A semantic slicing and two-stage reasoning method for massive communication data

Abstract

This invention discloses a semantic slicing and two-stage reasoning method for massive communication data, relating to the field of mobile communication technology. The specific steps of this method are as follows: First, standardize and preprocess multi-type data from cellular networks, extracting four types of fields and removing invalid data to generate a standard sequence; then, calculate a comprehensive score based on protocol boundaries and time interval features to select the optimal cutting point, and generate a slice set through context overlap; after local reasoning, each slice is deduplicated and integrated into an intermediate dataset using timestamps and identifiers; finally, global reasoning is used to correlate slices, reconstruct the communication process, and output the analysis results. This invention addresses semantic fragmentation by standardizing and preprocessing mobile cellular network communication data, employing adaptive semantic slicing, and using local and global two-stage reasoning to process slices in parallel and reconstruct end-to-end communication links across slices. It accurately identifies anomalies and bottlenecks, balancing processing efficiency and analysis depth, providing reliable support for network operation and maintenance optimization.

Description

Technical Field

[0001] This invention relates to the field of communication data processing technology, specifically to a semantic slicing and two-stage reasoning method for massive communication data. Background Technology

[0002] With the continuous development of mobile cellular network technology and its expanding application scope, the scale of signaling interaction data, protocol status data, network event data, and user behavior data generated during network operation is surging, forming massive communication data resources with strong time sequence and high correlation. This data comprehensively records core information such as network protocol interactions, device operating status, and service execution processes, serving as a crucial foundation for communication network operation and maintenance management, fault diagnosis, performance optimization, and service quality assurance. The parsing and analysis of communication data requires ensuring data standardization, semantic coherence, and comprehensive reasoning to accurately reconstruct communication processes and identify operational problems. Faced with the rapid growth in data volume, traditional processing models struggle to cope with the processing pressure brought by large-scale data. There is an urgent need for communication data processing and analysis technologies that are adaptable to massive scenarios and balance efficiency and accuracy to support the efficient operation and intelligent management of modern cellular networks.

[0003] Traditional communication data processing and reasoning methods have significant shortcomings and cannot meet the processing needs of massive amounts of communication data. Traditional data slicing often relies on fixed rules or manual division, failing to consider communication protocol characteristics and temporal distribution patterns. This easily leads to fragmented data semantics, loss of contextual information, and affects the accuracy of analysis results. Directly performing overall reasoning on massive amounts of data consumes a large amount of computing power, resulting in slow processing speed, high response latency, and difficulty in achieving real-time analysis. Local analysis results lack efficient deduplication and integration mechanisms, easily leading to data redundancy and duplicate results, resulting in distorted analysis conclusions. At the same time, traditional methods lack the ability to perform cross-slice semantic association and logical sorting, failing to reconstruct the complete communication interaction process, making it difficult to accurately locate communication anomalies, performance bottlenecks, and system failures. The mismatch between slice size and data distribution further reduces processing efficiency and cannot provide reliable support for network operation and maintenance. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies by providing a semantic slicing and two-stage reasoning method for massive communication data. This method standardizes and preprocesses various types of cellular network communication data, eliminating invalid information and forming a standardized time-series data sequence. Relying on an adaptive semantic slicing mechanism, it determines the optimal slicing point by combining protocol boundaries, time intervals, and data scale characteristics, preserving the contextual relationships between slices and avoiding semantic fragmentation. A two-stage mode combining local and global reasoning is employed. First, each slice is analyzed independently and deduplicated and integrated. Then, through cross-slice semantic association and logical organization, the communication interaction process is fully reconstructed, accurately identifying anomalies, bottlenecks, and faults, and outputting comprehensive analysis results. This method effectively solves the problems of low processing efficiency, incomplete analysis, and semantic incoherence in massive communication data processing, balancing processing efficiency and analytical accuracy, and providing reliable technical support for intelligent operation and maintenance of communication networks.

[0005] To solve the above-mentioned technical problems, this invention provides the following technical solution: a semantic slicing and two-stage reasoning method for massive communication data, the specific steps of which are as follows: S1, Data Standardization Preprocessing: The signaling interaction data, protocol status data, network event data, and user behavior data generated by the mobile cellular network are uniformly organized. Four types of fields are extracted from each record: timestamp, message identifier, message name, and message content. Invalid data with missing timestamps, duplicate message identifiers, or incomplete content are removed. After sorting by time, a standard communication data sequence is generated. S2, Adaptive Semantic Slicing: Taking standard communication data sequences as objects, protocol boundary feature values ​​are calculated through a protocol anchor point evaluation function, time interval feature values ​​are calculated through a time discontinuity evaluation function, and a comprehensive score for each candidate slice point is calculated based on the two feature values ​​and the scale adaptation result through a slice comprehensive scoring function. The scale adaptation result is determined by the matching degree between the current slice's cumulative data scale and the preset target scale. The highest score point is selected as the optimal slice point, and the starting position of adjacent slices is calculated through a context overlap start point update function to retain context information, generating a set of communication data slices. S3, Local Reasoning and Deduplication Integration of Results: Each slice in the communication data slice set is input independently, and the local reasoning model generates a set of local analysis results. The set of local analysis results is merged and deduplicated using timestamp and message identifier as the joint primary key to generate an intermediate analysis dataset. S4, Global Inference Output: Input the intermediate analysis dataset into the global inference model as a whole, perform cross-slice semantic association and logical sorting, restore the complete communication process, and generate the final analysis results.

[0006] Further, in S2, during the process of generating a set of communication data slices in adaptive semantic slicing, the starting position of the current slice is initialized to 0, the search offset range of candidate cutting points is initialized to a preset number of records offset, and the minimum slice size, target slice size, and maximum slice size are initialized to preset data size values; the search offset range is shifted backward from the starting position of the current slice to define the search interval of candidate cutting points; the cumulative data size from the starting position of the current slice to each candidate cutting point is calculated, and candidate cutting points with a cumulative data size greater than or equal to the minimum slice size and less than or equal to the maximum slice size are selected.

[0007] Furthermore, in S2, in the adaptive semantic slicing, after filtering candidate cut points whose cumulative data size is greater than or equal to the minimum size of the slice and less than or equal to the maximum size of the slice, the protocol boundary feature value and time interval feature value are calculated for each candidate cut point after filtering, the scale adaptation result corresponding to each candidate cut point is calculated, and the comprehensive score of each candidate cut point is calculated in combination with the weight parameters. All candidate cut points are then sorted in descending order according to the comprehensive score from high to low.

[0008] Furthermore, in S2, in the adaptive semantic slicing, after sorting all candidate cut points in descending order according to the comprehensive score, the candidate cut point ranked first is selected as the end position of the current slice. The communication data between the start position and the end position of the current slice is extracted to generate a single communication data slice. The start position of the next slice is calculated through the context overlap start point update function. The process of defining the candidate cut point search interval and calculating the start position of the next slice is repeated until all communication records in the standard communication data sequence have been processed.

[0009] Furthermore, in S3, the local reasoning and result deduplication integration process, the local reasoning model sequentially includes a slice data input layer, a communication semantic parsing layer, a feature structured conversion layer, and a local result output layer. The slice data input layer receives individual slice data from the communication data slice set. The communication semantic parsing layer parses the received slice data for text and protocol fields. The feature structured conversion layer converts the parsed communication information into data units with a fixed format. The local result output layer outputs the converted data units. The local reasoning model obtains a single slice within the communication data slice set, performs line-by-line parsing on all communication records within the slice, obtains the protocol interaction information and event features contained in the communication records, and generates a structured data unit composed of a timestamp, message identifier, and analysis content. All structured data units are combined according to the original order of the communication records within the slice to form a local analysis result set.

[0010] Furthermore, in S3, during the local reasoning and result deduplication integration, when performing a merge deduplication operation on the local analysis result set using the timestamp and message identifier as a combined primary key, all analysis items in the local analysis result set are traversed, and the timestamp and message identifier corresponding to each analysis item are read sequentially. The timestamp and message identifier are combined to form a combined primary key, and the recorded combined primary key information is compared. The analysis item corresponding to the first occurrence of the combined primary key is retained, and the analysis item corresponding to the repeated occurrence of the combined primary key is removed, thus completing the merge deduplication process.

[0011] Furthermore, in S4, during the global inference output, when the intermediate analysis dataset is input into the global inference model for cross-slice semantic association and logical sorting, the global inference model reads the timestamps, message identifiers, and analysis content of all analysis results in the intermediate analysis dataset. Based on the chronological order of the timestamps, it establishes the temporal correspondence between different analysis results; based on the matching relationship of the message identifiers, it establishes the semantic correspondence between different slice analysis results; based on the temporal correspondence and semantic correspondence, it establishes the temporal logical association between different slice analysis items, sorts out the state evolution process of each stage in the communication process, and restores the complete end-to-end communication session interaction link.

