CN116055330A - Digital twin network slicing method and device based on knowledge graph - Google Patents
Digital twin network slicing method and device based on knowledge graph Download PDFInfo
- Publication number
- CN116055330A CN116055330A CN202310089901.1A CN202310089901A CN116055330A CN 116055330 A CN116055330 A CN 116055330A CN 202310089901 A CN202310089901 A CN 202310089901A CN 116055330 A CN116055330 A CN 116055330A
- Authority
- CN
- China
- Prior art keywords
- network
- network slice
- knowledge
- triplet
- graph
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 61
- 239000013598 vector Substances 0.000 claims description 89
- 238000013528 artificial neural network Methods 0.000 claims description 54
- 230000006870 function Effects 0.000 claims description 33
- 230000009471 action Effects 0.000 claims description 19
- 230000004927 fusion Effects 0.000 claims description 12
- 230000008569 process Effects 0.000 claims description 11
- 238000012795 verification Methods 0.000 claims 1
- 239000002699 waste material Substances 0.000 abstract description 5
- 238000007689 inspection Methods 0.000 abstract 1
- 230000000875 corresponding effect Effects 0.000 description 21
- 230000002787 reinforcement Effects 0.000 description 21
- 238000004364 calculation method Methods 0.000 description 20
- 238000010586 diagram Methods 0.000 description 11
- 238000004873 anchoring Methods 0.000 description 8
- 102100033591 Calponin-2 Human genes 0.000 description 6
- 101000945403 Homo sapiens Calponin-2 Proteins 0.000 description 6
- 238000013527 convolutional neural network Methods 0.000 description 6
- 102100033620 Calponin-1 Human genes 0.000 description 5
- 101000945318 Homo sapiens Calponin-1 Proteins 0.000 description 5
- 238000010276 construction Methods 0.000 description 4
- 238000001514 detection method Methods 0.000 description 4
- 238000012546 transfer Methods 0.000 description 4
- 238000004891 communication Methods 0.000 description 3
- 230000003993 interaction Effects 0.000 description 3
- 239000000203 mixture Substances 0.000 description 3
- 238000012549 training Methods 0.000 description 3
- 102100023855 Heart- and neural crest derivatives-expressed protein 1 Human genes 0.000 description 2
- 102100038654 Heart- and neural crest derivatives-expressed protein 2 Human genes 0.000 description 2
- 101000905239 Homo sapiens Heart- and neural crest derivatives-expressed protein 1 Proteins 0.000 description 2
- 101001031594 Homo sapiens Heart- and neural crest derivatives-expressed protein 2 Proteins 0.000 description 2
- 101150039208 KCNK3 gene Proteins 0.000 description 2
- 230000002159 abnormal effect Effects 0.000 description 2
- 238000006243 chemical reaction Methods 0.000 description 2
- 125000004122 cyclic group Chemical group 0.000 description 2
- 238000013075 data extraction Methods 0.000 description 2
- 230000006872 improvement Effects 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 238000013468 resource allocation Methods 0.000 description 2
- 239000007787 solid Substances 0.000 description 2
- 238000001228 spectrum Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 230000002596 correlated effect Effects 0.000 description 1
- 238000013499 data model Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000005265 energy consumption Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000000977 initiatory effect Effects 0.000 description 1
- 230000002452 interceptive effect Effects 0.000 description 1
- 238000002372 labelling Methods 0.000 description 1
- 238000010295 mobile communication Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000006855 networking Effects 0.000 description 1
- 238000003062 neural network model Methods 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 238000003672 processing method Methods 0.000 description 1
- 230000008439 repair process Effects 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
Images
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
- H04L41/08—Configuration management of networks or network elements
- H04L41/0893—Assignment of logical groups to network elements
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/36—Creation of semantic tools, e.g. ontology or thesauri
- G06F16/367—Ontology
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
- H04L41/16—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks using machine learning or artificial intelligence
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Databases & Information Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Signal Processing (AREA)
- Computer Networks & Wireless Communication (AREA)
- Software Systems (AREA)
- Animal Behavior & Ethology (AREA)
- Medical Informatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention belongs to the technical field of network slicing, and particularly relates to a digital twin network slicing method and device based on a knowledge graph. Comprising the following steps: constructing a thing set of a data center, and constructing an initial triplet composed of things, relations and things according to the thing set; constructing a knowledge graph of each data center by using the initial triples meeting the inspection conditions in each data center and the triples established according to the prior art, and constructing a knowledge graph of the cross-domain data center by fusing the knowledge graphs of all the data centers once; predicting deployment positions of the network slice VNF and the network slice PNF according to the network running state of the network slice, transmitting deployment decisions back to the knowledge graph, and adjusting the network running state according to the deployment decisions by the knowledge graph. The invention can realize dynamic reasonable deployment of the network slice PNF and the network slice VNF and reduce waste of network resources.
Description
Technical Field
The invention belongs to the technical field of network slicing, and particularly relates to a method and a device for digital twin network slicing based on a knowledge graph.
Background
With the development of 5G, the speed and the bandwidth of the network are improved, the cloud network has more facing services, wide demand range, complex network environment and difficult network slicing deployment, so that the network networking is more complex, and the network functions are more; therefore, the original network slicing algorithm based on load balancing and minimum time delay is reduced in calculation accuracy, and the existing complex network planning cannot be processed. On the other hand, the physical device integrates more functional modules.
When each network element of different equipment providers and different communication providers bear different service operations, different hardware devices and different operating systems are adopted. The utilization rate of the physical equipment and the virtual equipment can be periodically or irregularly changed along with the service change and the different using time periods, so that the peak load requirement of a certain single service is often not met at present, and the physical resources of part of network elements are accumulated and the physical resources of part of network elements are extremely stressed.
At present, most of research on network slicing focuses on the deployment of the network slicing in terms of virtual network functions (virtualized network function, abbreviated as VNFs), but ignores the situation that network operation faults on the network slicing are caused by unreasonable allocation of physical network functions (physical network function, abbreviated as PNFs) resources and waste of a large amount of physical resources.
In view of this, overcoming the drawbacks of the prior art is a problem to be solved in the art.
Disclosure of Invention
Aiming at the above defects or improvement demands of the prior art, the invention provides a method and a device for digital twin network slicing based on a knowledge graph, which aim to construct the knowledge graph of a data center in a triplet form, and reasonably base on the PNF and the VNF of the digital twin network slicing of the knowledge graph according to different interactive learning of the knowledge graph and a reinforcement learning module, thereby solving the technical problems of network operation faults and resource waste on the network slicing caused by unreasonable resource allocation of the PNF and the VNF.
To achieve the above object, according to one aspect of the present invention, there is provided a method for digital twin network slicing based on knowledge-graph, specifically including: constructing a thing set of a data center, and constructing an initial triplet composed of things, relations and things according to the thing set; checking the rationality and the existence of the initial triples, constructing a knowledge graph of each data center by using the initial triples meeting the checking conditions in each data center and the triples established according to the prior art, and fusing the knowledge graphs of all the data centers at one time to construct a knowledge graph of a cross-domain data center; and representing the network operation state of the network slice by the knowledge graph of the cross-domain data center, predicting the deployment positions of the network slice VNF and the network slice PNF according to the network operation state of the network slice, transmitting the deployment decision back to the knowledge graph, and adjusting the network operation state by the knowledge graph according to the deployment decision.
Preferably, the specific method for constructing the things set of the data center includes: selecting keywords and conventional words from the text data, converting the voice data or the picture data into corresponding text data, and then selecting the keywords and the conventional words; calculating the character distance between the conventional word and the keyword, and comparing to obtain the minimum character distance between the conventional word and the keyword; selecting the key words as keys, selecting the conventional words corresponding to the key words in the minimum character distance as sting, and constructing a key-sting set; the transaction set of the data center includes at least one key-sting set.
Preferably, the method constructs an initial triplet composed of things, relations and things according to the things set, and specifically includes: selecting a first object and a second object in the key-sting set, marking the first object and the second object on text, voice or picture data, and presetting a first relation between the first object and the second object; identifying, by a neural network, the first thing, the second thing, and the first relationship according to the annotation; judging whether the preset first relation is established or not according to a loss value generated in the neural network identification process; if the relationship is established, the first thing, the second thing and the first relationship construct an initial triplet.
Preferably, the neural network identifies the first thing, the second thing and the first relation according to the label, and the specific method includes: the neural network identifying annotations for the first thing and the second thing, generating a first loss value; the neural network identifies the first relationship and generates a second loss value; and obtaining a neural network identification loss value according to the first loss value and the second loss value.
Preferably, the method for determining whether the preset first relationship is established according to the loss value generated in the neural network identification process includes: presetting a loss threshold value, wherein the loss threshold value is obtained by multiplying or adding the first loss value and the second loss value; if the neural network identification loss value is smaller than the loss threshold value, the first relation is established; and if the neural network identification loss value is greater than or equal to the loss threshold value, the first relation is not established.
Preferably, the method further includes determining whether the preset first relationship is established according to the loss value generated in the neural network identification process, and the method further includes: and if the first relation is not established, sequentially changing the first relation from a relation set preset in a data center, and judging whether the changed first relation is established or not by the neural network identification loss value.
Preferably, the method includes using an initial triplet of each data center satisfying a test condition and constructing a knowledge graph of each data center according to a well-known established triplet, and determining whether a distance between a relationship center vector in the triplet and a relationship center vector in the initial triplet exceeds a fusion distance threshold; and if the distance between the relationship center vector in the triplet and the relationship center vector in the initial triplet exceeds the fusion distance threshold, fusing the triplet with the initial triplet.
Preferably, the predicting the deployment positions of the network slice VNF and the network slice PNF according to the network operation state of the network slice specifically includes: the network running state of the network slice specifically comprises the position, the resource and the relation information of the network slice PNF and the network slice VNF; the network operation states of the network slice PNF and the network slice VNF are represented by triples of the knowledge spectrogram, and the network operation states of the network slice PNF and the network slice VNF are converted into a triplet vector representation; if the triplet vector of the network slice VNF is simultaneously smaller than the triplet vector of the network slice PNF, the network slice VNF is deployed above the network slice PNF; and if the triplet vector of the network slice VNF is not smaller than the triplet vector of the network slice PNF, replacing the network slice PNF until the network slice VNF is deployed on the network slice PNF.
Preferably, the deployment decision is returned to the knowledge graph, and the knowledge graph adjusts the network operation state according to the deployment decision, and the specific method comprises the following steps: and forming an action set according to the number of the network slice PNFs and the network slice VNs, setting a feedback function of a network running state for the network slice which is completely deployed, and adjusting the knowledge graph according to the action set and the feedback function.
According to another aspect of the present invention, there is provided a device for slicing a digital twin network based on a knowledge-graph, wherein the device comprises: at least one processor; and a memory communicatively coupled to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being programmed to perform the method of knowledge-graph based digital twin network slicing provided in the first aspect.
In general, compared with the prior art, the above technical solution conceived by the present invention has the following beneficial effects:
according to the invention, the knowledge graph of the data center is constructed in the form of the triples, the network operation state of the network slice is expressed in the form of the knowledge graph, the network operation state of the network slice is defined by the reinforcement learning module, the network operation state of the network slice is expressed in the form of the triples, the reinforcement learning module is combined with the triples of the knowledge graph to continuously interact, so that dynamic reasonable deployment of the network slice PNF and the network slice VNF is realized, and the waste of network resources is reduced.
Drawings
Fig. 1 is a schematic flow chart of a method for providing a digital twin network slice based on a knowledge-graph according to the first embodiment;
FIG. 2 is a schematic diagram of a transaction flow diagram for constructing a data center according to one embodiment;
FIG. 3 is a schematic diagram of a process for constructing an initial object-relationship-object triplet according to the object set according to an embodiment;
FIG. 4 is a schematic diagram of labeling a first object and a second object according to one embodiment;
FIG. 5 is a schematic diagram of a neural network identifying a first object and a second object and a relationship between the first object and the second object;
FIG. 6 is a schematic diagram of a neural network for determining initial status triad rationality and existence in accordance with an embodiment;
FIG. 7 is a schematic diagram of a user initiating a network slice request to a control layer;
fig. 8 is a schematic diagram of a knowledge graph structure provided in the third embodiment;
FIG. 9 is a schematic diagram of triples used in a knowledge graph provided in embodiment three;
fig. 10 is a schematic diagram of another knowledge graph structure provided in embodiment three;
fig. 11 is a schematic diagram of an apparatus for knowledge-graph-based digital twin network slicing according to the fourth embodiment.
Detailed Description
The present invention will be described in further detail with reference to the drawings and examples, in order to make the objects, technical solutions and advantages of the present invention more apparent. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the invention. In addition, the technical features of the embodiments of the present invention described below may be combined with each other as long as they do not collide with each other.
Embodiment one:
at present, most of research on network slicing focuses on the deployment of the network slicing in terms of virtual network functions (virtualized network function, abbreviated as VNFs), and ignores the situation that network operation faults on the network slicing are caused by unreasonable resource allocation of physical network functions (physical network function, abbreviated as PNFs) and waste of a large amount of physical resources.
A Knowledge Graph (KG) is a net-shaped Knowledge base, and is composed of "entity-relationship-entity" triples, and entity and its related attribute-value pairs. Knowledge maps describe concepts, entities and their relationships in the objective world in such structured form, providing a better ability to organize, manage and understand information.
The first embodiment provides a method for slicing a digital twin network based on a knowledge graph, as shown in fig. 1, the method includes the following steps:
s101, constructing a thing set of a data center, and constructing an initial triplet composed of things, relations and things according to the thing set.
The embodiment of the invention provides three network triplet states and evolution. The triplets (things 1, relations, things 2) in the network slicing platform are based on the relations as cores, the things are auxiliary, and the relations among different things are connected in series through the relations. The triplet states are divided into: initial state, checking state and establishing and completing three states. An initial state triplet refers to a triplet that is built from a single data center thing set that has not yet been rationalized. The check state triplets refer to triplets in which a vector representation of things and relationships is obtained in the same vector space after the initial state triplets are input into the anchor network, and the vector representation is performed. The triplet to establish the completion state includes two parts: one part is a triplet constructed according to general knowledge and expert experience which have been determined and accepted by the communication profession, and the other part is a triplet which is output after the check state triplet passes through the decision network.
The specific building process of the triples is as follows:
(1) The triplets are established based on the determined and accepted general knowledge and expert experience of the communication profession for training the anchoring network for making network slicing decisions, and are obtained from the determined and accepted general knowledge and expert experience, so that the triplets are in an established completion state, being completion state triplets.
(2) And establishing triples according to the actual conditions of the network, wherein the data and the relations in the triples can be obtained by collecting text, voice and image data containing the same data center. The voice data are converted into corresponding text data through an acoustic model and a language model, and the image data are converted into corresponding text data through a picture content recognition and image understanding technology. And then counting the high-frequency words in the text data to construct a thing set of a single data center. The method is characterized in that a category relation is used as a core at first, a thing set of the same noun is built, and then other relations are used as cores step by step for building. The construction method comprises calculating average distance between rest nouns and selected nouns (keys) in text, setting a word occupation distance as 1, average distance as average value of sum of distance between the word occupation distance and nearest key word, and finally counting thring corresponding to the smallest average distance of the nouns as key to construct thing set (key) i -(thing i1 ,thing i2 ,..), then building initial state triples from these sets of things.
(3) And judging the rationality and the existence of the ternary combination in the initial state according to the completed ternary combination. The judgment can be made by a two-layer neural network composed of an anchor network and a judgment network as shown in fig. 5. Firstly, training an anchoring network through a first layer of anchoring network by using the completion status triplet established in the step (1), and adjusting parameters of the anchoring network to enable the parameters to meet t1_vector+relation_vector=t2_vector. Then the initial state triplet established in the step (2) is passed through the first layer anchoring network which is completed by training, so as to obtain the initial state tripletVectorization means that the initial state triplet evolves into a check state triplet at this time. The second layer of judging network judges the rationality and the existence of the ternary combination in the checking state, and each judging task only judges the rationality and the existence of the ternary combination under one relation. The decision rule is as follows: as shown in fig. 6, the set-up completion state triplet in step (1) and the initial state triplet in step (2) are vectorized in the same vector space by the anchor network, respectivelyOne of the relationships R is selected at each decision K By R K Is a shaft (or a rope)>A circle is formed for 360 degrees of edge rotation, and the circle range is the judging range. Selecting the same relationship r k (r k =R K ) By r k Is a shaft (or a rope)>Forming a circle for 360 degrees of edge rotation, (1) if the circle is completely within the decision rangeSatisfy R K The triples under the relationship make up the condition, but are triples that already exist in the build completion state; (2) if this circle intersects with the decision range but does not completely coincide, then +.>Satisfy R K The triples form conditions under the relation and are a new triplet; (3) if this circle has no intersection with the decision range +.>Not meeting R K Triplet composition condition under relation,/->The triplet is unreasonable. And (3) outputting the triples in the step (2) after the judgment is carried out by the judgment network according to the rules, wherein the triples after the output are the triples in the established state, and the triples in the step (1) form a knowledge graph of the single data center. For example, selecting keywords and regular words in the literal data;
calculating the character distance between the conventional word and the keyword, and comparing to obtain the minimum character distance between the conventional word and the keyword;
selecting the key words as keys, selecting the conventional words corresponding to the minimum character distance as sting, and constructing a key-sting set;
The transaction set of the data center includes at least one key-sting set.
