CN-122020538-A - Multisource heterogeneous data fusion and processing method based on reinforcement learning mechanism

CN122020538ACN 122020538 ACN122020538 ACN 122020538ACN-122020538-A

Abstract

The invention discloses a multisource heterogeneous data fusion and processing method based on an reinforcement learning mechanism, which relates to the technical field of communication network operation and maintenance and data fusion, and the method comprises the steps of acquiring multisource heterogeneous data in a communication network in real time, abstracting and constructing a perception topological network, and configuring a first fusion strategy for feature nodes in the perception topological network; the method comprises the steps of introducing a hierarchical reinforcement learning mechanism to evaluate and analyze a perceived topological network, wherein the hierarchical reinforcement learning mechanism comprises an upper reinforcement strategy and a lower reinforcement strategy, wherein the upper reinforcement strategy is executed, a topological attention domain is determined through state evaluation, the lower reinforcement strategy is executed in the topological attention domain, a change instruction is generated, a first fusion strategy is changed, and a second fusion strategy is obtained.

Inventors

FENG XINGLONG
Zu Jinwei
LU YUSEN
ZHANG YAFEI

Assignees

西安星讯智能通信科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260128

Claims (10)

1. A multisource heterogeneous data fusion and processing method based on an reinforcement learning mechanism is characterized by comprising the following steps: the method comprises the steps of acquiring multi-source heterogeneous data in a communication network in real time, abstracting and constructing a perceived topological network, and configuring a first fusion strategy for characteristic nodes in the perceived topological network; Introducing a hierarchical reinforcement learning mechanism, and evaluating and analyzing a perceived topological network; The hierarchical reinforcement learning mechanism comprises an upper reinforcement strategy and a lower reinforcement strategy, wherein the upper reinforcement strategy is executed, a topological attention domain is determined through state evaluation, the lower reinforcement strategy is executed in the topological attention domain, a change instruction is generated, the first fusion strategy is changed, and the second fusion strategy is obtained.
2. The method for merging and processing heterogeneous data based on reinforcement learning mechanism according to claim 1, wherein the heterogeneous data comprises performance index, log text and operation instruction sequence of a communication network.
3. The multi-source heterogeneous data fusion and processing method based on the reinforcement learning mechanism of claim 1, wherein the abstract construction of the perceptual topology network comprises the following steps: Based on the multi-source heterogeneous data, extracting key elements, including a time sequence stream and a semantic stream; The method comprises the steps of mapping key elements into feature nodes of a perceived topological network, configuring a time sequence stream as node attributes of the feature nodes, configuring a semantic stream as text embedding of the feature nodes, constructing a topological edge between the feature nodes by monitoring the feature nodes with logical relations, and configuring a first fusion strategy for the feature nodes.
4. The method for multi-source heterogeneous data fusion and processing based on reinforcement learning mechanism of claim 2, wherein the first fusion strategy comprises a data structure, a weight adjustment category and an interactive gating, wherein the data structure is four-level and comprises a time stamp index, a data attribute, an entity identifier and a logic relationship, the weight adjustment category comprises weighting, weight reduction or shielding, and the interactive gating comprises a transfer gating threshold.
5. The method for multi-source heterogeneous data fusion and processing based on reinforcement learning mechanism of claim 1, wherein the executing the upper layer reinforcement strategy comprises: presetting a first time slice, and constructing an evaluation task of a second time slice in the process of evaluating the state of the first time slice, wherein the time stamp index of the second time slice is larger than that of the first time slice, and the first time slice is a waiting window with the first time sequence ranking in a current data buffer zone; extracting multi-source heterogeneous data in a first time slice to execute state evaluation, obtaining a first result, and storing a state score generated under the first time slice into a global feature library when the first result indicates that the state is stable; performing parameter broadcasting on the evaluation task of the second time slice, and identifying the scoring change track of each characteristic node to obtain a second result of the second time slice; and if the second result indicates the state oscillation, activating a linkage mechanism, identifying heterogeneous data with the same time sequence in the communication network, and determining a topological attention domain corresponding to the characteristic node generating the state oscillation.
6. The method for multi-source heterogeneous data fusion and processing based on reinforcement learning mechanism of claim 5, wherein performing network state evaluation to obtain a first result comprises: Extracting average value and fluctuation value of each performance index in a first time slice, generating a state score through weighted summation, and comparing the state score with a standard stability threshold value to dynamically trigger a score processing branch, wherein the score processing branch comprises a basic score branch and an enhanced score branch; Under the triggering of basic grading branches, extracting the deviation degree of the performance index and generating an abnormal fluctuation mask; Under the triggering of the enhanced scoring branch, a log text and an operation instruction sequence are called, and log abnormal characteristics and instruction intention characteristics are extracted to generate an abnormal event mask, wherein the abnormal fluctuation mask and the abnormal event mask are characterized in a matrix form; and combining the perceived topological network, judging that the abnormal fluctuation mask and the abnormal event mask are overlapped, and obtaining a first result.
7. The method for multi-source heterogeneous data fusion and processing based on reinforcement learning mechanism of claim 5, wherein the performing of the underlying reinforcement strategy in the topological interest domain comprises: extracting feature items of corresponding feature nodes based on the topological attention domain, wherein the feature items comprise performance index field values, log text keywords and operation instruction intents; Determining a single quantitative score of a characteristic item by adopting a nonlinear power function to obtain a corresponding oscillation factor, wherein the corresponding oscillation factor comprises a plurality of first oscillation factors, a plurality of second oscillation factors and a plurality of third oscillation factors, the first oscillation factors correspond to a performance index field value, the second oscillation factors correspond to a log text keyword, the third oscillation factors correspond to an operation instruction intention, and hierarchical relation data from top to bottom exists among the first oscillation factors, the second oscillation factors and the third oscillation factors; determining a plurality of candidate enhancement strategies matched with the oscillation factors, wherein the candidate enhancement strategies at least comprise modification parameters; Calculating the comprehensive score of each enhancement strategy in the plurality of candidate enhancement strategies, and sequencing the plurality of candidate enhancement strategies in a descending order to obtain a sequencing result, wherein the comprehensive score of each enhancement strategy is obtained by performing accumulation and averaging on oscillation factors of the same level to obtain a first score, and performing accumulation on the first scores of different levels to obtain a first score; and selecting the first candidate enhancement strategy of the sorting as a target strategy based on the sorting result, and extracting the corresponding change parameters to generate a change instruction.
8. The multi-source heterogeneous data fusion and processing method based on the reinforcement learning mechanism of claim 7, wherein the change parameters comprise a change type, a change content set and a change relation set, wherein the change type is a change type of a first fusion strategy, the change content set is the content of the first fusion strategy, and the change relation set is a logic relation among feature nodes.
9. The method for merging and processing multi-source heterogeneous data based on reinforcement learning mechanism according to claim 1, wherein the merging is performed on the multi-source heterogeneous data based on a second merging strategy.
10. A multi-source heterogeneous data fusion and processing system based on an reinforcement learning mechanism is characterized by comprising a data acquisition module, a data processing module and a data processing module, wherein the data acquisition module acquires multi-source heterogeneous data in a communication network in real time; the first processing module is used for abstracting and constructing a perceived topological network based on the multi-source heterogeneous data and configuring a first fusion strategy for the characteristic nodes in the perceived topological network; the second processing module introduces a hierarchical reinforcement learning mechanism to evaluate and analyze the perceived topology network; The hierarchical reinforcement learning mechanism comprises an upper reinforcement strategy and a lower reinforcement strategy, wherein the upper reinforcement strategy is executed, a topological attention domain is determined through state evaluation, the lower reinforcement strategy is executed in the topological attention domain, a change instruction is generated, the first fusion strategy is changed, and the second fusion strategy is obtained.