[0012] Furthermore, in S4, after restoring the complete end-to-end communication session interaction link in the global inference output, the global inference model identifies abnormal behaviors, performance bottlenecks and system failures in the communication process based on the complete communication flow, and generates a final analysis result including four parts: anomaly location, cause analysis, scope of impact and optimization suggestions.

[0013] Compared with existing technologies, this semantic slicing and two-stage reasoning method for massive communication data has the following advantages: This invention standardizes and preprocesses various types of communication data from mobile cellular networks, unifying data field specifications and eliminating invalid, missing, and duplicate abnormal data. It then organizes the data into a standard communication data sequence according to time sequence, ensuring data integrity, standardization, and consistency from the source, laying a high-quality data foundation for subsequent analysis. An adaptive semantic slicing mechanism is employed, comprehensively evaluating the optimal slicing point based on protocol boundary characteristics, time interval characteristics, and scale adaptation results. Context overlap settings preserve the correlation information between slices, effectively solving the problems of semantic fragmentation, scale imbalance, and context loss when slicing massive amounts of communication data. This ensures that slice division aligns with the semantic logic and actual distribution patterns of the communication data. Slices are independently input into a local inference model for analysis, and then merged and deduplicated using dual identifiers to generate an intermediate dataset. This achieves parallel and efficient processing of massive amounts of data while eliminating interference from repeated analysis, significantly improving the efficiency and accuracy of communication data parsing.

[0014] II. This invention employs a two-stage reasoning model to perform in-depth analysis of communication data. First, local reasoning achieves semantic parsing and structured output of single-slice communication data. Then, the integrated and deduplicated intermediate dataset is input into the global reasoning model. Based on temporal sequence and identifier associations, semantic and logical connections across slices are established, fully reconstructing the end-to-end communication session interaction chain, overcoming the limitation of single-slice analysis not covering the entire process. Global reasoning can comprehensively analyze the state evolution of the communication process, accurately identify abnormal behaviors, performance bottlenecks, and system faults during communication, and output complete analysis results including anomaly location, cause analysis, impact scope, and optimization suggestions. This method balances the efficiency of processing massive amounts of communication data with the depth of analysis, significantly improving the efficiency of troubleshooting communication network problems and providing comprehensive and reliable data support for the operation and maintenance optimization of communication systems.

[0015] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0017] Figure 1 A flowchart for a semantic slicing and two-stage reasoning method for massive communication data; Figure 2 A schematic diagram illustrating data transmission between each step of a semantic slicing and two-stage reasoning method for massive communication data. Figure 3 The flowchart shows the candidate cut-off point selection and size constraints. Detailed Implementation

[0018] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0019] Reference Figure 1One embodiment of the present invention proposes a semantic slicing and two-stage reasoning method for massive communication data. It adopts a progressive data processing mechanism of data standardization preprocessing, adaptive semantic slicing, local reasoning and result deduplication integration, and global reasoning output. It can adapt to the processing needs of massive signaling, protocol, event, and behavior communication data in mobile cellular networks. It reduces the reasoning load of a single-round model by semantic slicing and reduces the inference load by combining local and global two-stage reasoning to achieve cross-slice semantic association and complete communication process reconstruction. At the same time, it accurately identifies abnormal behavior, performance bottlenecks and system failures in the communication process and outputs complete analysis results including anomaly location, cause analysis, scope of impact and optimization suggestions.

[0020] The method described in this embodiment specifically includes: S1, Data Standardization Preprocessing: The signaling interaction data, protocol status data, network event data, and user behavior data generated by the mobile cellular network are uniformly standardized, and four types of fields, namely timestamp, message identifier, message name, and message content, are extracted. Invalid data is removed and sorted by time to generate a standard communication data sequence. S2, Adaptive Semantic Slicing: Taking standard communication data sequences as the processing object, the slice parameters are initialized and the candidate cut point search interval is defined. The candidate cut point scores are calculated through the protocol anchor point evaluation function, the time discontinuity evaluation function, and the slice comprehensive scoring function. The optimal cut point is selected to generate a single slice. The starting position of the next slice is determined by the context overlap start point update function. The process is repeated to generate a set of communication data slices. S3, Local Reasoning and Result Deduplication Integration: Each slice in the communication data slice set is independently input into the local reasoning model, and a local analysis result set is generated through multi-layer structure parsing and transformation. The merge and deduplication are completed using timestamp and message identifier as joint primary keys to obtain an intermediate analysis dataset with time-ordered and no duplicate entries. S4, Global Inference Output: The intermediate analysis dataset is input into the global inference model. Based on time sequence and message identifiers, cross-slice semantic relationships are established. The communication state evolution process is analyzed to reconstruct the complete communication session link. After identifying communication anomalies and bottlenecks, the final analysis results are generated, such as... Figure 2 As shown.

[0021] S1, Data standardization preprocessing: Specifically, data standardization preprocessing is the process of standardizing, cleaning, and organizing the raw communication data. It aims to eliminate format differences, redundant information, and invalid entries in the raw data, providing a unified and standardized data foundation for subsequent slicing and inference. This process sequentially completes five operations: data standardization, key field extraction, invalid data removal, format standardization, and time-series sorting.

[0022] First, we will carry out unified data standardization. For signaling interaction data, protocol status data, network event data, and user behavior data collected from mobile cellular networks, we will break down the format barriers of different data sources, integrate heterogeneous data into a unified processing object, ensure that all data follows the same processing rules, and eliminate processing obstacles caused by differences in data sources.

[0023] Subsequently, key fields were extracted. From each original communication record after normalization, four types of fields were accurately extracted: timestamp, message identifier, message name, and message content. Auxiliary fields that are irrelevant to communication semantic analysis were discarded to reduce invalid data interference and focus on key information of the communication process, so as to provide accurate data carriers for subsequent semantic analysis and reasoning.

[0024] Next, invalid data removal is performed, filtering and removing three types of invalid records according to preset rules: the first type is records with missing timestamps, which cannot be linked to a time sequence and have no value for communication flow analysis; the second type is records with duplicate message identifiers, which will cause subsequent deduplication and semantic association logic to become disordered; the third type is records with incomplete message content, which cannot complete semantic parsing and feature extraction. After removal, only communication records with complete fields, unique identifiers, and valid time sequences are retained.

[0025] Then, format standardization processing is performed, and the four core fields are standardized: the timestamps of all records are converted according to the ISO8601 standard time format to unify the representation of time data and ensure the accuracy of time sequence comparison; a globally unique 128-bit message identifier is assigned to each valid communication message to ensure that there are no duplicate identifiers in the entire communication data; the standard name of the communication protocol is uniformly used as the message name to eliminate parsing errors caused by non-standard expressions such as custom naming and abbreviations; binary garbled characters, newline characters, tab characters, and redundant padding symbols in the message content are cleaned up, and plain text protocol interaction information is retained to improve the efficiency and accuracy of subsequent semantic parsing.

[0026] Finally, time-series sorting is performed, arranging all the regularized, extracted, eliminated, and format-standardized communication records in ascending order of timestamp, forming a continuous and uninterrupted standard communication data sequence. Each record in this sequence consists of a fixed four-tuple structure composed of four elements: timestamp, message identifier, message name, and message content. The sequence has no time-series errors, no missing fields, and no format differences, and can be directly used for subsequent adaptive semantic slicing processing.

[0027] For example, when performing standardized preprocessing on communication data collected from mobile cellular networks, the following steps are taken: First, the signaling interaction data, protocol status data, network event data, and user behavior data are uniformly organized. Then, four types of fields—timestamp, message identifier, message name, and message content—are extracted line by line. Invalid records with missing timestamps, duplicate message identifiers, or incomplete message content are removed according to preset rules, while compliant communication data is retained. The timestamps of all records are converted to the ISO8601 standard format, and a globally unique 128-bit message identifier is assigned to each communication message. The message names are uniformly adjusted to the standard naming of the communication protocol, and binary garbled characters, newlines, and redundant padding symbols in the message content are removed. Finally, the processed communication records are arranged in ascending order of timestamps to form a continuous, uninterrupted, and structurally unified standard communication data sequence. Each record in the sequence uses a four-tuple structure consisting of timestamp, message identifier, message name, and message content, providing a stable data foundation for subsequent adaptive semantic slicing.

[0028] Optionally, during the data standardization preprocessing, field validation rules can be added according to the business type of the communication data, and dedicated validation logic can be configured for voice communication data and data service data respectively, to further improve data validity and adaptability.

[0029] S2, Adaptive semantic slicing: Specifically, adaptive semantic slicing is a process of intelligently segmenting standard communication data sequences based on three dimensions: protocol boundaries, time intervals, and data scale. It determines the optimal cutting point through multi-function evaluation while preserving contextual information between slices to avoid semantic breaks. This process sequentially completes eight operations: parameter initialization, candidate cutting point search interval definition, effective cutting point screening, multi-dimensional feature calculation, comprehensive score ranking, optimal cutting point selection, context overlap update, and iterative slice generation. Figure 3 As shown.

