Network optimization work order processing method and device, equipment and storage medium

Abstract

The disclosure provides a network optimization work order processing method and device, relating to the technical field of network optimization, which comprises the following steps: extracting the prompt of the network optimization work order to obtain network problem description information and network optimization object information; inputting the network problem description information and network optimization object information into a pre-trained optimization model to determine a network optimization intention, wherein the optimization model is a large language model; querying a network optimization platform for network performance indicators of the network optimization object network based on the network optimization intention; and performing perception analysis on the network performance indicators based on the optimization model to obtain optimization suggestion information for the network problem cause. Thus, a large language model in the field of wireless network optimization can be trained, and based on the powerful semantic understanding and intention translation capability of the large model, the network optimization intention can be accurately identified, and the network indicator query method and network analysis optimization strategy can be quickly recommended to the user.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of network optimization, and particularly relates to a network optimization work order processing method and device, equipment and a storage medium. BACKGROUND

[0002] Large language models (LLMs) have opened a new era of AI, and have greatly improved their ability to perform complex multi-step reasoning. Among various new network forms, intent-driven networks (IDNs) or intent-based networks (IBNs) are key technologies for network autonomy. Although large language models have rich general knowledge, they lack wireless network optimization-specific training data and have insufficient knowledge of wireless network optimization, making it difficult for them to understand network optimization intent and meet the needs of wireless network optimization professional scenarios. SUMMARY

[0003] The present disclosure aims to at least partially address one of the technical problems in the related art.

[0004] A first aspect of the present disclosure provides a network optimization work order processing method, comprising:

[0005] extracting prompt language from the network optimization work order to obtain network problem description information and network optimization object information;

[0006] inputting the network problem description information and the network optimization object information into a pre-trained optimization model to determine a network optimization intent, the optimization model being a large language model;

[0007] querying network performance indicators for the network optimization object network based on the network optimization intent;

[0008] performing perceptual analysis on the network performance indicators based on the optimization model to obtain optimization suggestion information for the network problem cause.

[0009] A second aspect of the present disclosure provides a network optimization work order processing device, comprising:

[0010] an obtaining module configured to extract prompt language from the network optimization work order to obtain network problem description information and network optimization object information;

[0011] a determining module configured to input the network problem description information and the network optimization object information into a pre-trained optimization model to determine a network optimization intent, the optimization model being a large language model;

[0012] query a network optimization platform for network performance indicators of the network optimization object network based on the network optimization intention;

[0013] analyze the network performance indicators based on the optimization model to obtain optimization suggestion information for a network problem cause.

[0014] A third aspect of the present disclosure provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, and when the processor executes the program, the processing method of the network optimization work order is implemented as described in the first aspect of the present disclosure.

[0015] A fourth aspect of the present disclosure provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the processing method of the network optimization work order is implemented as described in the first aspect of the present disclosure.

[0016] A fifth aspect of the present disclosure provides a computer program product, and when the instructions in the computer program product are executed by a processor, the processing method of the network optimization work order is implemented as described in the first aspect of the present disclosure.

[0017] In the embodiments of the present disclosure, first, the network optimization work order is prompted to extract the network problem description information and the network optimization object information, and then the network problem description information and the network optimization object information are input into the pre-trained optimization model to determine the network optimization intention. The optimization model is a large language model. Then, based on the network optimization intention, the network optimization platform is queried for network performance indicators of the network optimization object network. Then, based on the optimization model, the network performance indicators are analyzed to obtain optimization suggestion information for a network problem cause. In this way, by training a large language model in the field of wireless network optimization, based on the powerful semantic understanding and intention translation ability of the large model, the network optimization intention can be accurately identified. By using the understanding of the network optimization intention by the large language model and the understanding of expert optimization experience cases, when similar network faults occur, the user network indicator query method and network analysis optimization strategy can be quickly recommended.

[0018] Additional aspects and advantages of the present disclosure will be in part apparent and in part pointed out hereinafter. BRIEF DESCRIPTION OF DRAWINGS

[0019] The above and / or additional aspects and advantages of the present disclosure will become apparent and be readily appreciated from the following description of embodiments, taken in conjunction with the accompanying drawings, in which:

[0020] Figure 1A flowchart of a network optimization work order processing method provided by an embodiment of the present disclosure is shown in FIG. 1.

[0021] Figure 2 A flowchart of a wireless network optimization work order automatic processing method based on a large model is shown in FIG. 2.

[0022] Figure 3 A flowchart of another network optimization work order processing method provided by an embodiment of the present disclosure is shown in FIG. 3.

[0023] Figure 4 A flowchart of a wireless network optimization large model training and use method provided by an embodiment of the present disclosure is shown in FIG. 4.

[0024] Figure 5 A structural diagram of a network optimization work order processing device provided by an embodiment of the present disclosure is shown in FIG. 5.

[0025] Figure 6 A block diagram of an electronic device for network optimization work order processing according to an exemplary embodiment is shown in FIG. 6. DETAILED DESCRIPTION

[0026] Embodiments of the present disclosure are described in detail below with reference to the accompanying drawings, examples of which are shown in the drawings, in which the same or similar notations are used to denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by reference to the accompanying drawings are exemplary and are intended to explain the present disclosure, and cannot be understood as limiting the present disclosure.

[0027] In the related art, for the optimization of daily network problems of a wireless network, there is a set of wireless network optimization management platforms, which assign network optimization experts in the form of work orders through the platform. The network optimization experts analyze and formulate a solution on the basis of the network problem described in the work order, combined with their years of network optimization experience and the use of various professional auxiliary tools for data collection, data analysis, network problem re-identification, analysis and formulation of a solution. The wireless network optimization management platform records the solution and continues to circulate the work order to network front-line operators or an automatic execution system for execution of the solution until the problem is solved and the work order is closed. However, due to the large size of the network and the complex causes of network problems, and the non-standard description of wireless network work order problems, the data of the entire service process is disorganized, it is difficult to extract effective data information, and the network optimization intention cannot be accurately understood.

[0028] In addition, since the rules are set based on artificial experience to process centralized optimization work orders, the application convenience is greatly reduced, the rules cannot be updated and the judgment logic cannot be adjusted in time to adapt to changes in network conditions, and the method is not suitable for scenarios where abnormal judgment standards are not clear, and artificial rules or all standard processes cannot be defined in advance.

[0029] Embodiments of the present disclosure provide a network optimization work order processing method, device, equipment and storage medium, which are used to solve at least one of the above problems.

[0030] In the embodiments of the present disclosure, the network optimization work order processing method can be executed by a network optimization work order processing device, which is not limited herein.

[0031] To solve the above problems, the present disclosure provides a network optimization work order processing method and device.

[0032] The network optimization work order processing method and device of the embodiments of the present disclosure will be described below in combination with the accompanying drawings.

[0033] Figure 1 A flowchart of a network optimization work order processing method provided by the embodiments of the present disclosure is shown in FIG. 1.

[0034] As shown in FIG. 2, the network optimization work order processing method can include the following steps: Figure 1

[0035] Step 101: Extracting prompt words from the network optimization work order to obtain network problem description information and network optimization object information.

[0036] As an example, in the network problem processing scenario, the network optimization work order at least contains the following fields: work order ID, work order date, city, object type, object number, object name, problem description, network type, home area, coverage type, coverage scenario, work frequency band and equipment manufacturer.

[0037] ​Among them, in the network optimization work order, the above-mentioned fields contain rich information, which can be used to describe and locate the network problem to be optimized in detail. The specific explanations are as follows: 1. Work order ID: a unique number that identifies a network optimization task, facilitating management and tracking. 2. Work order date: records the time point when the network optimization task is created or submitted, facilitating work order processing and historical record retrieval according to time sequence. 3. City: indicates the geographical area where the network problem occurs, which helps to analyze regional network conditions and resource allocation. 4. Object type: indicates the object category that needs to be optimized, such as base station, switch, router, server, etc. 5. Object number: a unique identifier for a specific network object, facilitating accurate search and positioning. 6. Object name: the specific name or description of the network object, which helps to further understand the network structure. 7. Problem description: details of the problems or fault phenomena in the network, which is the focus of prompt extraction to find out the problem core and take appropriate optimization measures. 8. Network type: such as 4G, 5G, Wi-Fi, wired network, etc. Different types of networks may face different optimization challenges. 9. Belonging area: more detailed geographical location information, such as cell, park, business district, etc., which helps to determine the specific implementation range of optimization work. 10. Coverage type: such as macro station, micro station, indoor distribution system, etc. Different coverage types may require different optimization strategies. 11. Coverage scenario: such as residential area, commercial area, transportation hub, etc. Different application scenarios have different requirements for network quality and optimization focus. 12. Work frequency band: the working frequency range of communication network, which is very important for spectrum resource optimization and interference coordination. 13. Equipment manufacturer: the name of the company that manufactures or supplies network equipment. Different manufacturers' products may have different characteristics and compatibility, which may have some impact on problem diagnosis and optimization.

[0038] In the embodiments of the present disclosure, descriptive keywords, phrases or sentence fragments closely related to network problems can be accurately extracted from the text content in the work order as key information, and the prompt may include but is not limited to: 1. Specific manifestations of network failure (such as packet loss, delay, disconnection, inability to access specific services, etc.); 2. Time, place and frequency of failure occurrence; 3. Affected device model, configuration information, software version; 4. User's network environment (intranet, extranet, Wi-Fi, 4G / 5G, etc.); 5. Basic information of optimization object, such as server IP address, port number, network topology structure, etc.; 6. Preliminary diagnosis or troubleshooting attempts mentioned in the work order and their results.

[0039] The network problem description information refers to a specific description of the current network problem when reporting a network failure or requesting network optimization, which can include: 1, failure phenomenon: such as network disconnection, slow Internet speed, data transmission interruption, video call lag, application response delay, etc. 2, failure time: the specific time period when the problem occurs, whether there is regularity, etc. 3, affected range: how many users, which areas, and what services are affected by this problem. 4, network performance indicators: actual measurement values of packet loss rate, delay, throughput, etc. compared with normal range. 5, comparison before and after failure: network state changes before and after the problem occurs. 6, troubleshooting attempts and results: repair measures that have been tried and effects.

[0040] The network optimization object can refer to network components, devices, configurations, protocols, or services that need to be improved or optimized, such as: 1, network devices: switches, routers, firewalls, servers, base stations, etc. 2, network architecture: network topology, routing table, subnet division, load balancing strategy, etc. 3, network configuration: IP address allocation, bandwidth limitation, QoS policy, security policy, etc. 4, network protocol: specific protocols in the TCP / IP protocol stack such as TCP, UDP, DNS, DHCP, etc. 5, specific services: VoIP services, streaming media services, online gaming services, etc., and performance bottlenecks of these services.

[0041] Step 102, input the network problem description information and network optimization object information into the pre-trained optimization model to determine the network optimization intent. The optimization model is a large language model.

[0042] Specifically, through the pre-trained optimization model, the nature of the network problem and the required objects and direction of optimization can be analyzed according to the network problem description information and network optimization object information. In the embodiment of the present disclosure, the optimization model can use the open-source llama-lora model, use a large amount of wireless network optimization behavior data for incremental pre-training, and further perform prompt tuning Prompt-Tuning on this basis, construct a wireless network optimization scheme generation model based on a local knowledge base, and realize wireless network problem cause analysis, optimization scheme generation, and other wireless network optimization knowledge question and answer (such as based on the description of the optimization work order network problem, optimization object basic information, and network performance indicators, through natural language understanding capability, wireless network optimization scheme recommendation is provided).

[0043] Step 103, query the network performance indicators of the network optimization object network based on the network optimization intent to the network optimization platform.

[0044] The network performance indicators can include, but are not limited to, signal strength (RSSI or RSRP), signal-to-noise ratio (SNR), packet loss ratio, network latency (ping), throughput, link utilization, user experience perception indicators (QoS / QoE), and the like, without limitation.

[0045] For different network optimization intentions and network optimization objects, the corresponding network performance indicators can also be different.

[0046] Specifically, according to the network optimization intention, the key indicators can be selected from the many network performance indicators provided by the platform.

[0047] Step 104, based on the optimization model, the network performance indicators are analyzed to obtain optimization suggestion information for the network problem causes.

[0048] Based on the optimization model, the network performance indicators are analyzed by monitoring and analyzing various performance data of the network (such as bandwidth utilization, delay, packet loss rate, throughput, and concurrent connection number).

[0049] For example, the network performance indicators of cell 460-00-2103557-1 before and after parameter optimization are queried, the wireless connection rate is improved from 93.7 to 99.67, the handover success rate is improved from 93.7 to 99.67, the wireless drop rate is improved from 93.7 to 99.67, the maximum number of RRC connections is improved from 93.7 to 99.67, and the like, and comparative analysis is performed. The optimization scheme is reasonable, and the network performance quality is improved.

[0050] Specifically, based on the results of the perception analysis, the optimization model can provide targeted optimization suggestions, such as adjusting network configuration, increasing bandwidth, replacing or upgrading equipment, optimizing routing strategy, and the like. Through the above process, active identification and timely intervention of network problems can be achieved, and the stability and overall performance of the network can be improved.

[0051] The process of perception analysis mainly includes the following processes:

[0052] For example, if the user provides the current network performance indicators of cell 460-00-2103557-1, including wireless connection rate, wireless drop rate, handover success rate, etc. The system can analyze the degradation of these indicators and point out possible problems, such as low wireless connection rate, high wireless drop rate, etc., and prompt the user to call the root cause diagnosis service module to further locate the network problem. In the scheme optimization process, according to the specific cell network problem cause and parameter information described by the user, such as device manufacturer, working frequency band, etc., the system gives targeted network optimization scheme according to the problem description and analysis results, including adjusting the maximum transmit power, downtilt angle and reference signal power, etc. Network parameters, and suggest querying performance indicators after completing parameter optimization to evaluate optimization effect.

[0053] The following is an example of sensing analysis of network performance indicators based on an optimization model: [

[0055] # Optimization intention recognition process: Recognize network optimization intention through input feedback ticket basic information, and prompt user to query network performance indicators.

[0056] {

[0057] "content": "The optimization object of ticket AF100094 is 5G cell 460-00-2103557-1, and the current network problem is 5G low wireless connection rate. Which network performance indicators should be queried for analysis to locate the network problem cause.",

[0058] "summary": "The current network performance indicators of cell 460-00-2103557-1 can be queried through the \"one-click query\" function module of the network optimization platform (address: http: / / xxx.xxx), including wireless connection rate, wireless drop rate, handover success rate,..., RRC connection maximum number, uplink weak coverage proportion, etc. Whether the indicators are degraded can be identified by calling the \"indicator abnormal fluctuation identification\" function module (address: http: / / xxx.xxx)."

[0059] },

[0060] # Optimization indicator analysis process: Prompt network problem cause by inputting cell network performance indicators and indicator degradation.

[0061] {

[0062] "content": "Cell 460-00-2103557-1 Current network performance indicators: wireless connection rate is 93.7, wireless drop rate is 0.15, handover success rate is 99.8,..., RRC connection maximum number is 3, uplink weak coverage ratio is 35.77. Through analysis, it is found that the wireless connection rate is low, the wireless drop rate is high, the overlapping coverage is high, and the RRC connection maximum number is low, which have deteriorated to some extent",

[0063] "summary": "Through analysis of the index deterioration of the cell, the main network problem is "over coverage-over far access_TA", which can be further located by calling the "one-key root cause diagnosis" function module (address: http: / / xxx.xxx)."

[0064] },

[0065] # Optimization process: Through the input feedback of the work order basic information network problem reason, prompt the network optimization scheme, and give the optimization evaluation opinion.

[0066] {

[0067] "content": "The outdoor 5G cell 460-00-2103557-1 of the bottom residents in the main urban area of Nanping City, the equipment manufacturer of the cell is Huawei, the working frequency band is 700M, the current network problem is low wireless connection rate of 5G, through analysis, it is found that the wireless connection rate is low, the wireless drop rate is high, the overlapping coverage is high, and the RRC connection maximum number is low, which have deteriorated to some extent, the network problem is "over coverage-over far access_TA". How to optimize the cell parameters to improve the network performance quality.",

[0068] "summary": "Query the maximum transmit power, downtilt angle and reference signal power of cell 460-00-2103557-1, suggest to adjust the parameter values to 448, 12 and 132 respectively, after completing the parameter optimization for 3 days, the wireless connection rate, wireless drop rate, RRC connection maximum number and other performance indicators before and after optimization can be queried."

[0069] } ]

[0071] Figure 2 A wireless network optimization work order automatic processing flowchart based on a large model is provided, as shown in Figure 2As shown, the specific implementation process is as follows: after the optimization work order is generated, the prompt extraction is performed, the prompt extraction is mainly to extract the network problem description key information and the network optimization object basic information in the optimization work order, and feedback to the large language model. Among them, the large language model understands the network problem description content and the network environment of the optimization object, identifies the network optimization intention, gives the query instruction, queries the network performance indicators of the optimization object network through the network optimization system and feeds back to the large model, and the large model perceives and analyzes the network performance indicators of the optimization object, gives the network problem reason, and generates the network optimization suggestion, and then displays it in the work order system.

[0072] In the embodiment of the present disclosure, first, the prompt extraction is performed on the network optimization work order to obtain network problem description information and network optimization object information, and then the network problem description information and the network optimization object information are input into the optimization model pre-trained to determine the network optimization intention. The optimization model is a large language model, and then the network optimization platform is queried based on the network optimization intention to obtain the network performance indicators of the network optimization object network, and then the network performance indicators are perceived and analyzed based on the optimization model to obtain the optimization suggestion information for the network problem reason. Therefore, by training the large language model in the field of wireless network optimization, based on the powerful semantic understanding and intention translation ability of the large model, the network optimization intention can be accurately identified, and the understanding of the large language model for the network optimization intention and the understanding of the expert optimization experience cases can be used to quickly recommend the network indicator query method and the network analysis optimization strategy when encountering similar network faults.

[0073] Figure 3 The flowchart of another network optimization work order processing method provided by the embodiment of the present disclosure is shown.

[0074] As Figure 3 shown, the network optimization work order processing method can include the following steps:

[0075] Step 201, performing prompt extraction on the network optimization work order to obtain network problem description information and network optimization object information.

[0076] It should be noted that the specific implementation of step 201 can refer to the above-mentioned embodiments, which will not be described here.

[0077] Step 202, obtaining a training data set, wherein the training data set at least contains text description data of multiple work orders, text description data of network problems and network environments of poor cells, text description data of adjustment strategies of network optimization engineers for each network problem, and text description data of evaluation information of the network optimization engineers on the optimization scheme, wherein the evaluation information is an evaluation conclusion made according to the changes of the optimization target indicators before and after optimization.

[0078] As shown in Table 1 of the following table 1, Table 1 is a schematic table of work order information, mainly including optimization objects and related basic attribute information thereof:

[0079] Table 1

[0080]

[0081] In order to provide effective data samples for large models, the data of the above table 1 is converted into text description, and the text description data is as follows: the optimization object of the network optimization work order AF100094 is the outdoor 5G cell 460-00-2103557-1 of the bottom residents in the main urban area of S city, the equipment manufacturer of the cell is A manufacturer, the working frequency band is 700M, and the current network problem is 5G low wireless connection rate.

[0082] Optionally, it is also necessary to collect network optimization engineer query poor quality cell related performance indicators, combine the network problems and network environment of the poor quality objects to analyze the index degradation and further locate the network poor quality reasons, for example, the main information of the network problems and network environment of the poor quality cell is shown in Table 2 as follows:

[0083] Table 2

[0084]

[0085] As above, the text description data corresponding to the network problems and network environment of the poor quality cell reflected by the above table 2 is converted into text, such as: the current network performance index of the cell 460-00-2103557-1 is wireless connection rate of 93.7, wireless drop rate of 0.15, handover success rate of 99.8,..., RRC connection maximum number of 3, uplink weak coverage proportion of 35.77. Through analysis, it is found that the wireless connection rate is low, the wireless drop rate is high, the overlapping coverage degree is high, and the RRC connection maximum number is low, and these indexes have been degraded to a certain extent, and by calling the "one-key root cause diagnosis" function module, the network reason is found to be "over coverage-over far access TA".

[0086] Optionally, after collecting the network optimization engineer positioning network problem root cause, combining the current value of network performance index, according to the experience of querying the related parameters of optimization object configuration, and formulating adjustment strategy to optimize network quality. For example, a table of network optimization engineer adjustment strategy for each network problem is shown in Table 3 as follows:

[0087] Table 3

[0088]

[0089] The text description data of the adjustment strategy of the network optimization engineer in Table 3 is: querying the network parameters of the maximum transmission power, downtilt angle and reference signal power of cell 460-00-2103557-1, the parameter values are 478, 9 and 132 respectively, wherein the maximum transmission power and the downtilt angle need to be adjusted as they do not reach the optimal parameter value, and are adjusted to 448 and 12 respectively.

[0090] Optionally, the evaluation behavior of the network optimization engineer on the optimization scheme also needs to be collected, and the optimization scheme evaluation conclusion is made by querying the changes of the optimization target indicators before and after optimization. For example, a table for embodying the evaluation information of the network optimization engineer on the optimization scheme is shown in Table 4 as follows:

[0091] Table 4

[0092]

[0093] The text description information of the evaluation information of the network optimization engineer on the optimization scheme in Table 4 is: querying the network performance indicators of cell 460-00-2103557-1 before and after parameter optimization, the wireless connection rate is improved from 93.7 to 99.67, the handover success rate is improved from 93.7 to 99.67, the wireless drop rate is improved from 93.7 to 99.67, the maximum number of RRC connections is improved from 93.7 to 99.67, etc., and comparative analysis is performed, and the optimization scheme is reasonable and improves the network performance quality.

[0094] Optionally, the training data set can also include product documents, protocol specifications, network indicator definition specifications, wireless network optimization cases and other field knowledge of network problems, and the large language model for wireless network optimization scenarios can be researched and implemented through the training data set, which can realize the ability of knowledge learning, analysis and reasoning in complex wireless network optimization scenarios, support wireless network optimization in wireless network perception, network problem cause analysis, network problem optimization scheme generation, network optimization field knowledge question and answer, etc., improve the work efficiency of wireless network optimization personnel, reduce the cost of wireless network optimization, and integrate large model technology and wireless network optimization knowledge.

[0095] Step 203: text segmentation according to the training data set to obtain a plurality of files.

[0096] Optionally, line feed character segmentation can be performed.

[0097] Step 204: vectorization processing is performed on each file to obtain a plurality of first vectors to constitute a vector database and build an index between the vector databases.

[0098] Specifically, the index between each first vector in the vector database can be established based on product documents of network problems, protocol specifications, network index definition specifications, wireless network optimization cases and other domain knowledge.

[0099] As shown in Figure 4 Figure 4 is a training flowchart of a wireless network optimization large model. After vectorization Embedding processing of each module text data, the data is stored in a Faiss vector database. The Embedding layer mainly encodes the network optimization data. This part can include multiple data sources. The purpose is data alignment. The Self Attention attention layer is used to process the data through a multi-head attention mechanism to represent the relationship between data features. When searching, the wireless network optimization problem Query is embedded and the most similar K files Document are found through semantic retrieval. Then the related Document can be assembled into a prompt text Context, which is fed back to the wireless network optimization large model for wireless network optimization knowledge question and answer.

[0100] In step 205, the initial large language model is trained using the vector database. Different loss functions are used in different stages of training.

[0101] Optionally, the present application can calculate the loss loss using different loss functions in different stages. In the optimization intention recognition and optimization index analysis stage, the accuracy of wireless cell network problem cause positioning can be regarded as a classification problem, and the model is evaluated by the accuracy. The cross-entropy loss function is a commonly used loss function in multi-classification problems. Cross-entropy is used to evaluate the difference between the probability distribution obtained by the current training and the true distribution. It describes the distance between the actual output (probability) and the expected output (probability), that is, the smaller the cross-entropy value, the closer the two probability distributions.

[0102] Optionally, the cross-entropy loss function can be used to calculate the loss for predicting the network problem reason based on analysis behavior, and the MSE loss can be used to calculate the loss for predicting the best network optimization scheme based on optimization behavior.

[0103] In the scheme optimization stage, the effect of the model is evaluated based on the improvement degree of the network performance index before and after network optimization, so the mean square error (MSE) loss function is used to calculate the loss.

[0104] ​It can be understood that the present application can adopt different task heads to adapt to different tasks, and the total loss is calculated for backpropagation training. Therefore, training can be carried out in a standard language model pipeline without any special adaptation, and only one process supervision forward pass is required for the entire solution to determine the prediction of each stage during testing. For the correct solution, both the result and the process supervision provide the same information, i.e., each step is correct. For the wrong solution, both methods reveal the existence of at least one error, and the process supervision can also show the specific location of the error. This decision also keeps the artificial annotation cost similar and does not rely on the final answer that is easy to check, and determining the correctness of the solution is equivalent to determining its first error.

[0105] In step 206, the reasoning steps of the initial large language model are rewarded or punished based on the process supervision reward learning method.

[0106] It should be noted that general large models generally use outcome-supervised reward models (ORMs), i.e., only the final result of the model thought chain is used for training, which may cause logical illusions. In order to train a more reliable large model in the field of wireless network optimization, the present application adopts a process supervision reward model, i.e., accepts the reward of each step in the thought chain. The best optimization solution for wireless network problems is achieved by rewarding each correct reasoning step (“process supervision”) instead of simply rewarding the correct final answer (“outcome supervision”).

[0107] Optionally, the initial large language model can first determine the reasoning answer results corresponding to each reasoning task in the reasoning process based on pre-defined reasoning tasks, then determine the reward and punishment strategy corresponding to the reasoning answer results based on pre-set rules, and finally update the model parameters of the initial large language model based on the reward and punishment strategy.

[0108] It can be understood that the reward and punishment standards can be defined in advance according to the types of reasoning tasks. For example, for question answering tasks, rewards can be given according to the accuracy and completeness of the answers, and punishments can be given for incorrect or incomplete answers. When the model performs reasoning steps, its behavior can be supervised, and rewards or punishments can be given according to the actual results. According to the supervision results, the model is rewarded or punished. For correct reasoning steps, positive rewards are given, such as increasing its score or providing additional rewards; for incorrect reasoning steps, negative punishments are given, such as reducing its score or providing punitive feedback. Finally, the model parameters can be updated according to the results of the reward and punishment processing to improve its reasoning ability and effect. Reinforcement learning algorithms such as policy gradient methods can be used to optimize the behavior of the model.

[0109] Further, the inference task can be first defined, for example, in the network optimization scenario, the inference task can be to solve the poor quality cell problem or to improve the network environment. Then, for each inference task, the inference answer result needs to be defined, that is, the answer given by the model after completing the inference.

[0110] Then, the initial large language model can be used for inference to obtain the inference answer result of each inference task. These results can be the answers of the model to specific problems, such as the recommended optimization strategy or problem solution.

[0111] Optionally, a series of rules can be preset to determine the inference answer result to which each inference result belongs, and determine the corresponding reward and punishment strategy. These rules can be based on domain knowledge, expected results and experience judgment. For example, if the inference result is a solution to the poor quality cell problem, and the solution proves effective in practice, a positive reward is given; if the inference result is an unrealistic or unfeasible solution, a negative reward and punishment is given.

[0112] Finally, according to the obtained reward and punishment strategy, the parameters of the initial large language model can be updated. For the inference result that obtains a positive reward, the weight of the corresponding parameter can be increased to enhance the model's ability to correctly answer similar inference tasks; for the inference result that obtains a negative reward, the weight of the corresponding parameter can be reduced to reduce the model's preference for incorrect answers.

[0113] Step 207, the trained initial large language model is used as an optimization model.

[0114] Step 208, the network problem description information and the network optimization object information are input into the pre-trained optimization model to determine the network optimization intention, and the optimization model is a large language model.

[0115] Step 209, based on the network optimization intention, the network performance indicators of the network optimization object network are queried from the network optimization platform.

[0116] Step 210, based on the optimization model, the network performance indicators are analyzed to obtain optimization suggestion information for the network problem cause.

[0117] It should be noted that the specific implementation of steps 207-210 can refer to the above embodiments, which will not be described here.

[0118] In the embodiments of the present disclosure, in order to increase the adaptability of the network, reduce the complexity of network management, and improve the efficiency of network operation, the field meaning and the description of the related attribute values in the data are used for pre-training of an open source large model by collecting the network optimization behavior data of historical network optimization experts and the network performance evaluation data before and after optimization, network optimization intention translation recognition is realized, optimization scheme suggestions for network performance degradation are given according to the optimization intention, and the optimization behavior of the optimization experts in each stage of network optimization is scored by considering the importance of training a reliable model and the high cost of artificial feedback, a process supervision method is used to train a reward model, and the model effect is improved. The large model technology and wireless network optimization knowledge are fused, based on product documents, protocol specifications, network index definition specifications, wireless network optimization cases and other field knowledge, a large language model for wireless network optimization scenarios is researched and implemented, the ability of knowledge learning, analysis and reasoning in a complex wireless network optimization scenario is realized, wireless network optimization is assisted in wireless network perception, network problem cause analysis, network problem optimization scheme generation, network optimization field knowledge question and answer and other aspects, the work efficiency of wireless network optimization personnel is improved, and the cost of wireless network optimization is reduced.

[0119] It should be noted that the present application also has the following advantages:

[0120] 1. The storage of data such as data query behavior, data analysis behavior and network optimization scheme formulation behavior of network optimization engineers provides effective data support for wireless intelligent optimization.

[0121] 2. The generative large model training method based on wireless network optimization learning can accurately understand the network optimization intention, give network problem causes and corresponding network optimization suggestions, greatly save network optimization time, reduce user complaints and improve network self-intelligence level.

[0122] 3. In order to train a more reliable large model in the field of wireless network optimization, a process supervision reward model is adopted, that is, the reward of each step in the thinking chain is accepted. Different loss functions are used to calculate loss in different stages, a single label form is adopted, and the logarithmic likelihood of these target labels is maximized in the training process. Rewards are given to each correct reasoning step instead of simply rewarding the correct final answer to achieve the best optimization scheme for solving wireless network problems.

[0123] Figure 5 The structure diagram of the network optimization work order processing device provided by the embodiments of the present disclosure is shown in Figure 5 As shown in the figure, the network optimization work order processing device 500 can include:

[0124] The acquisition module 510 is configured to extract the prompt of the network optimization work order to obtain network problem description information and network optimization object information.

[0125] The determination module 520 is configured to input the network problem description information and the network optimization object information into a pre-trained optimization model to determine a network optimization intention, the optimization model being a large language model;

[0126] The query module 530 is configured to query a network optimization platform for network performance indicators of a network of the network optimization object based on the network optimization intention.

[0127] The analysis module 540 is configured to perform perceptual analysis on the network performance indicators based on the optimization model to obtain optimization suggestion information for network problem causes.

[0128] Optionally, the determination module further includes:

[0129] The acquisition unit is configured to acquire a training data set, wherein the training data set at least includes text description data of multiple work orders, text description data of network problems and network environments of poor cells, text description data of adjustment strategies of network optimization engineers for the network problems, and text description data of evaluation information of the network optimization engineers on optimization schemes, wherein the evaluation information is an evaluation conclusion made according to changes in optimization target indicators before and after optimization.

[0130] The segmentation unit is configured to perform text segmentation according to the training data set to obtain multiple files.

[0131] The processing unit is configured to perform vectorization processing on each file to obtain multiple first vectors to constitute a vector database and build an index between the vector databases.

[0132] Optionally, the processing unit is further configured to:

[0133] train the initial large language model using the vector database, wherein different loss functions are used in different stages of training.

[0134] reward or punish the inference steps of the initial large language model based on a process supervision reward learning mode.

[0135] use the trained initial large language model as the optimization model.

[0136] Optionally, the processing unit is specifically configured to:

[0137] determine inference answer results of the initial large language model corresponding to each inference task in an inference process based on pre-defined inference tasks.

[0138] Determine, based on a preset rule, a reward and punishment strategy corresponding to a reasoning answer result to which the reasoning result belongs;

[0139] Update model parameters of the initial large language model based on the reward and punishment strategy.

[0140] Optionally, the training process of the initial large language model includes a network optimization intention recognition stage and a scheme optimization stage, the network optimization intention recognition stage adopts a cross-entropy loss function, and the scheme optimization stage adopts a mean square error loss function.

[0141] Optionally, the network optimization work order at least includes the following fields: work order ID, work order date, city to which it belongs, object type, object number, object name, problem description, network type, home area, coverage type, coverage scenario, work frequency band, and device manufacturer.

[0142] In the embodiments of the present disclosure, first, a network optimization work order is subjected to prompt extraction to obtain network problem description information and network optimization object information, and then the network problem description information and the network optimization object information are input into a pre-trained optimization model to determine a network optimization intention, the optimization model being a large language model, and then based on the network optimization intention, a network performance index of a network for the network optimization object is queried from a network optimization platform, and then based on the optimization model, a perception analysis is performed on the network performance index to obtain optimization suggestion information for a network problem cause. In this way, by training a large language model in the field of wireless network optimization, based on the powerful semantic understanding and intention translation capability of the large model, the network optimization intention can be accurately recognized, and by using the understanding of the network optimization intention by the large language model and the understanding of expert optimization experience cases, when a similar network fault occurs, a network index query method and a network analysis optimization strategy can be quickly recommended to the user.

[0143] To implement the above-mentioned embodiments, the present application further provides an electronic device, as shown in Figure 6 Figure 6 is a block diagram of an electronic device for processing a network optimization work order according to an exemplary embodiment.

[0144] As shown in Figure 6 , the above-mentioned electronic device 800 includes:

[0145] The memory 810 and the processor 820, the bus 830 connecting different components (including the memory 810 and the processor 820), the memory 810 storing a computer program, and the processor 820 executing the program to implement the processing method of the network optimization work order according to the embodiments of the present disclosure.

[0146] ​Bus 830 represents one or more of several types of bus structures, including a memory bus or memory controller, a peripheral bus, a graphics acceleration bus, a processor or local bus using any of a variety of bus architectures. By way of example, these architectures include Industry Standard Architecture (ISA) bus, Micro Channel Architecture (MCA) bus, Enhanced ISA (EISA) bus, Video Electronics Standards Association (VESA) local bus, and Peripheral Component Interconnect (PCI) bus.

[0147] Electronic device 800 typically includes a variety of electronic device readable media. These media can be any available media that is located either internally or externally to electronic device 800, including both volatile and nonvolatile media, removable and non-removable media.

[0148] Memory 810 also can include computer system readable media in the form of volatile memory, such as random access memory (RAM) 840 and / or cache memory 850. Electronic device 800 can further include other removable / non-removable, volatile / non-volatile computer system storage media. By way of example only, storage system 860 can be provided for reading from and writing to a non-removable, non-volatile magnetic media (e.g., a "hard drive"). Figure 6 not shown, is typically provided as residual storage across electronic device 800, and can be used for storing data that is both created by and / or maintained by electronic device 800, including index scores, etc. Electronic device 800 can also contain a computer system readable media reader 850 that can further read instructions from a computer system readable storage medium such as a Figure 6 not shown, is typically provided as residual storage across electronic device 800, and can be used for storing data that is both created by and / or maintained by electronic device 800, including index scores, etc. Electronic device 800 can also contain a computer system readable media reader 850 that can further read instructions from a computer system readable storage medium such as a

[0149] Program / utility 880 having a set (at least one) of program modules 870 can be stored in, for example, memory 810 by way of example, such program modules 870 include, but are not limited to, an operating system, one or more applications, other program modules, and program data, each of which or a combination thereof, can include implementation of a network environment. Program modules 870 generally carry out the functions and / or methodologies of embodiments of the disclosure as described herein.

[0150] The electronic device 800 can also communicate with one or more external devices 890 such as a keyboard or pointing device, a display, etc.; other devices such as a storage device (removable and / or non-removable) including, but not limited to, magnetic or optical disks or tape and software storage, etc. The communication can be through the input / output (I / O) interfaces 892. Additionally, the electronic device 800 can communicate with one or more networks, such as one or more local area networks (LANs), wide area networks (WANs), telecommunication networks, and / or the Internet, etc., through a network adapter 893. As depicted, the network adapter 893 communicates with the other components of the electronic device 800 through the bus 830. It should be appreciated that the network adapter 893 can also be connected to the other components of the electronic device 800 through a network link that is connected via an I / O interface 892. Figure 6 As depicted, the network adapter 893 communicates with the other components of the electronic device 800 through the bus 830. It should be appreciated that the network adapter 893 can also be connected to the other components of the electronic device 800 through a network link that is connected via an I / O interface 892. Figure 6 It should be appreciated that the network adapter 893 can also be connected to the other components of the electronic device 800 through a network link that is connected via an I / O interface 892. Other hardware and / or software modules that can be used in the electronic device 800, but are not shown in FIG. 8, include but are not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data archival storage systems, etc.

[0151] The processor 820 performs various function applications and data processing by running programs stored in the memory 810.

[0152] It should be noted that the implementation process and technical principles of the electronic device of the present embodiment are described above in the description of the network optimization work order processing method of the present embodiment, and will not be described here.

[0153] To implement the above-mentioned embodiments, the present application further provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the network optimization work order processing method of the above-mentioned embodiments.

[0154] To implement the above-mentioned embodiments, the present application further provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the network optimization work order processing method of the above-mentioned embodiments.

[0155] In the description of the present specification, the terms "first", "second" are only for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first", "second" can explicitly or implicitly include at least one of the features. In the description of the present application, the meaning of "multiple" is at least two, such as two, three, etc., unless otherwise specifically limited.

[0156] In the description of the specification, the description using the terms "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" etc. means that the particular feature, structure, material or characteristic being described is included in at least one embodiment or example of the present application. The illustrative appearances of the above-mentioned terms in various places in the specification are not necessarily referred to the same embodiment or example. Moreover, the particular features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples. Furthermore, the terms "comprise", "comprising", "include", "including", "contain", "containing" or variations thereof are used inclusively and do not exclude the additional inclusion of unrecited features, structures, materials, or characteristics.

[0157] Although the embodiments of the present application have been shown and described above, it is understood that the above-described embodiments are exemplary and are not to be interpreted as limiting the present application, and ordinary skilled people in the art can make changes, modifications, replacements and variations to the above-described embodiments within the scope of the present application.

Claims

1. A method for processing a network optimization work order, the method comprising: The method comprises the following steps: extracting prompt from a network optimization work order to obtain network problem description information and network optimization object information; inputting the network problem description information and the network optimization object information into a pre-trained optimization model to determine a network optimization intention, wherein the optimization model is a large language model; querying a network optimization platform for network performance indicators for the network optimization object information based on the network optimization intention; performing perception analysis on the network performance indicators based on the optimization model to obtain optimization suggestion information for a network problem cause; wherein, before the network problem description information and the network optimization object information are input into the pre-trained optimization model to determine the network optimization intention, the method further comprises the following steps: obtaining a training data set, wherein the training data set at least contains text description data of multiple work orders, text description data of network problems and network environment of poor quality cells, text description data of adjustment strategies of network optimization engineers for each network problem, and text description data of evaluation information of optimization schemes of network optimization engineers, wherein the evaluation information is an evaluation conclusion made according to the changes of optimization target indicators before and after optimization; performing text segmentation according to the training data set to obtain multiple files; performing vectorization processing on each file to obtain multiple first vectors to constitute a vector database and build an index between the vector databases; training an initial large language model using the vector database, wherein different loss functions are used in different stages of training; rewarding and punishing the reasoning steps of the initial large language model based on a process supervision reward learning method; using the trained initial large language model as the optimization model.

2. The method of claim 1, wherein, Wherein, the training process of the initial large language model includes a network optimization intention recognition stage and a scheme optimization stage, the network optimization intention recognition stage uses a cross-entropy loss function, and the scheme optimization stage uses a mean square error loss function.

3. The method of claim 1, wherein, Wherein, the network optimization work order at least contains the following fields: work order ID, work order date, city, object type, object number, object name, problem description, network type, attribution area, coverage type, coverage scenario, work frequency band, and device manufacturer.

4. The method of claim 1, wherein, The reward and punishment processing of the reasoning steps of the initial large language model based on the process supervision reward learning method comprises: determining reasoning answer results corresponding to each reasoning task in the reasoning process of the initial large language model based on pre-defined reasoning tasks; judging the reward and punishment strategy corresponding to the reasoning answer result to which the reasoning answer result belongs based on pre-set rules; updating the model parameters of the initial large language model based on the reward and punishment strategy.

5. A network optimization ticket processing apparatus, comprising: The method comprises the following steps: an acquisition module for extracting prompt from a network optimization work order to obtain network problem description information and network optimization object information; a determination module for inputting the network problem description information and the network optimization object information into a pre-trained optimization model to determine a network optimization intention, wherein the optimization model is a large language model; The query module is configured to query, from a network optimization platform, a network performance indicator for the network optimization object information based on the network optimization intention. The analysis module is configured to perform perceptual analysis on the network performance indicator based on the optimization model to obtain optimization suggestion information for a network problem cause. The determination module is further configured to: obtain a training data set, wherein the training data set at least includes text description data of a plurality of work orders, text description data of network problems and network environments of poor-quality cells, text description data of adjustment strategies of network optimization engineers for the network problems, and text description data of evaluation information of the network optimization engineers on optimization schemes, wherein the evaluation information is an evaluation conclusion made according to changes in optimization target indicators before and after optimization; perform text segmentation according to the training data set to obtain a plurality of files; perform vectorization processing on each file to obtain a plurality of first vectors to form a vector database and build an index between the vector databases; train an initial large language model by using the vector database, wherein different loss functions are used in different stages of training. perform reward and punishment processing on inference steps of the initial large language model based on a process supervision reward learning manner. use the trained initial large language model as the optimization model.

6. The apparatus of claim 5, wherein, The training process of the initial large language model includes a network optimization intention recognition stage and a scheme optimization stage, the network optimization intention recognition stage uses a cross-entropy loss function, and the scheme optimization stage uses a mean square error loss function. The network optimization work order at least includes the following fields: work order ID, work order date, city to which it belongs, object type, object number, object name, problem description, network type, home area, coverage type, coverage scenario, work frequency band, and device manufacturer.

7. The apparatus of claim 5, wherein, A computer program is stored in the memory and executable on the processor, and when the processor executes the program, the network optimization work order processing method of any one of claims 1-4 is implemented. The computer program is executed by the processor to implement the network optimization work order processing method of any one of claims 1-4.

8. An electronic device, comprising: ​ 9. A computer readable storage medium having stored thereon a computer program, characterized in that, ​

Images