5g wireless public network data completion method and device based on large model, equipment and medium

Abstract

The application discloses a 5G wireless public network data completion method and device based on a large model, equipment and medium, belonging to the field of big data quality management and artificial intelligence, the method is: through extracting the time sequence and numerical sequence of 5G wireless public network data to identify the time sequence data missing; according to the content feature library and the traversal matching result of 5G wireless public network data, the business attribute data missing is identified; the time domain feature and frequency domain feature of 5G wireless public network data are extracted, and the weighted fusion result of time domain feature and frequency domain feature is corrected, and the first completion value is obtained; the associated subgraph of missing data is extracted in the business logic knowledge graph, and the associated subgraph is used to generate a prompt word, the prompt word is input into a large language model for neural symbol reasoning, and a second completion value is obtained; the first completion value or the second completion value is used for backfilling the 5G wireless public network data. Therefore, by implementing the application, the accuracy of 5G wireless public network data completion can be improved.

Description

Technical Field

[0001] This invention relates to the fields of big data quality management and artificial intelligence, and in particular to a method, apparatus, device and medium for 5G wireless public network data completion based on a large model. Background Technology

[0002] With the digital transformation of the energy internet and new power systems, power communication networks are rapidly evolving towards a ubiquitous Internet of Things (IoT) represented by 5G wireless public networks. Power IoT terminals transmit critical business data in real time via these wireless public networks. However, due to limitations such as multipath fading in wireless channels, base station handover interference, terminal hardware failures, and complex network environments, data loss, missing fields, or logical breaks inevitably occur during power wireless public network transmission. The integrity of this data directly affects the accuracy of power service billing, the real-time nature of status monitoring, and the security of power grid dispatching.

[0003] Existing data completion and repair technologies typically employ statistical methods (such as mean imputation and linear interpolation) or traditional deep learning methods (such as RNNs and LSTMs). However, when facing data scenarios like public power grid wireless networks, which are characterized by high concurrency, strong business correlations, and multimodal features, existing technologies suffer from the following significant drawbacks: First, for the restoration of time-series data, most existing methods only focus on the numerical statistical patterns of the data itself, lacking the ability to integrate and perceive the inertia of physical motion and the semantics of the external environment. Traditional interpolation algorithms struggle to handle continuous missing data of variable length and often fail when faced with non-periodic abrupt changes caused by the environment. Second, for the missing discrete business attribute data, existing governance methods mainly rely on manually written rule bases or simple dictionary matching. This approach lacks deep logical reasoning and structured association analysis capabilities. In actual business operations, there are complex graph constraints among attribute data, and traditional methods cannot handle structural missing data. Furthermore, when faced with original wide tables with fuzzy headers or no metadata, traditional methods rely on keyword matching, which is prone to omissions or misjudgments. Summary of the Invention

[0004] This invention provides a method, apparatus, device, and medium for 5G wireless public network data completion based on a large model, which can improve the accuracy of 5G wireless public network data completion.

[0005] This invention provides a method for 5G wireless public network data completion based on a large model, including: Extract the time series and numerical series of 5G wireless public network data, and perform time series data missing identification on the 5G wireless public network data based on the time series and the numerical series to obtain a first identification result; The 5G wireless public network data is traversed and matched according to a preset content feature library, and the missing service attribute data of the 5G wireless public network data is identified based on the matching results to obtain a second identification result. If the first identification result indicates that there is missing time-series data, then the time-domain features and frequency-domain features of the 5G wireless public network data are extracted, and the weighted fusion result of the time-domain features and frequency-domain features is semantically corrected to obtain the first completion value; If the second identification result indicates that there is missing business attribute data, then the association subgraph of the missing data is extracted from the pre-constructed business logic knowledge graph, and a prompt word is generated based on the association subgraph. The prompt word is then input into the pre-trained large language model for neural symbol reasoning to obtain the second completion value. When the first or second completion value exists, the 5G wireless public network data is backfilled with completion values ​​to obtain the completion result.

[0006] This invention, by separating and identifying missing time-series data and missing business attribute data, can classify and handle two different types of missing data problems, avoiding poor completion results caused by using a single method. By extracting and fusing time-domain and frequency-domain features and then performing semantic correction, it can comprehensively utilize local change trends, global periodic patterns, and external environmental semantics to improve the accuracy of time-series data completion. By extracting relational subgraphs from the business logic knowledge graph and using a large language model for reasoning, it can leverage the structured relational constraints between known business entity data to logically complete missing attributes, improving the accuracy of business attribute data completion. Compared to existing technologies that are prone to failure and omissions when processing 5G wireless public network data completion, this application can improve the accuracy of 5G wireless public network data completion.

[0007] Further, the step of extracting the time series and numerical series of 5G wireless public network data, and performing time-series data missing identification on the 5G wireless public network data based on the time series and the numerical series to obtain a first identification result, includes: Extract the time series and numerical series corresponding to each data sequence in the 5G wireless public network data, and calculate the mutual information entropy between each set of time series and numerical series. If there is a mutual information entropy greater than a preset random noise threshold, the first identification result is that there is missing time series data, and the data sequence corresponding to the mutual information entropy is marked as a missing time series data sequence. The normalized variance of each numerical sequence is calculated. If there is a normalized variance in a preset high fluctuation range, the first identification result is that there is missing time series data, and the data sequence corresponding to the normalized variance is marked as a missing time series data sequence.

[0008] The embodiments of the present invention use mutual information entropy and normalized variance as dual criteria for temporal missing data identification, which can effectively distinguish between real data missing patterns and random noise or static constants.

[0009] Further, the step of traversing and matching the 5G wireless public network data according to a preset content feature library, and identifying missing service attribute data in the 5G wireless public network data based on the matching results to obtain a second identification result includes: The matching confidence between each data sequence in the 5G wireless public network data and the regular expression of each business entity data in the content feature library is calculated respectively. If all matching confidences corresponding to any data sequence are less than or equal to the preset validity threshold, the second identification result is that there is missing business attribute data, and the data sequence is marked as a data sequence with missing business attributes. If there is a matching confidence score greater than the validity threshold, it is determined that the data sequence corresponding to the matching confidence score matches the business entity data. The matching value between each data in the data sequence and the regular expression of the business entity data is calculated according to the indicator function. If there is a matching value of zero, the second identification result is that there is missing business attribute data, and the data sequence is marked as a data sequence with missing business attributes. If there is unmatched business entity data after the traversal matching is completed, the second identification result is that there is missing business attribute data, and all unmatched business entity data is marked as a sequence of missing business attribute data.

[0010] This invention identifies missing business attributes by using regular expression-based rule matching and completeness checks, enabling it to identify various types of missing business attribute data.