[0030] First, initialize the slice processing parameters, set the current slice start position to the start position of the standard communication data sequence, set the candidate cut point search offset range to the preset number of records offset, and configure the minimum slice size, slice target size, and slice maximum size to preset data size values. By using preset size parameters to constrain the data volume of a single slice, it adapts to the processing capabilities of the local inference model.

[0031] Then, the candidate cutting point search range is defined. Based on the current slice start position, the search offset range is shifted backward to determine the search range of candidate cutting points. Only valid cutting points are selected within this range, thus narrowing the cutting point selection range and improving the efficiency of slice processing.

[0032] Next, we screen for valid candidate cut points. We calculate the cumulative data size of each candidate cut point from the current slice start position to the search interval, and retain only candidate cut points whose cumulative data size is greater than or equal to the minimum slice size and less than or equal to the maximum slice size. We remove invalid cut points that are too small to complete the semantic parsing or too large to exceed the model's processing capacity, thus ensuring the reasonableness of the slice size.

[0033] For the selected valid candidate cut points, the three types of feature values ​​and the fitting results are calculated sequentially, as shown in the following function: Protocol anchor evaluation function: used to quantify the protocol boundary feature value of the i-th communication record, and determine whether the record is a critical boundary point of the communication protocol. The mathematical expression is:

[0034] in, Indicates the first Protocol boundary characteristic value of each communication record Let A represent the i-th communication record, and let A represent the protocol anchor event set, which includes key communication protocol events such as terminal attachment, bearer establishment, location update, service initiation, and terminal separation.

[0035] The time-discontinuity evaluation function is used to quantify the time interval characteristic value of the i-th communication record and determine whether the time difference between this record and the previous record exceeds a preset threshold. Its mathematical expression is:

[0036] in, This represents the time interval characteristic value of the i-th communication record. This represents the time difference between the i-th communication record and the (i-1)-th communication record. This indicates the preset time-to-time threshold.

[0037] Size fit result: This measures the degree of matching between the cumulative data size corresponding to the current candidate cut point and the preset target size. The closer the cumulative data size is to the preset target size, the higher the corresponding size fit result value.

[0038] Scale adaptation results The mathematical expression is: , TargetSize represents the cumulative data size corresponding to the i-th candidate cut point, MaxSize represents the preset slice target size, and MaxSize represents the preset slice maximum size. The slice comprehensive scoring function calculates the comprehensive score of the i-th candidate slice based on three indicators: fusion protocol boundary, time interval, and scale adaptation. The mathematical expression is:

[0039] in, This represents the overall score of the i-th candidate cutting point. , , These represent the weight parameters for the protocol boundary feature value, the time interval feature value, and the scale adaptation result, respectively. This represents the size adaptation result of the i-th candidate cut point; Protocol boundary feature value, time interval feature value, and weight parameters of scale adaptation results , , The calculation is performed using a normalization algorithm based on feature contribution, and the formulas are as follows: ; ; ;in This represents the numerical value of the feature contribution of protocol boundary features in the semantic slice of communication data. This represents the numerical value of the time interval feature in the semantic slice of communication data. This represents the feature contribution value of the scale adaptation result in the semantic slice of communication data. The three weight parameters satisfy... The constraints.

[0040] After the comprehensive score calculation is completed, all valid candidate cut points are sorted in descending order of score from high to low. The candidate cut point ranked first is selected as the end position of the current slice. The communication data between the start position and the end position of the current slice is extracted to generate a single communication data slice.

[0041] Then, the starting position of the next slice is calculated using the context overlap start point update function, preserving the context information between slices and avoiding semantic loss across slices. The mathematical expression is:

[0042] in, This indicates the starting position of the (k+1)th slice. This indicates the end position of the k-th slice, L represents the preset context overlap length, and the max function ensures that the starting position is not less than 0 to avoid data overflow.

[0043] Finally, the above operations are repeated, starting from the updated starting position, repeatedly defining the search interval, filtering the cut-off point, calculating the score, generating slices, and updating the starting position, until all communication records in the standard communication data sequence have been processed, and finally a complete set of communication data slices is generated.

[0044] For example, when performing adaptive semantic slicing on a standard communication data sequence, the slicing processing parameters are first initialized. The starting position of the current slice is set as the starting position of the sequence, and the candidate cut-point search offset range is set to a preset number of records. Simultaneously, the minimum slice size, target slice size, and maximum slice size are configured as preset data size values. Using the current slice starting position as a reference, the candidate cut-point search interval is defined by shifting backward by the preset search offset range. The cumulative data size of each candidate cut-point within the search interval from the starting position is calculated, and valid candidate cut-points whose cumulative data size conforms to the preset size range are selected. The protocol boundary feature value of the valid candidate cut-points is calculated based on the protocol anchor evaluation function, and the time interval feature value is calculated based on the time discontinuity evaluation function. Combining preset weight parameters and size adaptation results, a comprehensive score for each candidate cut-point is obtained through a slice comprehensive scoring function. The valid candidate cut-points are sorted from high to low according to their comprehensive scores, and the cut-point with the highest score is selected as the end position of the current slice. The corresponding interval of communication data is extracted to generate a single communication data slice. The starting position of the next slice is calculated by updating the context overlap starting point function. The context information between slices is preserved. The operations of candidate cutting point screening, score calculation, slice generation and starting position update are repeated until all records in the standard communication data sequence have been sliced. Finally, a set of communication data slices containing multiple context overlap information is generated.

[0045] Optionally, during the adaptive semantic slicing process, the slice size parameters and weight parameters can be dynamically adjusted according to the network load to adapt the slice processing efficiency and system resource consumption under different load conditions, ensuring the stable operation of the overall processing flow.

[0046] S3, Local Reasoning and Deduplication Integration of Results: Specifically, the local reasoning and result deduplication integration is divided into two processing stages: local reasoning analysis and merging deduplication. First, the semantic parsing and structure transformation of a single slice are completed through the local reasoning model. Then, the duplicate results caused by context overlap are eliminated by deduplication through the joint primary key, ensuring the uniqueness and temporal integrity of the intermediate dataset.

[0047] Local reasoning analysis: The local inference model adopts a four-layer structure: slice data input layer, communication semantic parsing layer, feature structure transformation layer, and local result output layer. Each layer works together to complete the parsing and result generation of a single slice. Slice data input layer: responsible for receiving individual slice data from the communication data slice set, passing the quadruple structure record in the slice into the next layer in the original time sequence, and keeping the original time sequence relationship of the data unchanged; Communication semantic parsing layer: Performs line-by-line text parsing and protocol field decomposition on the input slice data, extracts semantic information such as protocol type, interaction status, event parameters, and business type from each record, and fully identifies the communication logic and event characteristics within a single slice; Feature structuring transformation layer: Converts the parsed unstructured communication semantic information into fixed-format data units containing timestamps, message identifiers, and analysis content, unifying the structural form of all analysis results and facilitating subsequent integration and processing; Local Result Output Layer: Combines all fixed-format data units according to the original order of communication records within the slice, and outputs the local analysis results corresponding to the slice.

[0048] The local inference model processes each slice in the communication data slice set independently, parses all communication records in the slice line by line, extracts complete protocol interaction information and event features, and finally generates a local analysis result set consisting of the results corresponding to all slices. Each result in the set contains an independent timestamp, message identifier, and analysis content field, and maintains a one-to-one correspondence with the communication records in the slice.

[0049] Merge and deduplication: The deduplication process uses timestamps and message identifiers as a composite primary key. The specific steps are as follows: First, iterate through all analysis items in the local analysis result set, reading the timestamp and message identifier corresponding to each analysis item in turn; concatenate the timestamp and message identifier to form a unique composite primary key; create a composite primary key record table, storing the first occurrence of the composite primary key and its corresponding analysis item in the record table, and directly removing the analysis items corresponding to duplicate composite primary keys; after the iteration is complete, arrange all analysis items in the record table in chronological order of timestamps to generate an intermediate analysis dataset.

[0050] This dataset has two characteristics: all analysis results are arranged in complete order of communication occurrence, without any disorder; each analysis result corresponds to a unique composite primary key, and there are no duplicate entries within the dataset, making it directly usable as input for global inference models.

[0051] For example, during the local inference and result deduplication integration process, each slice in the communication data slice set is input into the local inference model one by one. The slice data input layer receives the data of a single slice while maintaining the temporal sequence. The communication semantic parsing layer disassembles the protocol fields line by line and extracts semantic information. The feature structuring transformation layer converts unstructured information into fixed-format data units. The local result output layer outputs the local analysis results of a single slice in the original order. After all slices are processed, a complete set of local analysis results is formed. All analysis items in the local analysis result set are traversed, and the timestamp and message identifier of each item are read and concatenated into a composite primary key. The composite primary key record table retains the analysis item that appears for the first time and removes the analysis items that appear repeatedly. The retained analysis items are sorted in ascending order by timestamp, finally forming an intermediate analysis dataset with complete temporal sequence and no duplicate entries.

[0052] Optionally, during local inference, a batch parallel parsing mode can be adopted to process multiple communication data slices simultaneously, thereby shortening the overall time consumption of local inference and improving data processing efficiency.

[0053] S4, Global Inference Output: Specifically, the global inference output is the process of performing a holistic analysis of the intermediate analysis dataset. Through cross-slice semantic association and logical sorting, it restores the complete communication process, identifies communication problems, and generates standardized analysis results. This process sequentially completes five operations: data reading, establishment of temporal and semantic associations, restoration of communication process, identification of anomalies and bottlenecks, and generation of final results.

[0054] First, the global inference model reads three types of information from the intermediate analysis dataset: timestamps, message identifiers, and analysis content of all analysis results. It then loads all the data into the model processing unit, completing the data preprocessing and loading, and preparing for subsequent correlation analysis.

[0055] Then, the cross-slice relationship is established, which is done in two steps: 1. Establishing the temporal correspondence: Based on the order of timestamps, all analysis results are linked together in the order of the communication events to determine the order and evolution logic of the communication events; Second, semantic correspondence is established. Based on the matching relationship of message identifiers, analysis items belonging to the same communication session and the same protocol interaction in different slices are associated, breaking down semantic barriers between slices and realizing cross-slice information connectivity.

[0056] Based on the dual correspondence between timing and semantics, the global reasoning model sorts out the state evolution process of each stage in the communication process, from terminal access to the network, service initiation, data transmission, location switching, to terminal disconnection, and completely restores the end-to-end communication session interaction link. The restored link has no breaks, no semantic missing, and no timing disorder.

[0057] After reconstructing the communication process, the model performs a comprehensive inspection of the complete interaction link, identifying three types of communication problems: abnormal behavior, including signaling interaction failure, message loss, and unauthorized access; performance bottlenecks, including excessive data transmission latency, handover lag, and transmission rate failure; and system failures, including abnormal base station operation, incorrect protocol parameter configuration, and network congestion.

[0058] Finally, the model generates a final analysis result comprising four parts: anomaly localization, cause analysis, scope of impact, and optimization suggestions. Anomaly localization clarifies the timing, message identifier, and interaction link of the problem; cause analysis dissects the protocol, network, and device-level causes of the problem; scope of impact statistics include the number of affected terminals, service types, and runtime periods; and optimization suggestions provide specific implementation plans for parameter adjustment, device maintenance, and network optimization.

[0059] For example, when executing the global inference output, the intermediate analysis dataset is completely input into the global inference model. The model reads the timestamps, message identifiers, and analysis content of all analysis results, establishes a temporal correspondence between communication events based on timestamps, and establishes a semantic correspondence between different slices through message identifier matching. Based on the dual relationships, the model sorts out the communication state evolution process and reconstructs the complete end-to-end communication session interaction link. The model performs a comprehensive inspection of the complete interaction link, identifies abnormal behaviors, performance bottlenecks, and system failures in the communication process, determines the location, cause, and scope of impact for each identified problem, and generates corresponding optimization solutions based on the communication network operation rules. Finally, the results are integrated into a complete analysis result that includes anomaly location, cause analysis, scope of impact, and optimization suggestions, completing the global inference output.

[0060] Optionally, after the global inference output, the final analysis results can be synchronized to the communication network operation and maintenance platform to realize automatic alarm and closed-loop processing of abnormal problems, thereby improving the intelligence level of communication network operation and maintenance.

[0061] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A semantic slicing and two-stage reasoning method for massive communication data, characterized in that, The specific steps of this method are as follows: S1, Data Standardization Preprocessing: The signaling interaction data, protocol status data, network event data, and user behavior data generated by the mobile cellular network are uniformly organized. Four types of fields are extracted from each record: timestamp, message identifier, message name, and message content. Invalid data with missing timestamps, duplicate message identifiers, or incomplete content are removed. After sorting by time, a standard communication data sequence is generated. S2, Adaptive Semantic Slicing: Taking standard communication data sequences as objects, the protocol boundary feature value is calculated through the protocol anchor evaluation function, the time interval feature value is calculated through the time discontinuity evaluation function, and the comprehensive score of each candidate slicing point is calculated based on the two feature values ​​and the scale adaptation result through the slice comprehensive scoring function. The scale adaptation result is determined by the matching degree between the current slice cumulative data scale and the preset target scale. The highest score point is selected as the optimal cutting point. The starting positions of adjacent slices are calculated using the context overlap start point update function to preserve context information, and a set of communication data slices is generated. S3, Local Reasoning and Deduplication Integration of Results: Each slice in the communication data slice set is input independently, and the local reasoning model generates a set of local analysis results. The set of local analysis results is merged and deduplicated using timestamp and message identifier as the joint primary key to generate an intermediate analysis dataset. S4, Global Inference Output: Input the intermediate analysis dataset into the global inference model as a whole, perform cross-slice semantic association and logical sorting, restore the complete communication process, and generate the final analysis results.

2. The semantic slicing and two-stage reasoning method for massive communication data according to claim 1, characterized in that, In S2, during the process of generating a set of communication data slices in adaptive semantic slicing, the starting position of the current slice is initialized to 0, the search offset range of the candidate cutting point is initialized to a preset number of records offset, and the minimum slice size, the target slice size, and the maximum slice size are initialized to preset data size values. Based on the current slice start position, the candidate cut point search interval is defined by shifting the search offset range backward. Calculate the cumulative data size from the current slice start position to each candidate cut point, and filter candidate cut points whose cumulative data size is greater than or equal to the minimum slice size and less than or equal to the maximum slice size.

3. The semantic slicing and two-stage reasoning method for massive communication data according to claim 2, characterized in that, In S2, in adaptive semantic slicing, after filtering candidate cut points whose cumulative data size is greater than or equal to the minimum size of the slice and less than or equal to the maximum size of the slice, the protocol boundary feature value and time interval feature value are calculated for each candidate cut point after filtering. The scale adaptation result corresponding to each candidate cut point is calculated. The comprehensive score of each candidate cut point is calculated in combination with the weight parameters. All candidate cut points are sorted in descending order according to the comprehensive score from high to low.

4. The semantic slicing and two-stage reasoning method for massive communication data according to claim 2, characterized in that, In S2, in adaptive semantic slicing, after sorting all candidate cut points in descending order according to the comprehensive score, the candidate cut point with the highest score is selected as the end position of the current slice. The communication data between the start position and the end position of the current slice is extracted to generate a single communication data slice. The start position of the next slice is calculated by the context overlap start point update function. The process of defining the candidate cut point search interval and calculating the start position of the next slice is repeated until all communication records in the standard communication data sequence have been processed.

5. The semantic slicing and two-stage reasoning method for massive communication data according to claim 1, characterized in that, In S3, during the local reasoning and result deduplication integration, the local reasoning model sequentially includes a slice data input layer, a communication semantic parsing layer, a feature structuring transformation layer, and a local result output layer. The slice data input layer receives individual slice data from the communication data slice set. The communication semantic parsing layer parses the received slice data for text and protocol fields. The feature structuring transformation layer converts the parsed communication information into data units with a fixed format. The local result output layer outputs the transformed data units. The local reasoning model obtains a single slice within the communication data slice set, performs line-by-line parsing on all communication records within the slice, obtains the protocol interaction information and event features contained in the communication records, and generates structured data units composed of timestamps, message identifiers, and analysis content. All structured data units are combined according to the original order of the communication records within the slice to form a local analysis result set.

6. The semantic slicing and two-stage reasoning method for massive communication data according to claim 1, characterized in that, In S3, during the local reasoning and result deduplication integration, when performing a merge deduplication operation on the local analysis result set using the timestamp and message identifier as a combined primary key, all analysis items in the local analysis result set are traversed, and the timestamp and message identifier corresponding to each analysis item are read sequentially. The timestamp and message identifier are combined to form a combined primary key, and the recorded combined primary key information is compared. The analysis item corresponding to the first occurrence of the combined primary key is retained, and the analysis item corresponding to the duplicate occurrence of the combined primary key is removed, thus completing the merge deduplication process.

7. The semantic slicing and two-stage reasoning method for massive communication data according to claim 1, characterized in that, In S4, during the global inference output, when the intermediate analysis dataset is input into the global inference model for cross-slice semantic association and logical sorting, the global inference model reads the timestamps, message identifiers, and analysis content of all analysis results in the intermediate analysis dataset. Based on the chronological order of the timestamps, it establishes the temporal correspondence between different analysis results. Based on the matching relationship of the message identifiers, it establishes the semantic correspondence between different slice analysis results. Based on the temporal correspondence and semantic correspondence, it establishes the temporal logical association between different slice analysis items, sorts out the state evolution process of each stage in the communication process, and restores the complete end-to-end communication session interaction link.

8. The semantic slicing and two-stage reasoning method for massive communication data according to claim 7, characterized in that, In S4, the global inference output, after restoring the complete end-to-end communication session interaction link, the global inference model identifies abnormal behaviors, performance bottlenecks and system faults in the communication process based on the complete communication process, and generates the final analysis results including four parts: anomaly location, cause analysis, scope of impact and optimization suggestions.

Images