Description

Multisource heterogeneous data fusion and processing method based on reinforcement learning mechanism Technical Field The invention relates to the technical field of communication network operation and maintenance and data fusion, in particular to a multi-source heterogeneous data fusion and processing method based on an enhanced learning mechanism. Background As the communication network evolves towards 5G and cloud native architecture, the network topology structure is increasingly complex, so that the generated monitoring data presents high multi-source isomerism, and at present, the monitoring and fault management for the network state mainly depend on comprehensive analysis of performance indexes, log text and operation instructions, and a deep learning model is introduced, for example, a gating circulation unit or a Transformer architecture is selected to perform sequential modeling on equipment logs, network traffic, system call sequences and the like, so as to generate fusion features. The traditional data fusion technology has the defects that on one hand, the prior art processes heterogeneous data by adopting a static weighting or simple linear mapping mode, because the performance index, the log text and the instruction have essential differences in description granularity, if unified linear weighting is adopted, key signals are often submerged by massive random fluctuation signals, core features representing network oscillation are difficult to accurately extract, on the other hand, in a complex perception topological network, the global state space is extremely huge, the traditional deep learning model is easy to overfit in a single environment, in addition, single-layer reinforcement learning is usually adopted, when massive feature nodes are processed, searching is easy and difficult, model convergence is slow, a fault range is difficult to lock in the environment with fluctuation of network state quality and an effective change strategy is generated, meanwhile, a processing scheme has identification nodes, the generated fusion feature vector and execution logic of a bottom physical node have faults, and on the other hand, due to the lack of an effective feedback compensation mechanism, the system cannot represent the instability deviation of a precise calculation node according to the fused state, the control means often has hysteresis, and automatic inhibition of the network oscillation state is difficult to realize. Disclosure of Invention (One) solving the technical problems Aiming at the defects of the prior art, the invention provides a multi-source heterogeneous data fusion and processing method based on an reinforcement learning mechanism, which is used for acquiring multi-source heterogeneous data of performance indexes, log texts and operation instruction sequences in real time, constructing a perception topological network, presetting a first fusion strategy, realizing primary quantification of a communication network basic situation, introducing a layered reinforcement learning mechanism, wherein an upper reinforcement strategy is responsible for global state evaluation, accurately converging to a topological attention domain from global topological nodes, reducing calculation complexity, and a lower reinforcement strategy is used for generating a change instruction aiming at characteristic items of oscillation states in the topological attention domain, driving the dynamic evolution of the first fusion strategy to a second fusion strategy, and solving the problems in the background technology. (II) technical scheme In order to achieve the above purpose, the invention is realized by the following technical scheme: In a first aspect, the present application provides a method for fusion and processing of heterogeneous multi-source data based on an reinforcement learning mechanism, where the method includes: the method comprises the steps of acquiring multi-source heterogeneous data in a communication network in real time, abstracting and constructing a perceived topological network, and configuring a first fusion strategy for characteristic nodes in the perceived topological network; Introducing a hierarchical reinforcement learning mechanism, and evaluating and analyzing a perceived topological network; The hierarchical reinforcement learning mechanism comprises an upper reinforcement strategy and a lower reinforcement strategy, wherein the upper reinforcement strategy is executed, a topological attention domain is determined through state evaluation, the lower reinforcement strategy is executed in the topological attention domain, a change instruction is generated, the first fusion strategy is changed, and the second fusion strategy is obtained. Further, the multi-source heterogeneous data includes performance indicators, log text, and sequences of operation instructions of the communication network. Further, the abstracting construction of the perceived topology network includes: