executed by the processor, the steps of the network optimization method according to any one of claims 1 to 8 are implemented.

Images