Network state intelligent query method, device, equipment and medium

Abstract

The embodiment of the invention discloses a network state intelligent query method and device, equipment and a medium, and the method comprises the steps: receiving a first query statement inputted by a user, and determining a query intention of the user and first key information corresponding to the query intention based on the first query statement; if the first key information does not meet the preset query requirement, generating guide information according to a cascade relationship between the first key information, and sending the guide information to the user; determining second key information sent by the user based on the guide information, and obtaining associated data related to the query intention from the network state database according to the second key information; and generating a query result based on the associated data and the query intention, and sending the query result to the user. According to the network state intelligent query method provided by the invention, the user is helped to give accurate key information in time in a guiding manner. The problem that the efficiency is low due to the fact that a user is unfamiliar with query parameters can be effectively solved.

Description

Technical Field

[0001] The application relates to the field of computer technology, and in particular to a method, apparatus, device, and medium for intelligent querying of network status. Background Technology

[0002] In traditional network status query scenarios, users typically select query criteria through a graphical user interface (such as a webpage or app) to obtain the desired information. After the user completes the input, the frontend encapsulates these query criteria into request parameters (usually in JSON format) and sends them to the backend via an HTTP POST request. The backend dynamically generates SQL query statements based on the transmitted parameters, performs filtering, aggregation, and sorting operations on the database tables, and returns the query results to the frontend in JSON format for the user to view.

[0003] Therefore, the accuracy of traditional implementations depends on the user's familiarity with the query parameters, requiring step-by-step manual selection and a relatively fixed interaction method. This not only increases the user's learning curve but also leads to low query efficiency, failing to meet the needs for rapid response and problem troubleshooting. Summary of the Invention

[0004] This application addresses some of the deficiencies mentioned in the background art by providing a method, apparatus, device, and medium for intelligent network status query.

[0005] In a first aspect, embodiments of this application provide a method for intelligent network status query, comprising: Receive a first query statement input by the user, and determine the user's query intent and the first key information corresponding to the query intent based on the first query statement; If the first key information does not meet the preset query requirements, guidance information is generated and sent to the user based on the cascading relationship between the first key information. Determine the second key information sent by the user based on the guidance information, and obtain related data related to the query intent from the network status database based on the second key information; Based on the associated data and the query intent, a query result is generated and sent to the user.

[0006] In some embodiments of this application, receiving a first query statement input by a user, and determining the user's query intent and the first key information corresponding to the query intent based on the first query statement, includes: The first query statement is input into the first model, and the query intent and the first key information corresponding to the query intent are obtained through the first model.

[0007] In some embodiments of this application, the first query statement is input into a first model, and the query intent and the first key information corresponding to the query intent are obtained through the first model, including: The first query statement is input into the first network structure and the second network structure of the first model, respectively; The query intent and the first key information are obtained by merging the outputs of the first network structure and the second network structure in a predetermined manner. The second network structure is trained based on the output of the first network structure and the corresponding training data.

[0008] In some embodiments of this application, if the first key information does not meet the preset query requirements, generating guidance information based on the cascading relationship between the first key information and sending it to the user further includes: If the first key information does not meet the completeness requirement of the query intent, determine the descriptive information related to the first key information; Based on the description information and the first key information, guide information is generated and sent to the user; The step of determining the second key information sent by the user based on the guidance information, and obtaining related data related to the query intent from the network status database based on the second key information, includes: Determine the second key information confirmed by the user based on the guidance information; If the second key information satisfies the integrity requirement of the query intent, a second query statement is generated based on the second key information; Based on the second query statement, retrieve the associated data related to the query intent from the network state database.

[0009] In some embodiments of this application, if the second key information satisfies the integrity requirement of the query intent, a second query statement is generated based on the second key information, including: Determine the format requirements corresponding to the second key information and the logical requirements between the second key information; The second key information shall be corrected according to the format requirements and / or the logical requirements; A second query statement is generated based on the corrected second key information.

[0010] In some embodiments of this application, generating query results based on the associated data and the query intent, and sending the query results to the user, includes: The associated data and the query intent are input into the second model, and the second model generates the query results based on the associated data and the query intent.

[0011] In some embodiments of this application, the method further includes: Determine the verification data corresponding to the query results; The second training data is composed of the query results and the verification results; The second model is retrained based on the second training data within a predetermined time period.

[0012] In a second aspect, embodiments of this application also provide a network status intelligent query device, comprising: The intent analysis module is used to receive a first query statement input by the user, and determine the user's query intent and the first key information corresponding to the query intent based on the first query statement; The guidance information generation module is used to generate guidance information and send it to the user if the first key information does not meet the preset query requirements, based on the cascading relationship between the first key information. The associated data determination module is used to determine the second key information sent by the user based on the guidance information, and to obtain associated data related to the query intent from the network status database based on the second key information; The query result generation module is used to generate query results based on the associated data and the query intent, and send the query results to the user.