[0011] Further, if the first identification result indicates the presence of missing time-series data, then the time-domain and frequency-domain features of the 5G wireless public network data are extracted, and the weighted fusion result of the time-domain and frequency-domain features is semantically corrected to obtain a first completion value, including: Each of the aforementioned time-series missing data sequences is subjected to dynamic masking processing to obtain multiple binary mask vectors; The missing values ​​of each of the time-series missing data sequences are determined based on each of the binary mask vectors, and each of the missing values ​​is replaced with a preset vector to obtain multiple first time-series missing data sequences; Calculate the relative distance between any two time steps in each of the first time-series missing data sequences, encode each relative distance and add it to the corresponding first time-series missing data sequence to obtain multiple second time-series missing data sequences; Differential operations are performed on each of the second time-series missing data sequences to obtain multiple velocity sequences and multiple acceleration sequences. The local mean sequence and local variance sequence of each of the second time-series missing data sequences are calculated through a preset sliding window. Each second time-series missing data sequence is concatenated with the corresponding velocity sequence, acceleration sequence, local mean sequence, and local variance sequence to obtain multiple feature vector sequences. Each feature vector sequence is mapped to multiple time-domain token sequences and input into the Transformer encoder for feature extraction to generate multiple time-domain features. The second time-series missing data sequences are decomposed into low-frequency features, mid-frequency features, and high-frequency features using discrete wavelet transform. The corresponding low-frequency features, mid-frequency features, and high-frequency features are then mapped and concatenated to obtain multiple frequency domain token sequences. These sequences are then input into the Transformer encoder for feature extraction to generate multiple frequency domain features. The corresponding time-domain features and frequency-domain features are weighted and fused to obtain multiple repair sequences. Then, the semantics of each repair sequence are corrected to obtain multiple first completion values.

[0012] The embodiments of the present invention can improve the ability to repair data with variable length and irregular time sequence by introducing dynamic masking, relative position encoding and fusing multi-dimensional time domain features and frequency domain features.

[0013] Furthermore, the semantic correction of each of the repair sequences is performed to obtain multiple first completion values, including: Obtain the time data and geographic location data corresponding to each of the time-series missing data sequences, and retrieve external real-time knowledge bases based on each set of time data and geographic location data to construct environmental information prompt words corresponding to each of the time-series missing data sequences; Each of the environmental information prompts and the corresponding repair sequences are input into a large language model for analysis, generating multiple semantic correction bias sequences. Each repair sequence is added to its corresponding semantic correction bias sequence to obtain multiple first completion values.

[0014] The embodiments of the present invention, by retrieving external real-time knowledge bases and using large language models for semantic correction, enable the temporal completion results to better reflect the dynamic changes of actual scenarios.

[0015] Further, if the second identification result indicates missing business attribute data, then an association subgraph of the missing data is extracted from the pre-constructed business logic knowledge graph, and a prompt word is generated based on the association subgraph. The prompt word is then input into a pre-trained large language model for neural symbol reasoning to obtain a second completion value, including: Using the context attributes of the missing data sequences of each business attribute as query conditions, matching is performed in the business logic knowledge graph to obtain the association subgraphs corresponding to the missing data sequences of each business attribute. A depth-first search strategy is used to traverse the topology of each associated subgraph, and each topology is converted into a text description to obtain multiple prompt words; wherein, the prompt words contain graph position encoding information of each entity node and relation edge of the corresponding associated subgraph; Each of the aforementioned prompt words is input into a pre-trained large language model for neural symbol reasoning to obtain multiple second completion values.

[0016] This invention, through its embodiments, transforms the knowledge graph's associated subgraphs into text descriptions with positional encoding, effectively converting the graph's structured information into input that can be understood by a large language model.

[0017] Furthermore, the step of inputting each of the prompt words into a pre-trained large language model for neural symbolic reasoning yields multiple second completion values, including: Each of the aforementioned prompt words is input into a pre-trained large language model, so that the large language model can parse the graph symbol knowledge in each of the prompt words and perform chain inference in combination with general neural knowledge to generate a candidate entity set corresponding to each of the missing data. The confidence score of each candidate entity in each candidate entity set is calculated using the large language model, and the candidate entity with the highest confidence score in each candidate entity set is used as the second completion value of the corresponding business attribute missing data sequence, thus obtaining multiple second completion values.

[0018] The embodiments of the present invention generate candidate entities and calculate confidence scores through a large language model, which can improve the reliability of the data completion values ​​for missing business attributes.

[0019] Another embodiment of the present invention provides a 5G wireless public network data completion device based on a large model, including: a time-series data missing identification module, a service attribute data missing identification module, a time-series data missing completion module, a service attribute data missing completion module, and a completion value backfilling module; The time-series data missing identification module is used to extract the time series and numerical series of 5G wireless public network data, and to perform time-series data missing identification on the 5G wireless public network data based on the time series and the numerical series to obtain a first identification result; The business attribute data missing identification module is used to traverse and match the 5G wireless public network data according to the preset content feature library, and to identify the missing business attribute data of the 5G wireless public network data based on the matching results, so as to obtain a second identification result. The time-series data missing completion module is used to extract the time-domain features and frequency-domain features of the 5G wireless public network data if the first identification result indicates that there is a time-series data missing, and to perform semantic correction on the weighted fusion result of the time-domain features and frequency-domain features to obtain the first completion value. The business attribute data missing completion module is used to extract the association subgraph of the missing data in the pre-constructed business logic knowledge graph if the second identification result is that there is missing business attribute data, generate prompt words based on the association subgraph, input the prompt words into the pre-trained large language model for neural symbol reasoning, and obtain the second completion value. The completion value backfilling module is used to backfill the 5G wireless public network data with completion values ​​when the first completion value or the second completion value exists, so as to obtain the completion result.

[0020] Another embodiment of the present invention provides a terminal device, including: a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the steps of the 5G wireless public network data completion method based on a large model as described in the present invention.

[0021] Another embodiment of the present invention also provides a computer-readable storage medium item, including: a stored computer program, which, when the computer program is running, controls the device where the computer-readable storage medium is located to perform the steps of a 5G wireless public network data completion method based on a large model as described in the present invention. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating an embodiment of the 5G wireless public network data completion method based on a large model provided by the present invention. Figure 2 A flowchart illustrating an embodiment of the 5G wireless public network data missing type identification method provided by the present invention; Figure 3 A flowchart illustrating an embodiment of the method for generating time-series missing data completion values ​​provided by the present invention; Figure 4 A flowchart illustrating an embodiment of the method for generating missing business attribute data completion values ​​provided by the present invention; Figure 5 This is a flowchart illustrating another embodiment of the 5G wireless public network data completion method based on a large model provided by the present invention. Figure 6 This is a schematic diagram of one embodiment of the 5G wireless public network data completion device based on a large model provided by the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0024] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application; the terms “comprising” and “having”, and any variations thereof, in the specification, claims, and foregoing description of the drawings are intended to cover non-exclusive inclusion.

[0025] In the description of the embodiments of this application, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly defined.

[0026] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0027] In the description of the embodiments in this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.

[0028] In the description of the embodiments of this application, the term "multiple" refers to two or more (including two), similarly, "multiple sets" refers to two or more (including two sets), and "multiple pieces" refers to two or more (including two pieces).

[0029] See Figure 1 To address the issues of failure and omissions in existing technologies when processing 5G wireless public network data completion, an embodiment of the present invention provides a 5G wireless public network data completion method based on a large model, comprising steps S101 to S105: Step S101: Extract the time series and numerical series of 5G wireless public network data, and perform time series data missing identification on the 5G wireless public network data based on the time series and the numerical series to obtain the first identification result.

[0030] It should be noted that extracting the time series and numerical series of 5G wireless public network data, and identifying time-series data missing based on these time series and numerical series to obtain the first identification result means: for the 5G wireless public network data to be analyzed, extracting the corresponding time series and numerical series from each data series respectively. Analyzing and judging the inherent correlation between the time series and numerical series to identify whether there are discontinuities or missing data in the time dimension due to transmission interruptions, equipment failures, or channel interference. This process is accomplished by detecting statistical dependencies and fluctuation pattern changes between sequences, and finally outputting the first identification result of "time-series data missing exists" or "time-series data missing does not exist," providing a criterion for subsequent time-series completion.

[0031] Preferably, the step of extracting the time series and numerical series of 5G wireless public network data, and performing time-series data missing identification on the 5G wireless public network data based on the time series and the numerical series to obtain a first identification result includes: Extract the time series and numerical series corresponding to each data sequence in the 5G wireless public network data, and calculate the mutual information entropy between each set of time series and numerical series. If there is a mutual information entropy greater than a preset random noise threshold, the first identification result is that there is missing time series data, and the data sequence corresponding to the mutual information entropy is marked as a missing time series data sequence. The normalized variance of each numerical sequence is calculated. If there is a normalized variance in a preset high fluctuation range, the first identification result is that there is missing time series data, and the data sequence corresponding to the normalized variance is marked as a missing time series data sequence.

[0032] In one embodiment, the process for identifying time-series missing data is as follows: Figure 2 As shown in S11, let the time series of a data sequence be... The numerical sequence is The time series sum numerical sequence The mutual information entropy between them is ,in Time series The value in Numerical sequence The value in For all data The probability of occurrence For all data The probability of occurrence For all data and The probability of simultaneous occurrence. If the mutual information entropy is greater than the random noise threshold. ,Right now It was confirmed that the data sequence exhibited a strong dependency relationship over time, and the data sequence was determined to be a time-series missing data sequence.

[0033] In the above embodiments, the numerical sequence The normalized variance is ,in Let be the variance. If If the data is within a preset high fluctuation range, it indicates that the data changes dynamically over time rather than being a static value, thus determining that the data sequence is a time-series missing data sequence.

[0034] Step S102: The 5G wireless public network data is traversed and matched according to the preset content feature library, and the missing service attribute data of the 5G wireless public network data is identified based on the matching results to obtain a second identification result.

[0035] It should be noted that the process of traversing and matching the 5G wireless public network data according to a pre-built content feature library, and then identifying missing business attribute data based on the matching results, yields the second identification result. This means: a content feature library is pre-built, which stores standard format rules for various business entity data, typically represented by regular expressions or structured templates. Each data sequence in the 5G wireless public network data is traversed and matched against all rules in the pre-built content feature library. The business category is determined based on the degree of matching, and the format integrity is further verified. If a field cannot match any rule at all, has a low matching confidence, or matches a category but the specific value does not conform to the format specification, then it is determined that there is missing business attribute data at that location. In addition, after the traversal is completed, it is also necessary to check whether all expected business attributes have corresponding fields in the data, thereby identifying cases where an entire column is missing. Finally, this process outputs a second identification result of "missing business attribute data" or "missing business attribute data," providing a basis for subsequent completion based on business logic.

[0036] Preferably, the step of traversing and matching the 5G wireless public network data according to a preset content feature library, and identifying missing service attribute data in the 5G wireless public network data based on the matching results to obtain a second identification result includes: The matching confidence between each data sequence in the 5G wireless public network data and the regular expression of each business entity data in the content feature library is calculated respectively. If all matching confidences corresponding to any data sequence are less than or equal to the preset validity threshold, the second identification result is that there is missing business attribute data, and the data sequence is marked as a data sequence with missing business attributes. If there is a matching confidence score greater than the validity threshold, it is determined that the data sequence corresponding to the matching confidence score matches the business entity data. The matching value between each data in the data sequence and the regular expression of the business entity data is calculated according to the indicator function. If there is a matching value of zero, the second identification result is that there is missing business attribute data, and the data sequence is marked as a data sequence with missing business attributes. If there is unmatched business entity data after the traversal matching is completed, the second identification result is that there is missing business attribute data, and all unmatched business entity data is marked as a sequence of missing business attribute data.

[0037] In one embodiment, the process for identifying missing business attribute data is as follows: Figure 2 As shown in S12, suppose the pre-built content feature library contains A set of regular expression validation rules for various business entity data This includes specifications such as traffic record list format and IP address format. The content feature library is used to analyze a data sequence. Perform traversal matching, wherein the data sequence The amount of data is Define indicator functions When data Conforms to regular expression The value is 1 if it is true, and 0 otherwise.

[0038] In the above embodiments, the data sequence is calculated. No. Matching confidence of business entity data ,when When the value exceeds the preset validity threshold, it is determined to be the first... Business entity data. If the data sequence... If the data does not belong to any business entity in the content feature library, then the data sequence is determined. This is a data sequence with missing business attributes.

[0039] In the above embodiments, if the data sequence Determined as the first For business entity data, regular expressions are used. For the data sequence Perform validation for any data If the calculation result Then determine The data sequence does not conform to the format specifications and is therefore deemed invalid. This is a data sequence with missing business attributes.

[0040] In the above embodiments, after traversing and processing all data sequences, a completeness check is performed to record the business entity data that should exist in the content feature library but has not been matched, and all unmatched business entity data is determined to be a data sequence with missing business attributes.

[0041] Step S103: If the first identification result indicates that there is missing time-series data, then extract the time-domain features and frequency-domain features of the 5G wireless public network data, and perform semantic correction on the weighted fusion result of the time-domain features and frequency-domain features to obtain the first completion value.

[0042] It should be noted that if the first identification result indicates the presence of missing time-series data, then the time-domain and frequency-domain features of the 5G wireless public network data are extracted, and the weighted fusion result of the time-domain and frequency-domain features is semantically corrected to obtain the first completion value. This means that when the first identification result determines that there is missing time-series data in the 5G wireless public network data, multi-dimensional feature extraction is performed on the missing time-series data sequence: on the one hand, local statistical features, trend features, and relative time distance codes of the sequence are extracted from the time domain perspective; on the other hand, the sequence is decomposed into components of different frequencies through frequency domain transformation to capture its long-term periodicity, short-term fluctuations, and noise features. The extracted time-domain features and frequency-domain features are fused and input into the prediction model to generate a preliminary completion sequence. To further improve the rationality and contextual consistency of the completion result, it is also necessary to combine the time, geographical location, and other information corresponding to the missing time-series data sequence, search external knowledge bases, or use large language models to semantically correct the preliminary completion result. The final output is a "first completion value" that conforms to both numerical change patterns and the actual semantic environment.

[0043] Preferably, if the first identification result indicates missing time-series data, then the time-domain and frequency-domain features of the 5G wireless public network data are extracted, and the weighted fusion result of the time-domain and frequency-domain features is semantically corrected to obtain a first completion value, including: Each of the aforementioned time-series missing data sequences is subjected to dynamic masking processing to obtain multiple binary mask vectors; The missing values ​​of each of the time-series missing data sequences are determined based on each of the binary mask vectors, and each of the missing values ​​is replaced with a preset vector to obtain multiple first time-series missing data sequences; Calculate the relative distance between any two time steps in each of the first time-series missing data sequences, encode each relative distance and add it to the corresponding first time-series missing data sequence to obtain multiple second time-series missing data sequences; Differential operations are performed on each of the second time-series missing data sequences to obtain multiple velocity sequences and multiple acceleration sequences. The local mean sequence and local variance sequence of each of the second time-series missing data sequences are calculated through a preset sliding window. Each second time-series missing data sequence is concatenated with the corresponding velocity sequence, acceleration sequence, local mean sequence, and local variance sequence to obtain multiple feature vector sequences. Each feature vector sequence is mapped to multiple time-domain token sequences and input into the Transformer encoder for feature extraction to generate multiple time-domain features. The second time-series missing data sequences are decomposed into low-frequency features, mid-frequency features, and high-frequency features using discrete wavelet transform. The corresponding low-frequency features, mid-frequency features, and high-frequency features are then mapped and concatenated to obtain multiple frequency domain token sequences. These sequences are then input into the Transformer encoder for feature extraction to generate multiple frequency domain features. The corresponding time-domain features and frequency-domain features are weighted and fused to obtain multiple repair sequences. Then, the semantics of each repair sequence are corrected to obtain multiple first completion values.

[0044] In one embodiment, the process for generating imputation values ​​for time-series missing data is as follows: Figure 3 As shown, a time-frequency dual-stream Transformer network capable of adaptively handling variable-length and discontinuous missing data sequences is constructed for time-series missing data sequences. To enable the model to identify and handle variable missing amounts, dynamic masking is first applied to each time-series missing data sequence to generate a binary mask vector of the same length as each time-series missing data sequence, where a mask of 0 indicates missing data and a mask of 1 indicates valid data.

[0045] In the above embodiments, a globally shared special vector that can be updated via backpropagation is defined. For each missing point in a time-series missing data sequence, its original value is replaced with... For valid points, their original values ​​are retained, resulting in multiple temporally missing data sequences. To prevent long missing segments from causing the model to lose its sense of temporal distance, relative position encoding is introduced. The relative distance between any two time steps is calculated, and its encoding is added to the first temporally missing data sequence to obtain the second temporally missing data sequence.

[0046] In the above embodiments, the physical kinematic laws and statistical distribution laws of the data are fused and encoded, including: encoding the second time-series missing data sequence. Perform difference operations to calculate the velocity sequence. With acceleration sequence Capture the rate of change and trend of data; set a length of A sliding window is used to calculate the local mean sequence for each time step along the time axis. With local variance sequence .

[0047] In the above embodiments, the original numerical values, kinematic features, and statistical features are concatenated to form a feature vector sequence. A linear layer is used to map this vector into a high-dimensional vector, generating a time-domain token sequence. .Will Input a Transformer encoder and learn temporal features using a self-attention mechanism: .

[0048] In the above embodiment, the second time-series missing data sequence is decomposed into three features: low frequency, mid frequency, and high frequency using discrete wavelet transform. These three features are then mapped into high-dimensional features and concatenated to form a frequency domain token sequence. During the subsequent splicing process, high-frequency dimensions are compressed. Inputting a Transformer encoder, the frequency domain features are learned using a self-attention mechanism: .

[0049] In the above embodiments, the features output by the time-domain and frequency-domain Transformers are... and The process involves row-weighted fusion and then connecting it to a linear layer to output an initial repair sequence containing complete temporal information. .

[0050] Preferably, the semantic correction of each of the repair sequences to obtain multiple first completion values ​​includes: Obtain the time data and geographic location data corresponding to each of the time-series missing data sequences, and retrieve external real-time knowledge bases based on each set of time data and geographic location data to construct environmental information prompt words corresponding to each of the time-series missing data sequences; Each of the environmental information prompts and the corresponding repair sequences are input into a large language model for analysis, generating multiple semantic correction bias sequences. Each repair sequence is added to its corresponding semantic correction bias sequence to obtain multiple first completion values.

[0051] In the above embodiments, the reasoning ability of a large language model for external environmental information is utilized to... Optimization and correction were performed. Time and geographic location data corresponding to the missing time-series data sequences were obtained, external real-time knowledge bases were retrieved, and prompt words containing environmental information were constructed, including the current date attribute and the current emergency event.

[0052] In the above embodiments, environmental information prompts are combined with positionally encoded words with completion. Input a large language model, analyze the impact of environmental context on time series, and output a semantically corrected bias sequence. The inpainted value generated by the neural network is superimposed with the semantic correction bias generated by the large model to obtain the final inpainted value: .

[0053] Step S104: If the second identification result indicates that there is missing business attribute data, then extract the association subgraph of the missing data from the pre-constructed business logic knowledge graph, generate prompt words based on the association subgraph, and input the prompt words into the pre-trained large language model for neural symbol reasoning to obtain the second completion value.

[0054] It should be noted that if the second identification result indicates missing business attribute data, a related subgraph of the missing data is extracted from the pre-constructed business logic knowledge graph. Based on this subgraph, a prompt word is generated, and the prompt word is input into a pre-trained large language model for neural symbolic reasoning. The resulting second completion value refers to the following: When the second identification result determines that the current data has missing business attributes, the pre-constructed business logic knowledge graph, which structurally stores the business rules and related constraints of the 5G wireless public network domain in the form of entities and relationships, is used. Based on the contextual attributes of the record containing the missing data, a search and matching process is performed in the knowledge graph to extract a local related subgraph containing relevant entities and their relationships. The topology of this subgraph is transformed into a text description with location encoding information, forming a structured prompt word. This prompt word simultaneously contains symbolic knowledge and natural language description from the graph, so that it can be input into the pre-trained large language model. The large language model performs neural symbolic reasoning based on prompt words. That is, it combines its internal general semantic understanding capabilities with externally injected graph-based business rules, and generates a set of candidate entities that match the missing attributes through logical chain deduction, and outputs the confidence score of each candidate. Finally, the entity value with the highest confidence and that meets the graph constraints is selected as the "second completion value".

[0055] Preferably, if the second identification result indicates missing business attribute data, then an association subgraph of the missing data is extracted from the pre-constructed business logic knowledge graph, and a prompt word is generated based on the association subgraph. The prompt word is then input into a pre-trained large language model for neural symbol reasoning to obtain a second completion value, including: Using the context attributes of the missing data sequences of each business attribute as query conditions, matching is performed in the business logic knowledge graph to obtain the association subgraphs corresponding to the missing data sequences of each business attribute. A depth-first search strategy is used to traverse the topology of each associated subgraph, and each topology is converted into a text description to obtain multiple prompt words; wherein, the prompt words contain graph position encoding information of each entity node and relation edge of the corresponding associated subgraph; Each of the aforementioned prompt words is input into a pre-trained large language model for neural symbol reasoning to obtain multiple second completion values.

[0056] In one embodiment, the process for generating complete values ​​for missing business attribute data is as follows: Figure 4 As shown, for data sequences lacking business attributes, the entire process of wireless public network services is first analyzed to construct a business logic knowledge graph. Graph nodes include device list IP addresses, APNs, operator identifiers, and business department organizational trees. Graph edges represent deep-level relationships between different services and resources, such as the association between "APN" and "IP address." The business attribute data is then transformed into a standard triple format of "head entity, relation, tail entity" and stored in a graph database to form the business logic knowledge graph.

[0057] In the above embodiments, the contextual attributes of the missing business attribute data sequence are used as the basis for querying. The existing attribute set in the missing record is mapped to the query vector, and matching is performed in the business logic knowledge graph to extract the local association subgraph containing attribute association, hierarchical affiliation and physical binding.

[0058] In the above embodiments, a depth-first search strategy is used to traverse the topology of the local relational subgraph, converting the graph structure into a text description according to a predefined method. During the conversion process, graph position encoding information is introduced for each entity node and relation edge, and the distance between the node and the central anchor point is marked, thereby generating structured prompt words containing topological structure information. .

[0059] Preferably, the step of inputting each of the prompt words into a pre-trained large language model for neural symbolic reasoning to obtain multiple second completion values ​​includes: Each of the aforementioned prompt words is input into a pre-trained large language model, so that the large language model can parse the graph symbol knowledge in each of the prompt words and perform chain inference in combination with general neural knowledge to generate a candidate entity set corresponding to each of the missing data. The confidence score of each candidate entity in each candidate entity set is calculated using the large language model, and the candidate entity with the highest confidence score in each candidate entity set is used as the second completion value of the corresponding business attribute missing data sequence, thus obtaining multiple second completion values.

[0060] In the above embodiments, structured prompt words will be used. Input a pre-trained large language model and guide it to make predictions using a neural symbolic reasoning paradigm. The model parses the spectrogram symbolic knowledge in the prompt words and combines it with the model's inherent general neural knowledge to execute chained reasoning logic, generating a set of N candidate entities for the missing field. At the same time, the model outputs each candidate entity. conditional probability distribution The confidence score of the candidate value is used as the final business attribute completion value.

[0061] Step S105: When the first or second completion value exists, the 5G wireless public network data is backfilled with completion value to obtain the completion result.

[0062] It should be noted that when the first or second complete value exists, the complete value backfilling for the 5G wireless public network data means that after identifying and completing the missing time-series data or service attribute data, if a corresponding first or second complete value is generated, the generated complete value is accurately backfilled into the original 5G wireless public network data according to its corresponding missing position. Finally, a complete result that has completed missing data repair, is data-complete, and conforms to business logic is output.

[0063] This invention, by separating and identifying missing time-series data and missing business attribute data, can classify and handle two different types of missing data problems, avoiding poor completion results caused by using a single method. By extracting and fusing time-domain and frequency-domain features and then performing semantic correction, it can comprehensively utilize local change trends, global periodic patterns, and external environmental semantics to improve the accuracy of time-series data completion. By extracting relational subgraphs from the business logic knowledge graph and using a large language model for reasoning, it can leverage the structured relational constraints between known business entity data to logically complete missing attributes, improving the accuracy of business attribute data completion. Compared to existing technologies that are prone to failure and omissions when processing 5G wireless public network data completion, this application can improve the accuracy of 5G wireless public network data completion.

[0064] Optionally, in this embodiment of the invention, the step of extracting the time series and numerical series of 5G wireless public network data, and performing time-series data missing identification on the 5G wireless public network data based on the time series and the numerical series to obtain a first identification result, includes: Extract the time series and numerical series corresponding to each data sequence in the 5G wireless public network data, and calculate the mutual information entropy between each set of time series and numerical series. If there is a mutual information entropy greater than a preset random noise threshold, the first identification result is that there is missing time series data, and the data sequence corresponding to the mutual information entropy is marked as a missing time series data sequence. The normalized variance of each numerical sequence is calculated. If there is a normalized variance in a preset high fluctuation range, the first identification result is that there is missing time series data, and the data sequence corresponding to the normalized variance is marked as a missing time series data sequence.

[0065] The embodiments of the present invention use mutual information entropy and normalized variance as dual criteria for temporal missing data identification, which can effectively distinguish between real data missing patterns and random noise or static constants.

[0066] Optionally, in this embodiment of the invention, the step of traversing and matching the 5G wireless public network data according to a preset content feature library, and identifying missing service attribute data in the 5G wireless public network data based on the matching results to obtain a second identification result includes: The matching confidence between each data sequence in the 5G wireless public network data and the regular expression of each business entity data in the content feature library is calculated respectively. If all matching confidences corresponding to any data sequence are less than or equal to the preset validity threshold, the second identification result is that there is missing business attribute data, and the data sequence is marked as a data sequence with missing business attributes. If there is a matching confidence score greater than the validity threshold, it is determined that the data sequence corresponding to the matching confidence score matches the business entity data. The matching value between each data in the data sequence and the regular expression of the business entity data is calculated according to the indicator function. If there is a matching value of zero, the second identification result is that there is missing business attribute data, and the data sequence is marked as a data sequence with missing business attributes. If there is unmatched business entity data after the traversal matching is completed, the second identification result is that there is missing business attribute data, and all unmatched business entity data is marked as a sequence of missing business attribute data.

[0067] This invention identifies missing business attributes by using regular expression-based rule matching and completeness checks, enabling it to identify various types of missing business attribute data.

[0068] Optionally, in this embodiment of the invention, if the first identification result indicates the presence of missing time-series data, then the time-domain and frequency-domain features of the 5G wireless public network data are extracted, and the weighted fusion result of the time-domain and frequency-domain features is semantically corrected to obtain a first completion value, including: Each of the aforementioned time-series missing data sequences is subjected to dynamic masking processing to obtain multiple binary mask vectors; The missing values ​​of each of the time-series missing data sequences are determined based on each of the binary mask vectors, and each of the missing values ​​is replaced with a preset vector to obtain multiple first time-series missing data sequences; Calculate the relative distance between any two time steps in each of the first time-series missing data sequences, encode each relative distance and add it to the corresponding first time-series missing data sequence to obtain multiple second time-series missing data sequences; Differential operations are performed on each of the second time-series missing data sequences to obtain multiple velocity sequences and multiple acceleration sequences. The local mean sequence and local variance sequence of each of the second time-series missing data sequences are calculated through a preset sliding window. Each second time-series missing data sequence is concatenated with the corresponding velocity sequence, acceleration sequence, local mean sequence, and local variance sequence to obtain multiple feature vector sequences. Each feature vector sequence is mapped to multiple time-domain token sequences and input into the Transformer encoder for feature extraction to generate multiple time-domain features. The second time-series missing data sequences are decomposed into low-frequency features, mid-frequency features, and high-frequency features using discrete wavelet transform. The corresponding low-frequency features, mid-frequency features, and high-frequency features are then mapped and concatenated to obtain multiple frequency domain token sequences. These sequences are then input into the Transformer encoder for feature extraction to generate multiple frequency domain features. The corresponding time-domain features and frequency-domain features are weighted and fused to obtain multiple repair sequences. Then, the semantics of each repair sequence are corrected to obtain multiple first completion values.

[0069] The embodiments of the present invention can improve the ability to repair data with variable length and irregular time sequence by introducing dynamic masking, relative position encoding and fusing multi-dimensional time domain features and frequency domain features.

[0070] Optionally, in this embodiment of the invention, the semantic correction of each of the repair sequences to obtain multiple first completion values ​​includes: Obtain the time data and geographic location data corresponding to each of the time-series missing data sequences, and retrieve external real-time knowledge bases based on each set of time data and geographic location data to construct environmental information prompt words corresponding to each of the time-series missing data sequences; Each of the environmental information prompts and the corresponding repair sequences are input into a large language model for analysis, generating multiple semantic correction bias sequences. Each repair sequence is added to its corresponding semantic correction bias sequence to obtain multiple first completion values.

[0071] The embodiments of the present invention, by retrieving external real-time knowledge bases and using large language models for semantic correction, enable the temporal completion results to better reflect the dynamic changes of actual scenarios.

[0072] Optionally, in this embodiment of the invention, if the second identification result indicates missing business attribute data, then an association subgraph of the missing data is extracted from a pre-constructed business logic knowledge graph, and a prompt word is generated based on the association subgraph. The prompt word is then input into a pre-trained large language model for neural symbol reasoning to obtain a second completion value, including: Using the context attributes of the missing data sequences of each business attribute as query conditions, matching is performed in the business logic knowledge graph to obtain the association subgraphs corresponding to the missing data sequences of each business attribute. A depth-first search strategy is used to traverse the topology of each associated subgraph, and each topology is converted into a text description to obtain multiple prompt words; wherein, the prompt words contain graph position encoding information of each entity node and relation edge of the corresponding associated subgraph; Each of the aforementioned prompt words is input into a pre-trained large language model for neural symbol reasoning to obtain multiple second completion values.

[0073] This invention, through its embodiments, transforms the knowledge graph's associated subgraphs into text descriptions with positional encoding, effectively converting the graph's structured information into input that can be understood by a large language model.

[0074] Optionally, in this embodiment of the invention, the step of inputting each of the prompt words into a pre-trained large language model for neural symbol reasoning to obtain multiple second completion values ​​includes: Each of the aforementioned prompt words is input into a pre-trained large language model, so that the large language model can parse the graph symbol knowledge in each of the prompt words and perform chain inference in combination with general neural knowledge to generate a candidate entity set corresponding to each of the missing data. The confidence score of each candidate entity in each candidate entity set is calculated using the large language model, and the candidate entity with the highest confidence score in each candidate entity set is used as the second completion value of the corresponding business attribute missing data sequence, thus obtaining multiple second completion values.

[0075] The embodiments of the present invention generate candidate entities and calculate confidence scores through a large language model, which can improve the reliability of the data completion values ​​for missing business attributes.

[0076] like Figure 5 As shown, based on the above-described method embodiments, another embodiment of the 5G wireless public network data completion method based on a large model is provided, including steps S1 to S5: Step S1: Identify the types of missing data in the original data of 5G wireless public network services using the data missing type identification module; wherein, executing step S1 is equivalent to executing steps S101 and S102.

[0077] Step S2: For time series missing data, generate time series completion values ​​through the time series data intelligent completion module; wherein, executing step S1 is equivalent to executing step S103.

[0078] Step S3: Construct a wireless public network service logic knowledge graph; wherein, the wireless public network service logic knowledge graph is equivalent to the service logic knowledge graph described in step S104.

[0079] Step S4: For missing business attributes, generate a business attribute completion entity through the wireless public network business logic knowledge graph and business attribute data intelligent completion module; wherein, executing step S4 is equivalent to executing step S104.

[0080] Step S5 involves data fusion and backfilling based on the time-series completion value and business attribute completion entity; wherein, executing step S5 is equivalent to executing step S105.

[0081] This invention first proposes a multi-dimensional feature adaptive missing data identification network to locate and identify time-series missing data and business attribute missing data, respectively. For time-series missing data, a large-scale model prediction architecture integrating time-frequency features is proposed to achieve accurate fitting of fluctuation trends and repair of outlier missing values. For business attribute missing data, a wireless public network business logic knowledge graph is constructed. Graph-based prompts guide a large language model to perform chain-like reasoning, uncovering implicit relationships between entities and automatically completing missing business attributes. This invention can achieve intelligent completion and repair of public network business data across the entire network, all operators, and all standards, effectively solving the problem of low efficiency in abnormal data processing using traditional methods.

[0082] like Figure 6 As shown, based on the above method embodiments, corresponding apparatus embodiments are provided; An embodiment of the present invention provides a 5G wireless public network data completion device based on a large model, including: a time-series data missing identification module 601, a service attribute data missing identification module 602, a time-series data missing completion module 603, a service attribute data missing completion module 604, and a completion value backfilling module 605. The time-series data missing identification module 601 is used to extract the time series and numerical series of 5G wireless public network data, and perform time-series data missing identification on the 5G wireless public network data based on the time series and the numerical series to obtain a first identification result; The business attribute data missing identification module 602 is used to traverse and match the 5G wireless public network data according to the preset content feature library, and to identify the missing business attribute data of the 5G wireless public network data based on the matching result, so as to obtain a second identification result. The time-series data missing completion module 603 is used to extract the time-domain features and frequency-domain features of the 5G wireless public network data if the first identification result indicates that there is a time-series data missing, and to perform semantic correction on the weighted fusion result of the time-domain features and frequency-domain features to obtain the first completion value. The business attribute data missing completion module 604 is used to extract the association subgraph of the missing data in the pre-constructed business logic knowledge graph if the second identification result is that there is missing business attribute data, generate prompt words based on the association subgraph, input the prompt words into the pre-trained large language model for neural symbol reasoning, and obtain the second completion value. The completion value backfilling module 605 is used to backfill the 5G wireless public network data with completion values ​​when the first completion value or the second completion value exists, so as to obtain the completion result.

[0083] Optionally, in this embodiment of the invention, the time-series data missing identification module 601 includes: a first time-series data missing identification submodule and a second time-series data missing identification submodule; The first time-series data missing identification submodule is used to extract the time series and numerical series corresponding to each data sequence in the 5G wireless public network data, and calculate the mutual information entropy between each group of time series and numerical series respectively. If there is a mutual information entropy greater than a preset random noise threshold, the first identification result is that there is a time-series data missing, and the data sequence corresponding to the mutual information entropy is marked as a time-series missing data sequence. The second time-series data missing identification submodule is used to calculate the normalized variance of each numerical sequence. If there is a normalized variance in a preset high fluctuation range, the first identification result is that there is time-series data missing, and the data sequence corresponding to the normalized variance is marked as a time-series missing data sequence.

[0084] The embodiments of the present invention use mutual information entropy and normalized variance as dual criteria for temporal missing data identification, which can effectively distinguish between real data missing patterns and random noise or static constants.

[0085] Optionally, in this embodiment of the invention, the business attribute data missing identification module 602 includes: a first business attribute data missing identification submodule, a second business attribute data missing identification submodule, and a third business attribute data missing identification submodule; The first business attribute data missing identification submodule is used to calculate the matching confidence between each data sequence in the 5G wireless public network data and the regular expression of each business entity data in the content feature library. If all matching confidences corresponding to any data sequence are less than or equal to a preset validity threshold, the second identification result is that there is missing business attribute data, and the data sequence is marked as a business attribute missing data sequence. The second business attribute data missing identification submodule is used to determine that the data sequence corresponding to the matching confidence level matches the business entity data if there is a matching confidence level greater than the validity threshold, and calculate the matching value between each data in the data sequence and the regular expression of the business entity data according to the indicator function. If there is a matching value with a value of zero, the second identification result is that there is missing business attribute data, and the data sequence is marked as a business attribute missing data sequence. The third business attribute data missing identification submodule is used to identify that if there is unmatched business entity data after completing the traversal matching, the second identification result is that there is missing business attribute data, and all unmatched business entity data is marked as a business attribute missing data sequence.

[0086] This invention identifies missing business attributes by using regular expression-based rule matching and completeness checks, enabling it to identify various types of missing business attribute data.

[0087] Optionally, in this embodiment of the invention, the time-series data missing completion module 603 includes: a binary mask vector submodule, a first time-series missing data sequence submodule, a second time-series missing data sequence submodule, a time-series feature sequence submodule, a time-domain feature submodule, a frequency-domain feature submodule, and a semantic correction submodule; The binary mask vector submodule is used to perform dynamic masking processing on each of the time-series missing data sequences to obtain multiple binary mask vectors. The first time-series missing data sequence submodule is used to determine the missing value of each time-series missing data sequence based on each of the binary mask vectors, and replace each of the missing values ​​with a preset vector to obtain multiple first time-series missing data sequences; The second time-series missing data sequence submodule is used to calculate the relative distance between any two time steps in each of the first time-series missing data sequences, encode each relative distance and add it to the corresponding first time-series missing data sequence to obtain multiple second time-series missing data sequences; The time-series feature sequence submodule is used to perform difference operations on each of the second time-series missing data sequences to obtain multiple velocity sequences and multiple acceleration sequences, and to calculate the local mean sequence and local variance sequence of each of the second time-series missing data sequences through a preset sliding window. The time-domain feature submodule is used to concatenate each of the second time-series missing data sequences with the corresponding velocity sequence, acceleration sequence, local mean sequence and local variance sequence to obtain multiple feature vector sequences, map each of the feature vector sequences to multiple time-domain token sequences and input them to the Transformer encoder for feature extraction to generate multiple time-domain features. The frequency domain feature submodule is used to decompose each of the second time-series missing data sequences into low-frequency features, mid-frequency features and high-frequency features using discrete wavelet transform, and to perform feature mapping and splicing on the corresponding low-frequency features, mid-frequency features and high-frequency features to obtain multiple frequency domain token sequences, which are then input into the Transformer encoder for feature extraction to generate multiple frequency domain features. The semantic correction submodule is used to perform weighted fusion of the corresponding time-domain features and frequency-domain features to obtain multiple repair sequences, and to perform semantic correction on each repair sequence to obtain multiple first completion values.

[0088] The embodiments of the present invention can improve the ability to repair data with variable length and irregular time sequence by introducing dynamic masking, relative position encoding and fusing multi-dimensional time domain features and frequency domain features.

[0089] Optionally, in this embodiment of the invention, the semantic correction submodule includes: an environmental information cue word unit, a semantic correction bias sequence unit, and a semantic correction unit; The environmental information prompt word unit is used to obtain the time data and geographical location data corresponding to each of the time-series missing data sequences, and to retrieve the external real-time knowledge base based on each set of time data and geographical location data to construct the environmental information prompt words corresponding to each of the time-series missing data sequences. The semantic correction bias sequence unit is used to input each of the environmental information prompt words and the corresponding repair sequence into the large language model for analysis, and generate multiple semantic correction bias sequences. The semantic correction unit is used to add each of the repair sequences to the corresponding semantic correction bias sequence to obtain multiple first completion values.

[0090] The embodiments of the present invention, by retrieving external real-time knowledge bases and using large language models for semantic correction, enable the temporal completion results to better reflect the dynamic changes of actual scenarios.

[0091] Optionally, in this embodiment of the invention, the business attribute data missing completion module 604 includes: an association subgraph submodule, a prompt word submodule, and a neural symbol reasoning submodule; The associated subgraph submodule is used to match the business logic knowledge graph with the context attributes of the missing data sequences of each business attribute as the query conditions, and obtain the associated subgraphs corresponding to the missing data sequences of each business attribute. The prompt word submodule is used to traverse the topology of each of the associated subgraphs using a depth-first search strategy, and convert each of the topology into a text description to obtain multiple prompt words; wherein, the prompt words contain graph position encoding information of each entity node and relation edge of the corresponding associated subgraph; The neural symbol reasoning submodule is used to input each of the prompt words into a pre-trained large language model for neural symbol reasoning to obtain multiple second completion values.

[0092] This invention, through its embodiments, transforms the knowledge graph's associated subgraphs into text descriptions with positional encoding, effectively converting the graph's structured information into input that can be understood by a large language model.

[0093] Optionally, in this embodiment of the invention, the business attribute data missing completion module 604 includes: a candidate entity set submodule and a confidence level submodule; The candidate entity set submodule is used to input each of the prompt words into a pre-trained large language model, so that the large language model can parse the graph symbol knowledge in each of the prompt words and perform chain inference in combination with general neural knowledge to generate a candidate entity set corresponding to each of the missing data. The confidence score module is used to calculate the confidence score of each candidate entity in each candidate entity set through the large language model, and take the candidate entity with the highest confidence score in each candidate entity set as the second completion value of the corresponding business attribute missing data sequence to obtain multiple second completion values.

[0094] The embodiments of the present invention generate candidate entities and calculate confidence scores through a large language model, which can improve the reliability of the data completion values ​​for missing business attributes.

[0095] It is understood that the above-described device embodiments correspond to the method embodiments of the present invention, and can implement the 5G wireless public network data completion method based on a large model provided by any of the above-described method embodiments of the present invention.

[0096] This invention separates and identifies missing time-series data and missing business attribute data through a time-series data missing identification module 601 and a business attribute data missing identification module 602, enabling the classification and handling of two different types of missing data and avoiding poor completion results due to using a single method. The time-series data missing completion module 603 extracts and fuses time-domain and frequency-domain features for semantic correction, comprehensively utilizing local change trends, global periodic patterns, and external environmental semantics to improve the accuracy of time-series data completion. The business attribute data missing completion module 604 extracts relational subgraphs from the business logic knowledge graph and uses a large language model for reasoning, leveraging the structured relational constraints between known business entity data to logically complete missing attributes, thus improving the accuracy of business attribute data completion. Compared to existing technologies that are prone to failure and omissions when processing 5G wireless public network data completion, this application improves the accuracy of 5G wireless public network data completion.

[0097] It should be noted that the device embodiments described above are merely illustrative, and some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can specifically be implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0098] Based on the above embodiment of a 5G wireless public network data completion method based on a large model, another embodiment of the present invention provides a terminal device, which includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements a 5G wireless public network data completion method based on a large model according to any embodiment of the present invention.

[0099] For example, in this embodiment, the computer program can be divided into one or more modules, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the terminal device.

[0100] The terminal device may be a desktop computer, laptop, handheld computer, or cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.

[0101] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the terminal device, connecting all parts of the terminal device via various interfaces and lines.

[0102] Based on the above-described method embodiments, another embodiment of the present invention provides a computer-readable storage medium including a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to execute the 5G wireless public network data completion method based on a large model as described in any of the above-described method embodiments of the present invention.

[0103] The modules / units integrated in the device / terminal equipment, if implemented as software functional units and sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0104] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for data completion in 5G wireless public networks based on a large model, characterized in that, include: Extract the time series and numerical series of 5G wireless public network data, and perform time series data missing identification on the 5G wireless public network data based on the time series and the numerical series to obtain a first identification result; The 5G wireless public network data is traversed and matched according to a preset content feature library, and the missing service attribute data of the 5G wireless public network data is identified based on the matching results to obtain a second identification result. If the first identification result indicates that there is missing time-series data, then the time-domain features and frequency-domain features of the 5G wireless public network data are extracted, and the weighted fusion result of the time-domain features and frequency-domain features is semantically corrected to obtain the first completion value; If the second identification result indicates that there is missing business attribute data, then the association subgraph of the missing data is extracted from the pre-constructed business logic knowledge graph, and a prompt word is generated based on the association subgraph. The prompt word is then input into the pre-trained large language model for neural symbol reasoning to obtain the second completion value. When the first or second completion value exists, the 5G wireless public network data is backfilled with completion values ​​to obtain the completion result.

2. The 5G wireless public network data completion method based on a large model as described in claim 1, characterized in that, The extraction of time series and numerical series of 5G wireless public network data, and the identification of missing time series data based on the time series and numerical series to obtain a first identification result, includes: Extract the time series and numerical series corresponding to each data sequence in the 5G wireless public network data, and calculate the mutual information entropy between each set of time series and numerical series. If there is a mutual information entropy greater than a preset random noise threshold, the first identification result is that there is missing time series data, and the data sequence corresponding to the mutual information entropy is marked as a missing time series data sequence. The normalized variance of each numerical sequence is calculated. If there is a normalized variance in a preset high fluctuation range, the first identification result is that there is missing time series data, and the data sequence corresponding to the normalized variance is marked as a missing time series data sequence.

3. The 5G wireless public network data completion method based on a large model as described in claim 1, characterized in that, The step of traversing and matching the 5G wireless public network data according to a preset content feature library, and identifying missing service attribute data in the 5G wireless public network data based on the matching results to obtain a second identification result includes: The matching confidence between each data sequence in the 5G wireless public network data and the regular expression of each business entity data in the content feature library is calculated respectively. If all matching confidences corresponding to any data sequence are less than or equal to the preset validity threshold, the second identification result is that there is missing business attribute data, and the data sequence is marked as a data sequence with missing business attributes. If there is a matching confidence score greater than the validity threshold, it is determined that the data sequence corresponding to the matching confidence score matches the business entity data. The matching value between each data in the data sequence and the regular expression of the business entity data is calculated according to the indicator function. If there is a matching value of zero, the second identification result is that there is missing business attribute data, and the data sequence is marked as a data sequence with missing business attributes. If there is unmatched business entity data after the traversal matching is completed, the second identification result is that there is missing business attribute data, and all unmatched business entity data is marked as a sequence of missing business attribute data.

4. The 5G wireless public network data completion method based on a large model as described in claim 2, characterized in that, If the first identification result indicates missing time-series data, then the time-domain and frequency-domain features of the 5G wireless public network data are extracted, and the weighted fusion result of the time-domain and frequency-domain features is semantically corrected to obtain a first completion value, including: Each of the aforementioned time-series missing data sequences is subjected to dynamic masking processing to obtain multiple binary mask vectors; The missing values ​​of each of the time-series missing data sequences are determined based on each of the binary mask vectors, and each of the missing values ​​is replaced with a preset vector to obtain multiple first time-series missing data sequences; Calculate the relative distance between any two time steps in each of the first time-series missing data sequences, encode each relative distance and add it to the corresponding first time-series missing data sequence to obtain multiple second time-series missing data sequences; Differential operations are performed on each of the second time-series missing data sequences to obtain multiple velocity sequences and multiple acceleration sequences. The local mean sequence and local variance sequence of each of the second time-series missing data sequences are calculated through a preset sliding window. Each second time-series missing data sequence is concatenated with the corresponding velocity sequence, acceleration sequence, local mean sequence, and local variance sequence to obtain multiple feature vector sequences. Each feature vector sequence is mapped to multiple time-domain token sequences and input into the Transformer encoder for feature extraction to generate multiple time-domain features. The second time-series missing data sequences are decomposed into low-frequency features, mid-frequency features, and high-frequency features using discrete wavelet transform. The corresponding low-frequency features, mid-frequency features, and high-frequency features are then mapped and concatenated to obtain multiple frequency domain token sequences. These sequences are then input into the Transformer encoder for feature extraction to generate multiple frequency domain features. The corresponding time-domain features and frequency-domain features are weighted and fused to obtain multiple repair sequences. Then, the semantics of each repair sequence are corrected to obtain multiple first completion values.

5. The 5G wireless public network data completion method based on a large model as described in claim 4, characterized in that, The semantic correction of each repair sequence is performed to obtain multiple first completion values, including: Obtain the time data and geographic location data corresponding to each of the time-series missing data sequences, and retrieve external real-time knowledge bases based on each set of time data and geographic location data to construct environmental information prompt words corresponding to each of the time-series missing data sequences; Each of the environmental information prompts and the corresponding repair sequences are input into a large language model for analysis, generating multiple semantic correction bias sequences. Each repair sequence is added to its corresponding semantic correction bias sequence to obtain multiple first completion values.

6. The 5G wireless public network data completion method based on a large model as described in claim 3, characterized in that, If the second identification result indicates missing business attribute data, then an association subgraph of the missing data is extracted from the pre-constructed business logic knowledge graph, and a prompt word is generated based on the association subgraph. The prompt word is then input into a pre-trained large language model for neural symbol reasoning to obtain a second completion value, including: Using the context attributes of the missing data sequences of each business attribute as query conditions, matching is performed in the business logic knowledge graph to obtain the association subgraphs corresponding to the missing data sequences of each business attribute. A depth-first search strategy is used to traverse the topology of each associated subgraph, and each topology is converted into a text description to obtain multiple prompt words; wherein, the prompt words contain graph position encoding information of each entity node and relation edge of the corresponding associated subgraph; Each of the aforementioned prompt words is input into a pre-trained large language model for neural symbol reasoning to obtain multiple second completion values.

7. The 5G wireless public network data completion method based on a large model as described in claim 6, characterized in that, The step involves inputting each of the aforementioned prompt words into a pre-trained large language model for neural symbol reasoning to obtain multiple second completion values, including: Each of the aforementioned prompt words is input into a pre-trained large language model, so that the large language model can parse the graph symbol knowledge in each of the prompt words and perform chain inference in combination with general neural knowledge to generate a candidate entity set corresponding to each of the missing data. The confidence score of each candidate entity in each candidate entity set is calculated using the large language model, and the candidate entity with the highest confidence score in each candidate entity set is used as the second completion value of the corresponding business attribute missing data sequence, thus obtaining multiple second completion values.

8. A 5G wireless public network data completion device based on a large model, characterized in that, include: The module includes a time-series data missing identification module, a business attribute data missing identification module, a time-series data missing completion module, a business attribute data missing completion module, and a completion value backfilling module. The time-series data missing identification module is used to extract the time series and numerical series of 5G wireless public network data, and to perform time-series data missing identification on the 5G wireless public network data based on the time series and the numerical series to obtain a first identification result; The business attribute data missing identification module is used to traverse and match the 5G wireless public network data according to the preset content feature library, and to identify the missing business attribute data of the 5G wireless public network data based on the matching results, so as to obtain a second identification result. The time-series data missing completion module is used to extract the time-domain features and frequency-domain features of the 5G wireless public network data if the first identification result indicates that there is a time-series data missing, and to perform semantic correction on the weighted fusion result of the time-domain features and frequency-domain features to obtain the first completion value. The business attribute data missing completion module is used to extract the association subgraph of the missing data in the pre-constructed business logic knowledge graph if the second identification result is that there is missing business attribute data, generate prompt words based on the association subgraph, input the prompt words into the pre-trained large language model for neural symbol reasoning, and obtain the second completion value. The completion value backfilling module is used to backfill the 5G wireless public network data with completion values ​​when the first completion value or the second completion value exists, so as to obtain the completion result.

9. A terminal device, characterized in that, The method includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements a 5G wireless public network data completion method based on a large model as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, include: A stored computer program, wherein, when the computer program is executed, it controls the device containing the computer-readable storage medium to perform a 5G wireless public network data completion method based on a large model as described in any one of claims 1-7.

Images