For example, a first object and a second object in the object set are selected, the first object and the second object are marked on the voice class and picture class data, and the relation between the first object and the second object is preset, so that an initial object-relation-object triplet is constructed.
S102, checking the rationality and the existence of the initial triples, constructing a knowledge graph of each data center by using the initial triples meeting the checking conditions in each data center and the triples established according to the prior art, and fusing the knowledge graphs of all the data centers at one time to construct a knowledge graph of the cross-domain data center.
The new triples are fused by the initial object-relation-object triples to construct the knowledge graph of the data center, the knowledge graph of the data center is continuously updated after the new triples are fused in the initial object-relation-object triples, and if the new triples cannot be fused, the original knowledge graph is considered to contain new triples information.
After each single data center knowledge graph is gradually constructed, the single data center knowledge graphs are sequentially fused to construct a cross-domain data center knowledge graph. The rule for judging whether the triples in the knowledge graph are fused is as follows: for the knowledge graph to be fused and the fused knowledge graph, 1) judging whether a new relation is generated in the fused knowledge graph, if so, all triples in the new relation are new triples, and directly fusing the new triples into the knowledge graph to be fused. 2) When the relation in the fused knowledge-graph is existing in the knowledge-graph to be fused, (1) respectively calculating the center vectors of the triples under the relation in the two knowledge-graphs, (2) then calculating the Hamming Distance (Hamming Distance) between the center vectors of the triples under the relation in the knowledge-graph to be fused, selecting the minimum and maximum Hamming distances to obtain a judging range for judging whether the triples in the knowledge-graph to be fused are new triples, and (3) calculating the Hamming Distance between the center vector of the triples under the relation in the fused knowledge-graph and the center vector of the triples under the relation in the knowledge-graph to be fused, if the Hamming Distance is within the judging range, judging that the triples under the relation in the fused knowledge-graph are the existing triples not fused, and if the Hamming Distance is not within the judging range, judging that the triples under the relation in the fused knowledge-graph are new triples, and fusing the triples into the knowledge-graph to be fused.
S103, representing the network operation state of the network slice by the knowledge graph of the cross-domain data center, predicting the deployment positions of the network slice VNF and the network slice PNF according to the network operation state of the network slice, transmitting the deployment decision back to the knowledge graph, and adjusting the network operation state by the knowledge graph according to the deployment decision.
The knowledge graph defines the network operation state of the network slice, the knowledge graph outputs triplet information about the network operation state of the network slice to the reinforcement learning module, and the reinforcement learning module judges whether PNF/VNF resources are excessive or not according to the triplet information of the network slice PNF/VNF, and whether the operation state is normal or not, so that the positions where the network slice PNF/VNF are deployed respectively are predicted.
The reinforcement learning module transmits deployment decisions of the network slice PNF/VNF back to the knowledge graph, the knowledge graph adjusts the network operation state according to the deployment decisions, the network operation state after the adjustment of the knowledge graph is transmitted to the reinforcement learning module, the reinforcement learning module calculates the adjusted network operation state, and the corresponding network operation state decisions are selected.
In the first embodiment, in order to obtain the object set of the data center, in combination with the embodiment of the present invention, there is also a preferred implementation scheme, specifically, as shown in fig. 2, the specific method for constructing the object set of the data center includes the following steps:
S201, selecting keywords and conventional words from the text data, converting the voice data or the picture data into corresponding text data, and then selecting the keywords and the conventional words.
For example in the following text:
the server is connected with a 5800 exchanger through an electric port, the 5800 exchanger is used as an access exchanger to be connected with a convergence exchanger 6800, one 6800 convergence exchanger controls a plurality of 5800 exchangers, and a plurality of convergence exchangers 6800 are controlled by a core exchanger 6800E to jointly form a small data center. "
And counting according to the text and book data to obtain high-frequency nouns in the upper text, wherein a server, a switch and the like appearing in the upper text are used as keywords.
And counting according to the text and book data to obtain conventional words in the upper text, such as ' electric port ', ' 5800 ', access ' and the like in the upper text.
S202, calculating the character distance between the conventional word and the keyword, and comparing to obtain the minimum character distance between the conventional word and the keyword.
For example, in the phrase "the server is connected to the 5800 switch through the electric port", the distance character between the "server" and the "switch" is 9, the distance character between the "electric port" and the "switch" is 5, and the distance character between the "5800" and the "switch" is 0.
And calculating the average distance between the conventional word and the keyword (switch) in the upper text, wherein the average distance is equal to the average value of the sum of the distances between the conventional word and the keyword (switch), and comparing to obtain the minimum character distance.
S203, selecting the key words as keys, selecting the conventional words corresponding to the key words in the minimum character distance as sting, and constructing a key-sting set.
Finally, the nouns '5800', '6800E' corresponding to the smallest distance among the nouns are counted as the sting, the switch is taken as the key, and a key (switch) -sting (5800, 6800E) set is constructed.
The transaction set of the data center comprises at least one key-sting set, and a plurality of key-sting sets form the transaction set of the data center.
In the first embodiment, in order to obtain the object-relationship-object triplet of the data center, in combination with the embodiment of the present invention, there is a preferred implementation scheme, specifically, as shown in fig. 3, the specific method includes:
s301, selecting a first object and a second object in the key-sting set, marking the first object and the second object on text, voice or picture data, and presetting a first relation between the first object and the second object.
As shown in fig. 4, the collected voice and picture class data is first labeled based on key (switch) -sting (5800, 6800E) set. The solid and dashed circles are denoted (R_switch, G_switch), R_switch represents the switch in the solid circle, G_switch represents the switch in the dashed circle, and R_switch, G_switch correspond to two THING in the key (switch) -THING (5800, 6800E) set.
S302, identifying the first thing, the second thing and the first relation by a neural network.
Secondly, constructing a small knowledge graph according to THING1 and THING2, mixing a picture and a section of voice video, passing the same picture through different neural networks such as CNN1 and CNN2, and identifying intermediate values (y) of THING1 and THING2 under keys (such as servers, firewalls, switches, etc.) 1 ,y 2 ). Taking fig. 5 as an example, fig. 4 is input into CNN1 and CNN2 shown in fig. 5, where in CNN1, a given tag is g_switch, and the object is to identify a portion in the virtual loop in fig. 4, and in CNN2, a given tagLabeled r_switch, the goal is to identify the portion of the real coil in fig. 4.
The first relation between the first object and the second object is preset, and n relations, such as "own", "located", "related", "interconnected", "connected", etc., are preset in relation set in the data center. In DNN1, the goal is to identify the intermediate value y of the "connected" relationship in FIG. 4 in the voice, picture class information 3 . Further through DNN2, the goal is to identify whether the "connected" relationship between the switches in the real and imaginary circles in fig. 4 holds.
S303, judging whether the preset first relation is established or not according to a loss value generated in the neural network identification process.
Because the neural network can generate loss when recognizing the relation among the first object, the second object and the first object and the second object in the voice class and picture class data, if the loss value is smaller than a preset threshold value, the neural network recognition accuracy is high, and the authenticity of the establishment of the relation is ensured. The neural network includes one or more of CNN, DNN, or RNN.
S304, if the relation is established, the first object, the second object and the first relation construct an initial triplet. Different things and different relations are respectively represented by different vectors.
In a first embodiment of the present invention, in order to calculate a neural network identification loss value, in combination with the embodiment of the present invention, there is also a preferred implementation scheme, specifically, the neural network identifies the first object, the second object, and the first relationship according to the label, and the specific method includes:
The neural network identifies annotations for the first thing and the second thing, generating a first loss value.