[0013] In a third aspect, embodiments of this application provide a computer device, including: At least one processor; and The memory stores computer instructions that can run on the processor, which, when executed by the processor, implement the steps of any one of the network status intelligent query methods described in the above embodiments.

[0014] In a fourth aspect, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of any one of the network status intelligent query methods described in the above embodiments.

[0015] This application provides an intelligent network status query method. First, it determines the user's query intent and first key information based on the user's initial query statement. Then, if the first key information does not meet the query requirements, it generates guiding information based on the cascading relationships between the first key information, guiding the user to input correct second key information to retrieve related data from the database. Finally, it generates the corresponding query results based on the related data. This guided approach helps users provide accurate key information promptly and effectively solves the inefficiency problem caused by users' unfamiliarity with query parameters. Attached Figure Description

[0016] Figure 1 A flowchart illustrating an intelligent network status query method provided in an embodiment of this application; Figure 2 This is a schematic diagram of the structure of a network status intelligent query model provided in an embodiment of this application; Figure 3 This is a schematic diagram of a hierarchical guidance process provided in an embodiment of this application; Figure 4 This is a schematic diagram of a hierarchical guided interaction process provided in an embodiment of this application; Figure 5 This is a schematic diagram of timing prediction results provided in an embodiment of this application; Figure 6 This is a schematic diagram of anomaly detection reasoning results provided in an embodiment of this application; Figure 7 This is a schematic diagram of anomaly detection reasoning results provided in an embodiment of this application; Figure 8 A schematic diagram of the structure of a network status intelligent query device provided in an embodiment of this application; Figure 9 This is a schematic diagram of the structure of a computer device provided in one embodiment of this application; Figure 10 This is a schematic diagram of the structure of a computer-readable storage medium proposed in one embodiment of this application. Detailed Implementation

[0017] The present application / disclosure will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present application / disclosure and are not intended to limit the scope of the present application / disclosure. Furthermore, it should be noted that, for ease of description, only the parts relevant to the present application / disclosure are shown in the accompanying drawings, not the entire structure.

[0018] This application aims to address the problems of low query efficiency and high learning costs in the field of network status optimization. Traditionally, network status is obtained by contacting back-end personnel by phone or using reporting functions in applications (APPs). This approach relies on human experience to optimize the query content, requiring manual queries or pre-configured database query statements. It cannot flexibly address the data query needs of different real-world situations.

[0019] like Figure 1 As shown, in a first aspect, this application provides a method for intelligent network status query, comprising: Step S1: Receive the first query statement input by the user, and determine the user's query intent and the first key information corresponding to the query intent based on the first query statement; Step S2: If the first key information does not meet the preset query requirements, generate guidance information based on the cascading relationship between the first key information and send it to the user; Step S3: Determine the second key information sent by the user based on the guidance information, and obtain related data related to the query intent from the network status database based on the second key information; Step S4: Generate query results based on related data and query intent, and send the query results to the user.

[0020] In the embodiments of this application, the first query statement is a natural language query statement or a question entered by the user.

[0021] Query intent refers to the data or results that a user wants to extract based on the natural language query input by the user.

[0022] The first key information refers to the keywords or key phrases extracted from the natural language query entered by the user. These keywords and key phrases can be used to determine data related to the user's query intent.

[0023] The second key information refers to the natural language statements or corresponding key information that the user re-enters with the help of the guidance information.

[0024] Cascading relationships refer to the hierarchical relationships between primary key information. These hierarchical relationships are determined by pre-setting logical relationships based on the real meaning of the data and by the actual representation of the corresponding data content.

[0025] Guidance information refers to prompts that guide the user to provide accurate information after correcting any unclear content in the initial key information entered by the user.

[0026] In step S1, the system receives a natural language query statement input by the user and analyzes the data the user wants to obtain and the key information of the natural language query statement input by the user. For example, if the user inputs "get the 5G network coverage rate of a certain area", the corresponding query intent is "coverage rate", and the corresponding key information is "a certain area" and "5G (5th Generation Mobile Networks)".

[0027] In step S2, after determining the first key information, it is determined whether the first key information meets the corresponding query requirements. If the first key information does not meet the corresponding query requirements, guidance information is generated based on the cascading relationship corresponding to the first key information. Specifically, the key information obtained in step S1 is queried in the database to determine whether an accurate and unique query result can be found. The corresponding database can be a structured database or a vector database. If the database query results are not unique, prompt information is generated based on the multiple query results and the cascading relationship of the corresponding key information. For example, if a user enters "get the 5G network coverage rate of region A", the user's input of "region A" is their common name for the corresponding region, which cannot be strictly aligned with the content recorded in the database. If multiple addresses containing this information are obtained from the database, the parent region to which "region A" belongs is used as the prompt information for this key information based on the hierarchical relationship of the addresses. This allows the user to receive the prompt information and confirm the region they want to query.

[0028] In step S3, after receiving the second key information input by the user based on the prompt, relevant content is retrieved from the corresponding database based on the second key information, serving as associated information. The number of associated information may be one or more pieces of associated data based on key information of different dimensions. For example, if the user queries the operational status data of a certain base station, it may return one piece of data for that base station; if the user queries the operational status data of that base station over a period of time, it may return multiple pieces of data. Therefore, the number of associated data is related to the user's query intent and the key information corresponding to the query intent.

[0029] In step S4, the corresponding query results are generated based on the associated data and sent to the user. For example, if the user wants to query "get the 5G network coverage rate of region A", the generated result will be the 5G network coverage rate of region A. This is accurate to a numerical value.

[0030] In some embodiments of this application, receiving a first query statement input by a user, and determining the user's query intent and the first key information corresponding to the query intent based on the first query statement, includes: The first query statement is input into the first model, and the query intent and the first key information corresponding to the query intent are obtained through the first model.

[0031] In this embodiment, to address the universality of query intent and key information extraction, a large language model is used to identify the intent and extract key information from the user's input natural language query. Specifically, the user's input natural language query and corresponding prompt words are input into the large language model, which then provides the corresponding query intent and key information.

[0032] In some embodiments of this application, the first query statement is input into a first model, and the query intent and the first key information corresponding to the query intent are obtained through the first model, including: The first query statement is input into the first network structure and the second network structure of the first model, respectively; The query intent and the first key information are obtained by merging the outputs of the first network structure and the second network structure in a predetermined manner. The second network structure is trained based on the output of the first network structure and the corresponding training data.

[0033] In this embodiment, to address the accuracy of the large language model's understanding of data in the domain to which this application is applied, this application fine-tunes and trains a second network structure based on the output data of the large language model. The user's natural language query is then input into both the first and second network structures. The outputs of the first and second network structures are then combined to obtain the corresponding query intent and key information. The first network structure is a traditional large language model, and the second network structure is a neural network with the same input-output size as the first network structure.

[0034] In the embodiments of this application, the first network structure adopts an open-source general-purpose large language model, and the second network structure is obtained by training on this general-purpose large language model through LoRA (Low-Rank Adaptation of Large Language Models). The specific network structure is as follows: Figure 2 As shown in the figure, PretrainedWeights on the left represents the weight data of the general large language model W. A and B on the right are obtained by fine-tuning the network state data based on the domain in which this application is applied. The input dimension of A and the output dimension of B are the same as the input and output dimensions of the general large language model, while the output dimension of A and the input dimension of B are values ​​much smaller than the original model's input and output dimensions. This significantly reduces the number of parameters to be trained.

[0035] During training, only the parameters of A and B are updated; the parameters of the pre-trained general-purpose language model remain fixed. During inference, the concept of reparameterization can be used to merge AB with W, thus avoiding the introduction of additional computation during inference. Therefore, for inference tasks in different application scenarios, it is only necessary to retrain AB based on the pre-trained model, which can also accelerate the training pace of large models.

[0036] Specifically, taking a certain 7B open-source large language model as an example, the parameters of the large language model are first loaded. Perform initialization and compute the maximum conditional language model probability update. ,Right now:

[0037] in Indicates model parameters, This represents the training dataset; It is the input data. It is the target sequence; Indicates the length of the target sequence; This represents all words in the target sequence before time t; Indicates that given input Given all words before time t, predict the current target. The conditional probability.

[0038] The parameter increments learned by this fine-tuning method Dimensions and pre-trained parameters It is consistent, therefore requiring a great deal of resources. However, it uses fewer parameters to represent what needs to be learned. That is the most economical and efficient method, namely The optimization goal will shift from finding Transform into searching ,Right now

[0039] The objective is then transformed into maximizing parameters. ,in Indicates model parameters, This represents the training dataset; It is the input data. It is the target sequence; Indicates the length of the target sequence.

[0040] This allows for fine-tuning of only a portion of the parameters, enabling the model to be trained using a small number of samples, thus making it more specialized in the vertical domain knowledge corresponding to the network state data.