First loss value error 0 The calculation formula of (2) is as follows:
wherein: i represents the ith, then, j represents the jth element in the Ti vector, y represents the intermediate quantity, T represents then, n represents the vector having n dimensions.
And generating a second loss value when the neural network identifies the first relation.
Wherein things, relationships and intermediate values (y 1 ,y 2 ) Represented by vectors, each vector has n dimensions, and things, relationships, and intermediate values are represented by vectors as: t (T) 1 (t 11 ,t 12 ,...,t 1n ),T 2 (t 21 ,t 22 ,...,t 2n ),R 1 (r 11 ,r 12 ,...,r 1n ),y 1 (y 11 ,y 12 ,...,y 1n ),y 2 (y 21 ,y 22 ,...,y 2n ). And then the same principle that a section of marked voice passes through DNN1 network shown in FIG. 5, a relation (Relationship R) contained between things in the voice is recognized 1 ) If so, only one relationship is identified per task. Then put T again 1 、T 2 、R 1 After passing DNN2 shown in FIG. 5, the output is T 1 ',T 2 ',R 1 '. Judgment (T) 1 ',T 2 ',R 1 ') whether a triplet is formed.
Second loss value error 1 The calculation formula of (2) is as follows:
wherein: t1 represents a first trailing vector; t2 represents a second lying vector; r1 represents a first relationship vector; t1i represents the i-th element in the first trailing vector; t2i represents the i-th element in the first lying vector.
In a first embodiment of the present invention, in order to determine whether a relationship between a first object and a second object is preset so as to facilitate a later establishment of a triplet, in combination with an embodiment of the present invention, there is also a preferred implementation scheme, specifically, the method for determining whether the first relationship is preset by using a neural network recognition loss value includes:
And presetting a loss threshold value, wherein the loss threshold value is obtained by multiplying or adding the first loss value and the second loss value.
In the first embodiment, the neural network identification loss value is obtained by multiplying the first loss value by the second loss value.
And if the neural network identification loss value is smaller than the loss threshold value, the first relation is established.
If the neural network identification loss value is set to 0.5, a first loss value error 0 And error 1 When the product result of (2) is 0.3, the predetermined relationship between the first object and the second object is established.
And if the neural network identification loss value is greater than or equal to the loss threshold value, the first relation is not established.
If the neural network identification loss value is set to 0.5, a first loss value error 0 And error 1 When the product result of (2) is 0.8, the predetermined relationship between the first object and the second object is established.
In a first embodiment, in order to further accurately determine whether a preset relationship between a first object and a second object is established, in combination with an embodiment of the present invention, there is also a preferred implementation scheme, specifically, the method further includes:
And if the first relation is not established, sequentially changing the first relation from a relation set preset in a data center, and judging whether the changed first relation is established or not by the neural network identification loss value.
If the predetermined relationship between the first object and the second object is not established, sequentially changing the relationship in the relationship set of the data center, changing the "connection" to be "owned", "located", "correlated", "interconnected", etc., and identifying a loss value (error) in the neural network 0 *error 1 ) Under the calculation condition of (1), corresponding model parameters in the neural network model are adjusted to obtain a knowledge-graph triplet (T) 1 ',R 1 ',T 2 ')。
For example, in order to recognize that the preset relationship is "connected" from the segment of speech, after recognition by the neural networks CNN1 and CNN2 in fig. 4, r_switch and g_switch are obtained respectively, and then the relationship between r_switch and g_switch is determined to be connected after recognition by the DNN 2. And forming a triplet (R_switch, connect, G_switch) so as to construct an initial triplet of things-relation-things of the data center, namely an initial knowledge graph.
In a first embodiment, in order to construct knowledge maps of a plurality of data centers, in combination with the embodiment of the present invention, there is also a preferred implementation scheme, specifically, the fusing the new triples generated by the object set to the initial triples, and constructing a knowledge map of the data center, where the specific method includes:
And judging whether the distance between the relationship center vector in the new triplet and the relationship center vector in the initial triplet exceeds a fusion distance threshold.
And if the distance between the relationship center vector in the new triplet and the relationship center vector in the initial triplet exceeds the fusion distance threshold, fusing the new triplet and the initial triplet.
The preset fusion distance threshold is a preset distance which can be fused by the relation center vector in the initial triplet, and if the distance exceeds the fusion distance threshold, the new triplet is fused with the initial triplet;
if the distance does not exceed the fusion distance threshold, the new triplet is not fused with the initial triplet.
And fusing new data center knowledge on the original basic knowledge graph. In D 2 The data center is exemplified by firstly collecting various data of the data center and constructing a knowledge-graph tripletAccording to the relationship->Sorting the confirmed categories, dividing the original data center +.>Triplet of relations, constituting vector->Computing +.>Vector R in relation j Center vector of +.>And wherein the distance is from the center vector distance. For example->And calculating the center vector in the connection relation triplet in the original data center for the connection relation, wherein the calculation mode of the ith element in the center vector is as shown in the formula (3).
If new data center triplesCorresponding vector->And->And if the distance between the two triples is larger than the fusion distance threshold value, the new data center triples are considered to be fused into the knowledge graph, otherwise, the original knowledge graph is considered to contain the information of the new data center triples, and the two triples are not fused. Obtaining a knowledge graph triplet according to calculation>Thereby constructing a knowledge graph of the data center.
In a first embodiment, in order to deploy a network slice according to a network operation state, in combination with the embodiment of the present invention, there is also a preferred implementation scheme, specifically, the network operation state of the network slice is represented by a knowledge graph of the data center, and the deployment positions of the VNF and PNF of the network slice are predicted according to the network operation state of the network slice, where the specific method includes:
the network operation state of the network slice specifically includes location, resource and relationship information of the network slice PNF and the network slice VNF.
The following definitions are given first:representing the network slice PNF in the physical cluster layer,representing links between network slice PNF and network slice PNF in physical cluster layer, e.g. +.>Representing network slice- >And network slice->And links between them. />Representing network components in the physical cluster layer, at network slice S s In (I)>Representing a network slice VNF node,/->Representing link connections between network slice VNF nodes. For example->Representing network slice->And network slice->And links between them.
The network operational states of the network slice PNF and the network slice VNF are characterized by triples of the knowledge spectrogram and are converted into a triplet vector representation.
For network slicingThe resources of the computing, storage, network of (a) are denoted +.> VNF node in network slice>The calculation, storage, network resources of (1) are denoted +.>Assume a total of b network slices { S ] 1 ,S 2 ,..S b )。
Network operating states are defined. Defining network operation state according to availability and resource condition of network slice PNF and network slice VNF, specifically selecting placement positions of network slice PNF and network slice VNF simultaneously when network slice is performed, ordering network slice PNF according to spatial position relationship, and assuming total k network slice PNFs (PNF) 1 ,PNF 2 ,...,PNF k ). r network slice VNFs (VNFs) 1 ,VNF 2 ,...,VNF r ). The initial states of the network slice PNF and the network slice VNF consist of four parts, the first part represents the sequence number of the previous network slice PNF or the network slice VNF, the first bit of the network slice PNF or the network slice VNF is 0, the second part represents the sequence number of the network slice PNF or the network slice VNF, and the second part represents 1. The third partial table network slices PNF or VNF calculates, stores, network resources. For a network slice PNF or a network slice VNF,1 indicates that there is a remaining computing/storage/network resource, and 0 indicates that it is exhausted. The fourth part represents the operational status of the network slice PNF or the network slice VNF, normal operation as 1, abnormal operation as 0.
For example, when there are 2 PNFs and 3 VNFs, where VNF1 and VNF2 are deployed on PNF1 and VNF3 is deployed on PNF2, the two PNFs operate normally and the resources are sufficient. The network operational state of each network slice is represented in binary form, as shown below, to facilitate later deployment of the set of actions and feedback functions for the network operational state.
PNF1 first part: 00 second section: 01 third section: 11 fourth section: 11, pnf1 initial state is 00011111;
PNF2 first part: 01 second part: 10 third section: 11 fourth section: 11, pnf2 initial state is 01101111.
VNF1 first part: 00 second section: 01 third section: 11 fourth section: 11; VNF1 initial state is 00011111;
VNF2 first part: 01 second part: 01 third section: 11 fourth section: 11; VNF2 initial state is 01011111;
VNF3 first part: 10 second part: 10 third section: 11 fourth section: 11.VNF3 initial state is 10101111.
Each network slice PNF or network slice VNF reliability is measured by average dead time (Mean Time To Failure, abbreviated MTTF), serviceability is measured by average repair time (Mean Time To Restoration, abbreviated MTTR), PNF (VNF) availability a PNF /A VNF As shown in equation (4).
The availability of the whole slice is equal to the minimum value of the availability of PNF or VNF constituting the network sliceThe network slice PNF and network slice VNF positional relationship is represented.
The expression of the formula (5) is given,the value of (2) is 0 or 1, if +.>At->On the other hand, the flow of qi is->Otherwise, 0.
Thus atThe triplet vector of locations/resources/relationships should satisfy the following conditions:
equation (6) indicates that the network slice VNF is deployed on top of the network slice PNF if the triplet vector of the network slice VNF is simultaneously smaller than the triplet vector of the network slice PNF.
The reinforcement learning module deploys a network slice VNF on top of a network slice PNF.
After the knowledge graph of the data center is built, all information of the physical clusters and the virtual clusters in the network can be obtained. And continuously selecting and extracting network knowledge from the knowledge graph based on the reinforcement learning module so as to promote the overall network state.
And if the triplet vector of the network slice VNF is not smaller than the triplet vector of the network slice PNF, replacing the network slice PNF until the network slice VNF is deployed on the network slice PNF.
The reinforcement learning module deploys the network slice VNF on top of the network slice PNF that meets the requirements of equation (6).
In the first embodiment, in order to continuously select a more suitable deployment position of a network slice according to a network operation state, the deployment decision is returned to the knowledge graph, and the knowledge graph adjusts the network operation state according to the deployment decision, and the specific method includes:
and forming an action set according to the number of the network slice PNFs and the network slice VNs, setting a feedback function of a network running state for the network slice which is completely deployed, and adjusting the knowledge graph according to the action set and the feedback function.
After the placement positions of the network slice PNF and the VNF are selected, a cyclic interaction process is continued between the reinforcement learning module and the knowledge graph, and the cyclic interaction process specifically comprises an action set and a feedback function.
A set of actions. And selecting one of the PNF/VNF deployment for the currently deployable PNF/VNF according to the state information of each moment. Each action in each action set corresponds to a node and, assuming that there are N PNFs and VNFs to be deployed according to the network slice, the action set is denoted as a= (a) 1 ,A 2 ,...,A N ). Action set each action represents a single adjustment to the deployment location of a network slice, with only one PNF or VNF network slice placed per deployment.
And (5) a feedback function. Different feedback values are set for the network operation states of the network slice VNF and the network slice PNF after placement, for example, when the resource usage of the VNF selection node exceeds the PNF resource limit, the feedback value is set to be-100, and when the VNF and the PNF are abnormal in operation, the feedback value is set to be-100. When the VNF and the PNF are normal in operation, a feedback function is established for the network at the time t and the network state at the time t+1 after the current action is executed. The feedback function is proportional to the availability of the network, load balancing, inverse proportion to the delay. The feedback value of this action is higher when the network availability is higher, the load is more balanced, and the delay is lower. Wherein the calculation of the feedback function may be fitted with a neural network, wherein the neural network structure may be as shown in fig. 5. Fig. 5 shows only one example, and other neural network structures may be used.
If the feedback function shows that the state of the network slice is good, the action set transmits actions to the knowledge graph, the knowledge graph adjusts the triples, the knowledge graph transmits the triples of the running state of the network slice to the reinforcement learning module after the adjustment is completed, and the reinforcement learning module makes a series of deployment decisions of the network slice.
And after multiple times of adjustment, calculating an optimal PNF and VNF deployment scheme, and finishing the optimal slicing of the network.
As shown in fig. 7, a user initiates a network slice request to a control layer, and the control layer module outputs a placement decision of the network slice according to the request description.
Virtual clusters (VNFs) and physical clusters (PNFs) respond to decisions and check available resources, reserving the available resources and feeding them back up to the control layer.
The control layer feeds back the current network state and available resources of the network slice to the knowledge graph module, the knowledge graph module constructs a current network knowledge graph, the current network knowledge graph comprises the network running state of the network slice, and the user requirements and the network knowledge graph information are transmitted to the reinforcement learning module.
The reinforcement learning module establishes a network running state and an action set based on the network triples transmitted by the knowledge graph module to make a series of decisions of network slicing, and transmits the decisions to the knowledge graph, and the knowledge graph adjusts the network state according to different decisions and transmits the adjusted network state back to the reinforcement learning module.
The reinforcement learning module evaluates the adjusted network state through a feedback function and a neural network, selects a decision corresponding to the best network state, and transmits the selected decision information to the control module.
Embodiment two:
in order to realize simultaneous deployment of the VNF and the PNF and better resource utilization rate, the patent provides a construction method of a network knowledge graph and realizes end-to-end slicing of the network based on the knowledge graph. A Knowledge Graph (KG) is a Knowledge database that integrates data using a graphical structured data model or topology. The basic element is a triplet, which represents the relationship between the first thing and the second thing. In a first embodiment, the first thing, the second thing, and the first relation may be regarded as a triplet in the indication map.
In the method provided in the first embodiment, a knowledge graph of the current network state and resources can be constructed according to the physical clusters and the virtual clusters, assuming that there are a plurality of data centers (D 1 ,D 2 ,..) to form a physical cluster and a corresponding virtual cluster, and the steps for constructing the knowledge graph are as follows:
(1) Construction of Single Domain data center D 1 Knowledge graph of (2)
(1) And defining a related entity set E and a related relation set R according to the literal data such as the technical document. And forming a model layer of the knowledge graph by using the entity set.
(2) Constructing instance set { key ] according to literal data i -(thing i1 ,thing i2 ,...),key i E, calculating the average distance between other nouns in the text data and key words, setting the occupied distance of one text as 1, setting the average distance as the average value of the sum of the distances between the word and the nearest key word, and finally counting that the noun with the smallest distance is the corresponding thong of the key words.
(3) According to the relation between the voice and picture data extraction examples, obtaining<Example 1, relationship, example 2>The triples thus construct a knowledge graph. First, the collected voice and picture data are marked, and two examples (hiking) are carried out each time 1 :T 1 ,thing 2 :T 2 ) The identification is performed, and the corresponding label is (y 1 ,y 2 ). As shown in fig. 2, identifying the marked picture data through different CNN networks, and the loss function of Task1 is:
wherein each vector has n dimensions, T 1 (t 11 ,t 12 ,...,t 1n ),T 2 (t 21 ,t 22 ,...,t 2n ),y 1 (y 11 ,y 12 ,...,y 1n ),y 2 (y 21 ,y 22 ,...,y 2n ). Similarly, a section of marked voice passes through DNN1 network to identify that the voice contains a certain relation R 1 (r 11 ,r 12 ,...,r 1n ) Only one relationship is identified for each task. Then T is taken up 1 、T 2 、R 1 After passing through the same DNN2 network, the output is T 1 '、T 2 '、R 1 '. Final judgment (T) 1 ',T 2 ',R 1 ') whether a triplet is formed, the Task2 loss function is:
in minimizing the loss function (error 0 *error 1 ) Under the condition of (1), parameters in CNN and DNN models are adjusted to obtain a knowledge-graph triplet (T) 1 ',R 1 ',T 2 '), thereby constructing a data center D 1 Network knowledge graph.
(2) Constructing a knowledge graph of a cross-domain data center
And (3) constructing knowledge graphs of the rest data centers according to the step (1). In D 2 For example, data center, choose D 2 Triplet with relation of R in knowledge graph of data centerBy vectors->Representing the calculation of the center vector of these triples +. >The calculation formula of the kth element in the center vector is shown in formula (3).
Simultaneously extract D 1 The relationship in the knowledge graph of the data center is also a triplet (T 1 j ,R j ,T 2 j ) By vectorsRepresenting the calculation of the center vector of these triples +.>The calculation method of the kth element in the center vector is the same as the formula (3). Calculate these center vectors +.>The hamming distance between them is counted to obtain the farthest hamming distance, if +.>And->The Hamming distance between the two is larger than the farthest Hamming distance, then D is considered as 2 The triplet of this in the knowledge graph of the data center is a new triplet, if +.>And->The Hamming distance between the two is smaller than the farthest Hamming distance, then D is considered 2 The triplet of this in the data center knowledge-graph already exists at D 1 And the knowledge graph of the data center is not fused. And so on to sum the rest data center knowledge graph and D 1 And obtaining the cross-domain data center knowledge graph after the data center knowledge graph is fused.
And finally, selecting and extracting network knowledge from the knowledge graph module based on the reinforcement learning module to make a network slicing decision, and feeding decision information back to the EMS to finish network slicing.
The EMS feeds back the updated network state and resource data to a knowledge graph module, the knowledge graph module constructs an incremental knowledge graph, repeatedly detects the incremental knowledge graph and the original knowledge graph, and fuses the incremental knowledge graph and the original knowledge graph to obtain an updated new knowledge graph for the next slicing. The duplicate detection method is as follows.
The current moment is marked as i, and the current knowledge graph is G i The incremental knowledge graph generated after the slicing task is completed is recorded as delta t . Will delta t Embedded in G i Is defined in the vector space of (a). For a triplet vector (h, r, t) ∈Δ t Projecting vectors h and t onto a hyperplane having a unit normal vector w r The projection is shown in formula (4).
Vector conversion onto hyperplane denoted d r The loss function is:
if the loss function is closer to 0, Δ t New triples and G i The higher the overlap ratio. Detection of no match with G i Repeated triplets are added to G i Obtaining a new knowledge graph G i+1 。
As shown in fig. 1, by two data centers (D 1 ,D 2 ) And constructing a knowledge graph of the current network state and the resources by the composed physical network and the corresponding virtual network.
(1) And defining a related entity set E and a related relation set R according to the literal data such as the technical document. Such as: entity set e= { server, switch, firewall, data center interconnect device, CPU, memory, storage capacity, bandwidth } and relationship set r= { owned, located, connected }. And forming a model layer of the knowledge graph by using the entity set.
(2) Constructing instance set { key ] according to literal data i -(thing i1 ,thing i2 ,...),key i E, calculating the average distance between other nouns in the text data and key words, setting the occupied distance of one text as 1, setting the average distance as the average value of the sum of the distances between the word and the nearest key word, and finally counting that the noun with the smallest distance is the corresponding thong of the key words.
For example, the average distance between other nouns and "switch" from the following text is shown in table 1:
the server is connected with 580 switches through electric ports, the 580 switches are connected with a convergence switch 680 as access switches, one 680 switch controls a plurality of 580 switches, and a plurality of convergence switches 680 are controlled by a core switch 690 to form a small data center together.
Table 1 average distance between other nouns and "switch
Then the corresponding hig of the switch (key) is (580, 680, 690).
(3) According to speechThe relation between the picture data extraction examples is obtained<Example 1, relationship, example 2>The triples thus construct a knowledge graph. First, the collected voice and picture data are marked, and two examples (hiking) are carried out each time 1 :T 1 ,thing 2 :T 2 ) Identifying, as shown in FIG. 3, the red and green circles in the graph are labeled switches, i.e., (R_switch, G_switch) are two instances of a physical switch (T 1 ,T 2 ) The corresponding label is (y 1 ,y 2 ). As shown in fig. 2, identifying the marked picture data through different CNN networks, and the loss function of Task1 is:
wherein each vector has n dimensions, T 1 (t 11 ,t 12 ,...,t 1n ),T 2 (t 21 ,t 22 ,...,t 2n ),y 1 (y 11 ,y 12 ,...,y 1n ),y 2 (y 21 ,y 22 ,...,y 2n ). Similarly, a section of marked voice passes through DNN1 network to identify that the voice contains a certain relation R 1 (r 11 ,r 12 ,...,r 1n ) Only one relationship is identified for each task. Then T is taken up 1 、T 2 、R 1 After passing through the same DNN2 network, the output is T 1 '、T 2 '、R 1 '. Final judgment (T) 1 ',T 2 ',R 1 ') whether a triplet is formed, the Task2 loss function is:
in minimizing the loss function (error 0 *error 1 ) Under the condition of (1), parameters in CNN and DNN models are adjusted to obtain a knowledge-graph triplet (T) 1 ',T 2 ',R 1 '). For example from this piece of speech the relationship is identified as connection,then, from fig. 2, through CNN1 and CNN2, r_switch and g_switch are identified, and after DNN2, connection between r_switch and g_switch is determined, so as to obtain a triplet (r_switch, connect, g_switch). Thereby constructing the data center D 1 Network knowledge graph.
(2) Constructing a knowledge graph of a cross-domain data center
Construction of D according to step (1) 2 Knowledge graph of data center. Selecting D 2 Triplet with relation of R in knowledge graph of data centerBy vectors->Representing the calculation of the center vector of these triples +.>The calculation formula of the kth element in the center vector is shown in formula (3).
Simultaneously extract D 1 The relationship in the knowledge graph of the data center is also a triplet (T 1 j ,R j ,T 2 j ) By vectorsRepresenting the calculation of the center vector of these triples +.>The calculation method of the kth element in the center vector is the same as the formula (3). Calculate these center vectors +. >The hamming distance between them is counted to obtain the farthest hamming distance, if +.>And->The Hamming distance between the two is larger than the farthest Hamming distance, then D is considered as 2 The triplet of this in the knowledge graph of the data center is a new triplet, if +.>And->The Hamming distance between the two is smaller than the farthest Hamming distance, then D is considered 2 The triplet of this in the data center knowledge-graph already exists at D 1 And the knowledge graph of the data center is not fused. D (D) 2 Knowledge graph of data center and D 1 And obtaining the cross-domain data center knowledge graph after the data center knowledge graph is fused.
And finally, selecting and extracting network knowledge from the knowledge graph module based on the reinforcement learning module to make a network slicing decision, and feeding decision information back to the EMS to finish network slicing.
The EMS feeds back the updated network state and resource data to a knowledge graph module, the knowledge graph module constructs an incremental knowledge graph, repeatedly detects the incremental knowledge graph and the original knowledge graph, and fuses the incremental knowledge graph and the original knowledge graph to obtain an updated new knowledge graph for the next slicing. The duplicate detection method is as follows.
The current moment is marked as i, and the current knowledge graph is G i The incremental knowledge graph generated after the slicing task is completed is recorded as delta t . Will delta t Embedded in G i Is defined in the vector space of (a). For a triplet vector (h, r, t) ∈Δ t Projecting vectors h and t onto a hyperplane having a unit normal vector w r The projection is shown in formula (4).
Vector conversion onto hyperplane denoted d r The loss function is:
if the loss function is closer to 0, Δ t New triples and G i The higher the overlap ratio. Detection of no match with G i Repeated triplets are added to G i Obtaining a new knowledge graph G i+1 。
According to the above example, according to the method for digital twin network slicing based on the knowledge graph provided in the first embodiment, the network slicing under the dynamic network environment can be completed quickly and effectively, the deployment cost is reduced to the maximum extent, and the performances of time delay, energy consumption and the like are optimized.
Embodiment III:
based on the methods provided in the implementation one and the embodiment two, a specific example of performing network slicing in a practical scenario is provided in this embodiment. In practical implementation, reference may be made to this example, which is implemented in combination with the actual requirements.
The scenario of this embodiment is based on the following text: the server is connected with a 5800 exchanger through an electric port, the 5800 exchanger is used as an access exchanger to be connected with a convergence exchanger 6800, one 6800 exchanger controls a plurality of 5800 exchangers, and a plurality of convergence exchangers 6800 are controlled by a core exchanger 6900 to form a small data center. "
Firstly, using high-frequency vocabulary such as a switch as a key, and secondly, calculating the average distance between the nouns of a server, an electric port, 5800, an access and the like and the switch in the text. For example, the distance between the server and the switch is 9, the distance between the electric port and the switch is 5, and the distance between 5800 and the switch is 0. Finally, the nouns with the smallest average distance among the nouns are countedFor the corresponding taking of the switch (key), the object set corresponding to the switch is obtained, namely (switch- (5800, 6800, 6900)), then a triplet (switch-category-5800), (switch-category-6800) and (switch-category-6900) can be constructed. By this method, a single data center D is obtained 1 After the object set, forming a triplet of the initial state according to the object set.