[0041] In some embodiments of this application, if the first key information does not meet the preset query requirements, generating guidance information based on the cascading relationship between the first key information and sending it to the user further includes: If the first key information does not meet the completeness requirement of the query intent, determine the descriptive information related to the first key information; Based on the description information and the first key information, guide information is generated and sent to the user; The step of determining the second key information sent by the user based on the guidance information, and obtaining related data related to the query intent from the network status database based on the second key information, includes: Determine the second key information confirmed by the user based on the guidance information; If the second key information satisfies the integrity requirement of the query intent, a second query statement is generated based on the second key information; Based on the second query statement, retrieve the associated data related to the query intent from the network state database.

[0042] In this embodiment, the user's first query (natural language query) is often a colloquial expression of the content to be queried. This expression has good information accuracy based on the user's geographical or time range, while the content stored in the database often requires the uniqueness of the data. This inherently contradicts the user's expression. Therefore, the key information obtained from the user's natural language query may not be able to accurately match the corresponding content, or the user's input may be incomplete, making it impossible to obtain the accurate data to be queried. For example, the user inputs a cell name, which needs to be converted into a base station cell name. This is not only because base station cells and physical residential cells have different naming conventions, for example, a base station cell might be named "JQ-Dunhuang City-Dunhu Community-ZEW-773891-Dunhu Community-EI-165," but the user only remembers "Dunhu Community" and will also input "Dunhu Community" verbally. However, when the model extracts "Dunhu Community" as a parameter, it does not meet the query conditions. In addition, the naming method of base station cells is inconsistent with the naming of residential communities in the physical world. For example, base station cells are named "BY-Baiyin District-Meteorological Bureau Base Station 1-ZEH-912110-Northwest Lead and Zinc Smelting Plant-EO-228", "BY-Huining County-Riverbank Convergence Room 1-ZEH-912143-Riverbank Town Mingde Community Remote-EO-230", etc.

[0043] Therefore, in this embodiment, it is first determined whether the extracted first key information meets the completeness requirements of the corresponding query intent. If the completeness requirements are not met, data related to the first key information is obtained as descriptive information. Descriptive information refers to data related to the corresponding key information. Then, corresponding guidance information is generated based on the descriptive information and sent to the user.

[0044] Furthermore, the system receives new natural language queries or second key information from the user following the guidance information. If the user provides a new natural language query, the system extracts the corresponding second key information from it. It then verifies whether the second key information meets the query requirements. If it does, it retrieves the corresponding related data from the database based on the second key information. If the query requirements are still not met, the system generates guidance information again and sends it to the user.

[0045] Specifically, such as Figure 3 As shown, Figure 3 The illustration shows a hierarchical query guidance diagram for a target cell provided in an embodiment of this application. By guiding the user to input parameters such as the city / county where the target cell belongs, network type, key parameters of the query object, and key parameters of the query time, the system helps the system quickly retrieve the relevant base station cell name from the database, so that the user can select and confirm the final base station cell name and thereby query the base station cell network quality data.

[0046] Furthermore, such as Figure 4 As shown, Figure 4 This diagram illustrates the interactive process of querying 5G coverage in a specific region according to an embodiment of this application. First, the user inputs a natural language query, "5G coverage of the Linxia production building." The intelligent tool implemented by the solution provided in this application generates prompts based on the district / county hierarchy of "Linxia." Then, the user determines the region of their target query, namely "Yongjing County." Further prompts are generated based on the descriptive information related to "production building," further narrowing the scope for the user. After the user inputs the specific address "Yongjing State Taxation Bureau," a query is performed based on "Yongjing State Taxation Bureau" to determine if the uniqueness requirement is met. Three relevant cells are found. Corresponding prompts are then generated and sent to the user. The user selects the accurate address to determine the second key information of the spatial range. The second key information of the time range corresponding to the user's query result is further confirmed. Corresponding prompts are generated to remind the user to input the exact time information. Finally, the second key information corresponding to the spatial and time range dimensions is determined. Then, the database is queried based on the second key information to obtain the corresponding related data. Finally, the query results are generated.

[0047] In some embodiments of this application, if the second key information satisfies the integrity requirement of the query intent, a second query statement is generated based on the second key information, including: Determine the format requirements corresponding to the second key information and the logical requirements between the second key information; The second key information shall be corrected according to the format requirements and / or the logical requirements; A second query statement is generated based on the corrected second key information.

[0048] In this embodiment, during parameter extraction from the large model, the generation mechanism may produce illusions, leading to extraction results that do not conform to the data interface format specifications and business logic requirements, thus affecting the system's stability and accuracy. To address this issue, a comprehensive parameter calibration strategy needs to be designed to improve the reliability and consistency of the extraction results. First, for time-related parameters (such as "start time" and "end time"), they should be uniformly converted to a standard format, and incomplete or ambiguous time information should be completed. For example, "November 10, 2024" should be converted to "2024-11-11". Second, in terms of geographic information processing, the differences between the extraction results and the interface storage format need to be standardized. For example, "Linxia" should be completed to "Linxia City", and consistency verification should be achieved through a place name standardization dictionary or contextual information. In addition, for the extraction results of free text parameters, character cleanup rules need to be designed to remove redundant characters or content that does not meet the interface requirements. For example, "Zhangjiakou (Hebei)" should be formatted as "Zhangjiakou". Simultaneously, the calibration strategy also needs to check the completeness and logical consistency of the extracted results, such as ensuring that the "end time" is not earlier than the "start time" and that the query granularity parameters meet business logic requirements. To cope with dynamic changes in data interface standards, the calibration strategy should have dynamic adaptability, achieving rapid synchronization by maintaining a parameter rule base, and optimizing calibration logic and model extraction capabilities through a feedback mechanism. Through the above multi-level calibration strategy, deviations in the large model parameter extraction process can be effectively eliminated, ensuring that the extracted results conform to the data interface specifications and improving the robustness and practicality of the system in complex scenarios. In some embodiments of this application, generating query results based on the associated data and the query intent, and sending the query results to the user, includes: The associated data and the query intent are input into the second model, and the second model generates the query results based on the associated data and the query intent.

[0049] In this embodiment, if the user's query intent is predicted based on multiple management data points, the corresponding related data and query intent are input into the corresponding second model, which then provides the prediction result. This prediction result is then used as the user's query result.

[0050] In some embodiments of this application, the method further includes: Determine the verification data corresponding to the query results; The second training data is composed of the query results and the verification results; The second model is retrained based on the second training data within a predetermined time period.

[0051] In this embodiment, validation data refers to the actual results corresponding to the predicted results output by the second model (i.e., the query results sent to the user). The query results and actual results are combined to form training data. Then, the second model is retrained at predetermined intervals to dynamically correct its prediction accuracy. In network state query optimization, this method can simultaneously predict network metrics and detect anomalies. When using the trained model, after confirming the user's query conditions, it only needs to automatically select the previous month's data as input time series data. It can then generate predicted values ​​and anomaly markers through a shared encoder and task-specific headers, providing comprehensive support to the user. This method not only improves efficiency and generalization ability but also simplifies the actual deployment process, making it a powerful tool for intelligent network management. Specifically, firstly, network optimization time-series data containing a small number of labeled anomalies from the past month are collected. After shared preprocessing (RevIN normalization, sliding window partitioning, and linear embedding), global temporal features are extracted by a shared Transformer encoder. Subsequently, the features are split into two task branches—the TIME-LLM branch generates future H-step metric prediction values ​​and calculates the prediction loss through block reprogramming, freezing LLM inference, and output projection. The Anomaly Transformer branch completes feature reconstruction and association difference calculation through the Anomaly-Attention dual-branch (prior association + sequence association), thereby obtaining the anomaly detection loss. Finally, the dual-branch losses are summarized through a joint loss function (α1=0.6 weighted prediction loss, α2=0.4 weighted anomaly detection loss), and the parameters of the shared encoder and dual branches are simultaneously optimized through backpropagation to achieve collaborative training that adapts prediction accuracy and anomaly detection. During the inference stage, TIME-LLM outputs future metric prediction results, and Anomaly Transformer outputs anomaly detection results based on the association difference threshold. The two are integrated into a unified feedback mechanism for network optimization scenarios that combines "predicting trends + anomaly alerts".

[0052] Training data composition: Data source: In response to user-initiated network optimization indicator query requests (query time granularity supports daily / weekly / monthly), the system automatically extracts network optimization time-series data for the 30 consecutive days prior to the current query time.

[0053] Data composition: The dataset contains 90%-95% normal network optimization index data and 5%-10% labeled abnormal data (data labeled after verifying the online query results output by the model based on the validation data); among which, normal data is used for basic feature learning of the model, and labeled abnormal data is used for loss calculation and model optimization for anomaly detection tasks.

[0054] Indicator types: Covering core optimization indicators of mobile communication networks, including but not limited to continuous time-series indicators such as user downlink rate, network latency, packet loss rate, signal strength, and cell access success rate.

[0055] Furthermore, shared data preprocessing is performed on the input 30-day network optimization time-series data to ensure data format standardization and feature validity: RevIN normalization: The Reversible Instance Normalization (RevIN) method is used to normalize each network optimization index separately, ensuring that the data meets the distribution characteristics of a mean of 0 and a standard deviation of 1, eliminating the influence of differences in the units of different indices. Sliding window partitioning: A fixed window size of 100 is set, and the normalized time-series data is partitioned by sliding window with a step size of 1 to obtain several continuous time-series subsequences, preserving the time dependencies of the indices. Linear embedding transformation: Each time-series subsequence is mapped to a fixed-dimensional space through a fully connected linear layer, with a unified output dimension of d_model=512, providing adaptive input for subsequent shared feature extraction.

[0056] Furthermore, shared feature extraction involves inputting the preprocessed temporal features into a shared Transformer encoder to extract global temporal features. Encoder structure: The shared Transformer encoder consists of three stacked encoder units, each containing a multi-head self-attention mechanism and a feed-forward network. The number of multi-head self-attention heads is set to 8, and the hidden layer dimension of the feed-forward network is set to 2048.

[0057] Feature extraction logic: Long-distance dependencies in time series data are captured through a multi-head self-attention mechanism. The features are then transformed nonlinearly by a feedforward neural network, and the final output is a global time series feature with a dimension of N×512 (N is the number of subsequences after the sliding window is divided). This feature contains the periodicity, trend and correlation information of network optimization indicators.

[0058] Furthermore, dual-task branch collaborative training is implemented, sharing global temporal features while simultaneously inputting into two parallel task branches to complete temporal prediction and anomaly detection training respectively. Collaborative optimization is achieved through a joint loss function. 1. TIME-LLM time series prediction branch: (1) Block reprogramming: The global temporal features are split into fixed lengths and aligned with the text prototype (text prototype dimension V'=1000). The mapping relationship between temporal features and text prototype is established through a multi-head cross attention mechanism to generate a feature format adapted to LLM input.

[0059] (2) Freezing LLM inference: The fine-tuned Qianwen 2.5 7B model is used as the basic large language model. The pre-training parameters of the model are frozen, and only the block reprogramming layer and the subsequent output layer are trained. The guiding prompt words are constructed by “Prompt-as-Prefix”, with the format: “Task: Network optimization index time series prediction; Domain knowledge: Mobile communication network index with periodic and trend features; Input statistics: Data length = NN, Time granularity = days; Prediction target: Future HH step index value”. The reprogrammed features and the prompt words are concatenated and input into LLM to obtain the intermediate semantic representation of the model.

[0060] (3) Output projection calculation, intermediate semantic representation of LLM output Slice the data and extract the portions corresponding to the time sequence blocks. Flattening process is performed to obtain ;Mapped to the target output dimension (dimension = H) through a linear projection layer. The preset prediction step size (configurable from 1 to 7 days) generates predicted values ​​for future H-step network optimization indicators. The formula is:

[0061] in Let be the projection weight matrix. This is a bias term.

[0062] (4) Calculate the prediction loss by using the mean squared error (MSE) as the loss function to calculate the predicted value. Compared with the true value The difference is used to obtain the predicted loss. The formula is:

[0063] in This is the predicted value at step h. This is the actual value at step h.

[0064] 2. Anomaly Detection Branch of Anomaly Transformer: This branch learns the association patterns of normal time series through the Anomaly-Attention dual-branch mechanism, and identifies anomalies by comparing reconstruction errors with association differences. The specific implementation is as follows: (1) Anomaly-Attention mechanism: Construct a two-branch attention structure, modeling the prior association branch and the sequence association branch separately: Prior association branch: Modeling the priors of the neighbor set of time series data using a learnable Gaussian kernel function, the formula is:

[0065] in: For the prior correlation matrix, This represents the prior association weight between the i-th time series point and the j-th time series point; The learnable scale parameters for the i-th time point (derived from the linear layer) generate, ; To perform normalization operations, ensure .

[0066] Sequence association branch: The true association between time-series subsequences is calculated using a self-attention mechanism, with the following formula:

[0067] in: For sequence incidence matrix, This represents the actual association weight between the i-th time series point and the j-th time series point; (This is the attention weight matrix).

[0068] (2) Feature reconstruction and difference calculation Feature Reconstruction: Based on Sequence Association Matrix AND-value matrix ), generate feature reconstruction values The formula is:

[0069] Association difference calculation: quantifying prior association distribution using symmetric KL divergence. Distribution associated with sequence The difference is used to obtain the correlation difference value. The formula is:

[0070] in Let KL divergence be the KL divergence. , They are respectively , The distribution vector of the i-th row.

[0071] (3) Calculation of abnormal losses Anomaly detection loss is calculated using a combined loss function. The normal pattern learning is enhanced by reconstructing errors, and the anomaly discrimination is amplified by associating differences. The formula is as follows:

[0072] in: The square of the Frobenius norm represents the feature reconstruction error; The L1 norm represents the global correlation difference; This is a balancing coefficient, with a value of 3, used to adjust the weight ratio of the two losses.

[0073] 3. Joint optimization training: The loss values ​​of the two branches are fused using a joint loss function to construct a global optimization objective, as shown in the formula:

[0074] in (Predicted loss weights) (Anomaly detection loss weights) achieve dual-task collaborative optimization by synchronously updating the shared Transformer encoder parameters, the trainable parameters of the TIME-LLM branch (block reprogramming layer, output projection layer), and all parameters of the Anomaly Transformer branch through backpropagation.

[0075] Furthermore, after training, the model performs inference operations based on user query requests and outputs integrated network optimization analysis results. The specific process and examples are as follows: 1. Two-branch independent reasoning: (1) Temporal prediction reasoning: The global temporal features extracted by shared features are processed by the TIME-LLM branch and then restored to the original data scale by RevIN inverse transformation. The network optimization index prediction results for the next H steps (H=7 days in this example) are output, including the specific numerical range, trend slope and confidence interval of each index (confidence level is set to 95%).

[0076] Example: For a user query "Average speed of 4G network in Gannan Prefecture on December 22, 2025 and prediction for the next 7 days", the time series prediction inference output result is as follows: Figure 5 As shown. (2) Anomaly detection inference: Based on the correlation difference value AssDis calculated by the Anomaly Transformer branch, the anomaly threshold is determined by a combination of "statistical method + validation set fine-tuning" (the initial threshold is set to the mean of AssDis of the training data + 3 times the standard deviation, and the optimal threshold after validation set adjustment in this example is 0.35); when the AssDis of the inference data exceeds the threshold, it is judged as an anomaly, and the abnormal time period, abnormal index type and abnormality degree quantification value are output.

[0077] Example: Perform anomaly detection and inference on the average speed data of the 4G network in Gannan Prefecture from November 22, 2025 to December 21, 2025 (the past 30 days). The output result is as follows: Figure 6 As shown.

[0078] 2. Results fusion processing: Based on the timestamp, the prediction sequence for the next 7 steps is spatiotemporally aligned with the anomaly detection results of the past 30 days to establish a correlation mapping between "historical anomalies and future trends". If the future prediction period and the historical anomaly period period overlap periodically (such as the same time of the week), the confidence interval of the prediction value is corrected to improve the prediction accuracy of high-risk periods.

[0079] Example: Based on spatiotemporal alignment analysis, the historical anomalous period (December 10, 2025, 08:00-10:00) falls within the Wednesday morning rush hour. The period from 08:00 to 10:00 on December 17, 2025 (Wednesday) falls within the same period in the next 7 days, indicating a potential anomaly risk. Therefore, the confidence interval for the corresponding predicted date (December 17, 2025) is adjusted. The fused result is as follows. Figure 7 As shown.

[0080] 3. Integrated feedback output: Integrate the above prediction results, anomaly detection results, and fusion analysis conclusions to generate a structured network optimization analysis report, which is fed back to users / network optimization personnel in natural language, while retaining structured data for further analysis.

[0081] Example feedback results: "The average speed of the 4G network in Gannan Prefecture on December 22, 2025 was 32.8 Mbps (historical query results). The predicted average speed for the next 7 days (December 23 to December 29, 2025) is expected to remain between 32.7 and 33.1 Mbps, showing a slight fluctuation trend. Data monitoring over the past 30 days shows an abnormal drop in average speed (30% lower than the normal average) between 08:00 and 10:00 on December 10, 2025. This period coincides with Wednesday morning peak hours. There is a potential risk of anomalies during the same time period on December 27, 2025 (Wednesday). It is recommended to check the base station load and signal coverage in the corresponding area in advance and prepare network optimization plans. Structured data has been exported and supports subsequent in-depth analysis." This application provides an intelligent network status query method. First, it determines the user's query intent and first key information based on the user's initial query statement. Then, if the first key information does not meet the query requirements, it generates guiding information based on the cascading relationships between the first key information, guiding the user to input correct second key information to retrieve related data from the database. Finally, it generates the corresponding query results based on the related data. This guided approach helps users provide accurate key information promptly and effectively solves the inefficiency problem caused by users' unfamiliarity with query parameters.

[0082] like Figure 8 As shown, in a second aspect, embodiments of this application also provide a network status intelligent query device, comprising: The intent analysis module 1 is used to receive a first query statement input by the user, and determine the user's query intent and the first key information corresponding to the query intent based on the first query statement; The guidance information generation module 2 is used to generate guidance information and send it to the user if the first key information does not meet the preset query requirements, based on the cascading relationship between the first key information. The associated data determination module 3 is used to determine the second key information sent by the user based on the guidance information, and to obtain associated data related to the query intent from the network status database based on the second key information; The query result generation module 4 is used to generate query results based on the associated data and the query intent, and send the query results to the user.

[0083] like Figure 9 As shown, in a third aspect, embodiments of this application provide a computer device, including: At least one processor 91; and The memory 92 stores computer instructions 93 that can be executed on the processor. When the processor 91 executes the instructions 93, it implements the steps of any one of the network status intelligent query methods in the above embodiments.

[0084] like Figure 10 As shown, in a fourth aspect, embodiments of this application provide a computer-readable storage medium 101, which stores a computer program 102. When the computer program 102 is executed by a processor, it implements the steps of any one of the network status intelligent query methods described in the above embodiments.

[0085] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0086] Computer instructions include computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. Computer-readable media can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in computer-readable media can be appropriately added to or removed according to the requirements of legislation and patent practice in a jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0087] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments in this specification are not limited to the described order of actions, because according to the embodiments in this specification, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments in this specification.

[0088] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0089] The preferred embodiments disclosed above are merely illustrative of this specification. The optional embodiments do not exhaustively describe all details, nor do they limit the invention to specific implementations. Clearly, many modifications and variations can be made based on the embodiments described in this specification. These embodiments are selected and specifically described in this specification to better explain the principles and practical applications of the embodiments, thereby enabling those skilled in the art to better understand and utilize this specification. This specification is limited only by the claims and their full scope and equivalents.

Claims

1. A method for intelligent network status query, characterized in that, include: Receive a first query statement input by the user, and determine the user's query intent and the first key information corresponding to the query intent based on the first query statement; If the first key information does not meet the preset query requirements, guidance information is generated and sent to the user based on the cascading relationship between the first key information. Determine the second key information sent by the user based on the guidance information, and obtain related data related to the query intent from the network status database based on the second key information; Based on the associated data and the query intent, a query result is generated and sent to the user.

2. The method according to claim 1, characterized in that, The step of receiving a first query statement input by the user and determining the user's query intent and the first key information corresponding to the query intent based on the first query statement includes: The first query statement is input into the first model, and the query intent and the first key information corresponding to the query intent are obtained through the first model.

3. The method according to claim 2, characterized in that, The step of inputting the first query statement into the first model and obtaining the query intent and the first key information corresponding to the query intent through the first model includes: The first query statement is input into the first network structure and the second network structure of the first model, respectively; The query intent and the first key information are obtained by merging the outputs of the first network structure and the second network structure in a predetermined manner. The second network structure is trained based on the output of the first network structure and the corresponding training data.

4. The method according to claim 1, characterized in that, If the first key information does not meet the preset query requirements, generating guidance information based on the cascading relationship between the first key information and sending it to the user further includes: If the first key information does not meet the completeness requirement of the query intent, determine the descriptive information related to the first key information; Based on the description information and the first key information, guide information is generated and sent to the user; The step of determining the second key information sent by the user based on the guidance information, and obtaining related data related to the query intent from the network status database based on the second key information, includes: Determine the second key information confirmed by the user based on the guidance information; If the second key information satisfies the integrity requirement of the query intent, a second query statement is generated based on the second key information; Based on the second query statement, retrieve the associated data related to the query intent from the network state database.

5. The method according to claim 4, characterized in that, If the second key information satisfies the integrity requirement of the query intent, a second query statement is generated based on the second key information, including: Determine the format requirements corresponding to the second key information and the logical requirements between the second key information; The second key information shall be corrected according to the format requirements and / or the logical requirements; A second query statement is generated based on the corrected second key information.

6. The method according to claim 1, characterized in that, The step of generating query results based on the associated data and the query intent, and sending the query results to the user, includes: The associated data and the query intent are input into the second model, and the second model generates the query results based on the associated data and the query intent.

7. The method according to claim 6, characterized in that, Also includes: Determine the verification data corresponding to the query results; The second training data is composed of the query results and the verification results; The second model is retrained based on the second training data within a predetermined time period.

8. A network status intelligent query device, characterized in that, include: The intent analysis module is used to receive a first query statement input by the user, and determine the user's query intent and the first key information corresponding to the query intent based on the first query statement; The guidance information generation module is used to generate guidance information and send it to the user if the first key information does not meet the preset query requirements, based on the cascading relationship between the first key information. The associated data determination module is used to determine the second key information sent by the user based on the guidance information, and to obtain associated data related to the query intent from the network status database based on the second key information; The query result generation module is used to generate query results based on the associated data and the query intent, and send the query results to the user.

9. A computer device, characterized in that, include: At least one processor; as well as A memory storing computer instructions executable on the processor, which, when executed by the processor, implement the steps of the method according to any one of claims 1-7.

10. A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method according to any one of claims 1-7.

Images