And judging the rationality and the existence of the ternary combination in the initial state. The judgment is made by a two-layer neural network composed of an anchor network and a judgment network as shown in fig. 6. The first layer is an anchor network comprising (CNN 1, CNN2, DNN 1), and the second layer is a decision network comprising (DNN 2). For example, taking a "category" Relationship as an example, the previously established triplets such as (switch-category-access) are subjected to supervised learning on the parameters of the first layer neural network through the first layer anchoring network, the parameters of the neural network are adjusted, and the parameters of the neural network are satisfied with t1_vector (switch) +relation_vector) =t2_vector, so that the first layer anchoring neural network parameter adjustment work is completed at this time. And then anchoring the triplets in the initial state, such as (exchanger-type-5800) through the trained first layer of anchored neural network, to obtain the vectorized representation of the triplets in the initial state, namely, the triplets in the initial state are evolved into the triplets in the checking state. Such as the triplet vectorized representation of figure 6,
R 1 =(-0.45594668,0.45142751,0.43007157,...,-0.19828),
R 1 =(0.13850508,0.35625942,0.38727191,...,-0.08883082),
R 1 =(0.35304537,-0.37963883,0.20548223,...,-0.28563719),
r 1 =(0.20457409,0.44506297,0.45373613,...,0.33640039),
The second layer of judging network judges the rationality and the existence of the ternary combination in the checking state, and only judges the ternary combination under one relation in each judging taskRationality and presence of the group. For example, the triplet under the relationship "category" is determined, as shown in the determination case R of (1) in FIG. 6 1 =r 1 Type =type, with R 1 As a result of the fact that the shaft,forming a circle for 360 degrees of edge rotation, wherein the circle range is the judging range of the relation 'category', and then r is used for 1 Is a shaft (or a rope)>A circle is formed for a 360 degree rotation of the edge, which is seen to be well within the decision range, so the triplet (switch-class-5800) can be established but is an already existing triplet. Similarly, the determination case R 2 =r 2 By r 2 Is a shaft (or a rope)>A circle formed for 360 degrees of edge rotation intersects with the judgment range but does not completely coincide, then +.>Satisfy R 2 The condition of the triplet composition under the relation is a new tripletOutputting as a triplet of the established state through a judging network; determination of the situation R 2 =r 3 By r 3 Is a shaft (or a rope)>The circle formed for 360 degrees of edge rotation has no intersection with the decision range, then +. >Not meeting R 2 The triples under the relationship form the condition,the triplet is unreasonable.
The judging network judges according to the rule and outputs the new triplet meeting the composition condition, the output triplet is the triplet in the established state, and the triplet in (1) form a single data center D together 1 Is a knowledge graph of (1). Then constructing single data center D by the same method 2 Is a knowledge graph of (1).
Constructing a cross-domain data center network knowledge graph. After each single data center network knowledge graph is gradually constructed, the single data center network knowledge graphs are sequentially fused to construct the cross-domain data center network knowledge graph. Data center D 1 The knowledge patterns of (a) are to-be-fused knowledge patterns, and the data center D 2 The knowledge graph of (2) is the fused knowledge graph. Firstly judging whether a new relation is generated in the fused knowledge graph, if so, all triples in the new relation are new triples, and directly fusing the new triples into the knowledge graph to be fused. When the relation R in the fused knowledge-graph is existing in the knowledge-graph to be fused, selecting a triplet with the relation R in the fused knowledge-graph, and using vectorsRepresenting the calculation of the center vector of these triples +.>The calculation formula of the kth element in the center vector is shown in formula (3).
Simultaneously extracting the triples with the relation of R in the knowledge graph to be fused, and using vectorsRepresenting the calculation of the center vector of these triples +.>The kth element in the center vectorThe calculation method of (2) is the same as that of the formula (3). Calculate these center vectors +.>And counting the minimum and maximum hamming distances among the hamming distances to judge whether the triples in the fused knowledge graph are the judging range of the new triples. If->And->If the Hamming distance between the two is not in the judging range, the triplet of the fusion knowledge graph is regarded as a new triplet, if +.>And->If the Hamming distance between the two knowledge patterns is within the judging range, the triplet of the fused knowledge patterns is considered to exist in the knowledge patterns to be fused, and fusion is not carried out. D (D) 2 Knowledge graph of data center and D 1 The data center knowledge graph is fused to obtain a cross-domain data center knowledge graph, for example, as shown in fig. 8.
(5) And the knowledge graph module generates a triplet according to the received slice demand data. For example, VR slice: 2 VNFs are needed, VNF1 needs computing resource 10, storage resource 10, transfer resource 100, VNF2 needs computing resource 10, storage resource 20, transfer resource 100, slice priority is 1. The generation of triples is shown in fig. 9: (VNF 1, calculate, 10), (VNF 1, store, 10), (VNF 1, transfer, 100), (VNF 2, calculate, 10), (VNF 2, store, 20), (VNF 1, transfer, 100), (VNF 1, connect, VNF 2), (VR slice, priority, 1), (VR slice, need, VNF 2). The triples are fused into the knowledge spectrum of the cross-domain data center of the last step to form the knowledge spectrum which is finally transmitted to the reinforcement learning module, for example, as shown in fig. 10.
Embodiment four:
a knowledge-graph-based digital twin network slicing apparatus, the apparatus comprising: at least one processor; and a memory communicatively coupled to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being programmed to perform the method of knowledge-graph based digital twin network slicing provided in embodiment one or embodiment two.
In the second embodiment, the device for slicing the digital twin network based on the knowledge graph includes one or more processors 21 and a memory 22. In fig. 11, a processor 21 is taken as an example.
The processor 21 and the memory 22 may be connected by a bus or otherwise, for example in fig. 8.
The memory 22 serves as a non-volatile computer readable storage medium for storing non-volatile software programs and non-volatile computer executable programs, such as the knowledge-graph-based digital twin network slice method in embodiment one. The processor 21 executes means for implementing knowledge-graph based digital twin network slicing by running non-volatile software programs and instructions stored in the memory 22.
The memory 22 may include high-speed random access memory, and may also include non-volatile memory, such as at least one magnetic disk storage device, flash memory device, or other non-volatile solid-state storage device. In some embodiments, memory 22 may optionally include memory located remotely from processor 21, which may be connected to processor 21 via a network. Examples of such networks include, but are not limited to, the internet, intranets, local area networks, mobile communication networks, and combinations thereof.
The program instructions/modules are stored in the memory 22 and when executed by the one or more processors 21 perform the method of knowledge-graph-based digital twin network slicing of the above embodiment, for example, performing the steps shown in fig. 1 described above.
It should be noted that, because the content of information interaction and execution process between modules and units in the above-mentioned device and system is based on the same concept as the processing method embodiment of the present invention, specific content may be referred to the description in the method embodiment of the present invention, and will not be repeated here.
Those of ordinary skill in the art will appreciate that all or a portion of the steps in the various methods of the embodiments may be implemented by a program that instructs associated hardware, the program may be stored on a computer readable storage medium, the storage medium may include: read Only Memory (ROM), random access Memory (RAM, random Access Memory), magnetic or optical disk, and the like.
It will be readily appreciated by those skilled in the art that the foregoing description is merely a preferred embodiment of the invention and is not intended to limit the invention, but any modifications, equivalents, improvements or alternatives falling within the spirit and principles of the invention are intended to be included within the scope of the invention.
Claims (10)
1. A method for slicing a digital twin network based on a knowledge graph, the method comprising:
constructing a thing set of a data center, and constructing an initial triplet composed of things, relations and things according to the thing set;
checking the rationality and the existence of the initial triples, constructing a knowledge graph of each data center by using the initial triples meeting the checking conditions in each data center and the triples established according to the prior art, and fusing the knowledge graphs of all the data centers at one time to construct a knowledge graph of a cross-domain data center;
and representing the network operation state of the network slice by the knowledge graph of the cross-domain data center, predicting the deployment positions of the network slice VNF and the network slice PNF according to the network operation state of the network slice, transmitting the deployment decision back to the knowledge graph, and adjusting the network operation state by the knowledge graph according to the deployment decision.
2. The knowledge-based digital twin network slicing method of claim 1, wherein the constructing the set of things for the data center comprises:
selecting keywords and conventional words from the text data, converting the voice data or the picture data into corresponding text data, and then selecting the keywords and the conventional words;
calculating the character distance between the conventional word and the keyword, and comparing to obtain the minimum character distance between the conventional word and the keyword;
selecting the key words as keys, selecting the conventional words corresponding to the key words in the minimum character distance as sting, and constructing a key-sting set;
the transaction set of the data center includes at least one key-sting set.
3. The knowledge-graph-based digital twin network slicing method of claim 1, wherein constructing an initial triplet of things, relationships, and things from the set of things comprises:
selecting a first object and a second object in the key-sting set, marking the first object and the second object on text, voice or picture data, and presetting a first relation between the first object and the second object;
Identifying, by a neural network, the first thing, the second thing, and the first relationship according to the annotation;
judging whether the preset first relation is established or not according to a loss value generated in the neural network identification process;
if the relationship is established, the first thing, the second thing and the first relationship construct an initial triplet.
4. The method of knowledge-graph-based digital twin network slicing of claim 3, wherein said identifying said first thing, said second thing, and said first relationship by said neural network based on said annotations comprises:
the neural network identifying annotations for the first thing and the second thing, generating a first loss value;
the neural network identifies the first relationship and generates a second loss value;
and obtaining a neural network identification loss value according to the first loss value and the second loss value.
5. The method for slicing a digital twin network based on a knowledge-graph of claim 3, wherein said determining whether said first predetermined relationship is established by a loss value generated during a neural network recognition process comprises:
Presetting a loss threshold value, wherein the loss threshold value is obtained by multiplying or adding the first loss value and the second loss value;
if the neural network identification loss value is smaller than the loss threshold value, the first relation is established;
and if the neural network identification loss value is greater than or equal to the loss threshold value, the first relation is not established.
6. The knowledge-graph-based digital twin network slicing method of claim 3, wherein said determining whether said predetermined first relationship is established from a loss value generated during neural network recognition, further comprises:
and if the first relation is not established, sequentially changing the first relation from a relation set preset in a data center, and judging whether the changed first relation is established or not by the neural network identification loss value.
7. The method for slicing a digital twin network based on a knowledge-graph of claim 1, wherein the constructing a knowledge-graph of each data center using an initial triplet in each data center satisfying a verification condition and a triplet established according to a well-known method comprises:
judging whether the distance between the relationship center vector in the triplet and the relationship center vector in the initial triplet exceeds a fusion distance threshold;
And if the distance between the relationship center vector in the triplet and the relationship center vector in the initial triplet exceeds the fusion distance threshold, fusing the triplet with the initial triplet.
8. The knowledge-graph-based digital twin network slice method of claim 1, wherein predicting deployment locations of the network slice VNF and a network slice PNF based on a network operational state of the network slice comprises:
the network running state of the network slice specifically comprises the position, the resource and the relation information of the network slice PNF and the network slice VNF;
the network operation states of the network slice PNF and the network slice VNF are represented by triples of the knowledge spectrogram, and the network operation states of the network slice PNF and the network slice VNF are converted into a triplet vector representation;
if the triplet vector of the network slice VNF is simultaneously smaller than the triplet vector of the network slice PNF, the network slice VNF is deployed above the network slice PNF;
and if the triplet vector of the network slice VNF is not smaller than the triplet vector of the network slice PNF, replacing the network slice PNF until the network slice VNF is deployed on the network slice PNF.
9. The method of knowledge-based digital twin network slicing according to claim 1, wherein the deployment decision is returned to the knowledge-graph, the knowledge-graph adjusting the network operational state according to the deployment decision, the method comprising:
and forming an action set according to the number of the network slice PNFs and the network slice VNs, setting a feedback function of a network running state for the network slice which is completely deployed, and adjusting the knowledge graph according to the action set and the feedback function.
10. A knowledge-graph-based digital twin network slicing apparatus, comprising: at least one processor; and a memory communicatively coupled to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being programmed to perform the knowledge-graph based digital twin network slice method of any one of claims 1-9.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202310089901.1A CN116055330A (en) | 2023-01-17 | 2023-01-17 | Digital twin network slicing method and device based on knowledge graph |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202310089901.1A CN116055330A (en) | 2023-01-17 | 2023-01-17 | Digital twin network slicing method and device based on knowledge graph |
Publications (1)
Publication Number | Publication Date |
---|---|
CN116055330A true CN116055330A (en) | 2023-05-02 |
Family
ID=86119898
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202310089901.1A Pending CN116055330A (en) | 2023-01-17 | 2023-01-17 | Digital twin network slicing method and device based on knowledge graph |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN116055330A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117668528A (en) * | 2024-02-01 | 2024-03-08 | 成都华泰数智科技有限公司 | Natural gas voltage regulator fault detection method and system based on Internet of things |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111339321A (en) * | 2020-05-18 | 2020-06-26 | 中国人民解放军国防科技大学 | Triple authenticity detection method and device in knowledge graph |
US20200313969A1 (en) * | 2018-03-29 | 2020-10-01 | Huawei Technologies Co., Ltd. | Network Slice Deployment Method and Apparatus |
CN112365366A (en) * | 2020-11-12 | 2021-02-12 | 广东电网有限责任公司 | Micro-grid management method and system based on intelligent 5G slice |
CN113347641A (en) * | 2020-03-02 | 2021-09-03 | 中国电信股份有限公司 | Network deployment method, device and computer readable storage medium |
US20220191648A1 (en) * | 2021-12-20 | 2022-06-16 | Ned M. Smith | Digital twin framework for next generation networks |
CN115051926A (en) * | 2022-06-22 | 2022-09-13 | 烽火通信科技股份有限公司 | Digital twin architecture, model evaluation system and model operation method |
-
2023
- 2023-01-17 CN CN202310089901.1A patent/CN116055330A/en active Pending
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20200313969A1 (en) * | 2018-03-29 | 2020-10-01 | Huawei Technologies Co., Ltd. | Network Slice Deployment Method and Apparatus |
CN113347641A (en) * | 2020-03-02 | 2021-09-03 | 中国电信股份有限公司 | Network deployment method, device and computer readable storage medium |
CN111339321A (en) * | 2020-05-18 | 2020-06-26 | 中国人民解放军国防科技大学 | Triple authenticity detection method and device in knowledge graph |
CN112365366A (en) * | 2020-11-12 | 2021-02-12 | 广东电网有限责任公司 | Micro-grid management method and system based on intelligent 5G slice |
US20220191648A1 (en) * | 2021-12-20 | 2022-06-16 | Ned M. Smith | Digital twin framework for next generation networks |
CN115051926A (en) * | 2022-06-22 | 2022-09-13 | 烽火通信科技股份有限公司 | Digital twin architecture, model evaluation system and model operation method |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117668528A (en) * | 2024-02-01 | 2024-03-08 | 成都华泰数智科技有限公司 | Natural gas voltage regulator fault detection method and system based on Internet of things |
CN117668528B (en) * | 2024-02-01 | 2024-04-12 | 成都华泰数智科技有限公司 | Natural gas voltage regulator fault detection method and system based on Internet of things |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111625361B (en) | Joint learning framework based on cooperation of cloud server and IoT (Internet of things) equipment | |
CN104995870B (en) | Multiple target server arrangement determines method and apparatus | |
CN113705959B (en) | Network resource allocation method and electronic equipment | |
CN110995487A (en) | Multi-service quality prediction method and device, computer equipment and readable storage medium | |
CN108365969B (en) | Self-adaptive service combination method based on wireless sensor network | |
CN108075974B (en) | Flow forwarding control method and device and SDN architecture system | |
CN108111335A (en) | A kind of method and system dispatched and link virtual network function | |
Candan et al. | A dynamic island model for adaptive operator selection | |
Cao et al. | Collaborative attributes and resources for single-stage virtual network mapping in network virtualization | |
CN116055330A (en) | Digital twin network slicing method and device based on knowledge graph | |
CN114978908A (en) | Computing power network node evaluation and operation method and device | |
CN112330519A (en) | Data processing method and device | |
Li et al. | Network topology optimization via deep reinforcement learning | |
Li et al. | Study of manufacturing cloud service matching algorithm based on OWL-S | |
Zhou et al. | Multi-task deep learning based dynamic service function chains routing in SDN/NFV-enabled networks | |
CN115220916A (en) | Automatic computing power scheduling method, device and system for video intelligent analysis platform | |
CN113133038B (en) | Power Internet of things link backup method, device, equipment and storage medium | |
Mezni et al. | A negotiation‐based service selection approach using swarm intelligence and kernel density estimation | |
CN113139639A (en) | MOMBI-based smart city application-oriented multi-target calculation migration method and device | |
CN115412401B (en) | Method and device for training virtual network embedding model and virtual network embedding | |
CN115016950B (en) | Data analysis method and system based on multithreading model | |
CN116109058A (en) | Substation inspection management method and device based on deep reinforcement learning | |
CN112906745B (en) | Integrity intelligent network training method based on edge cooperation | |
Li et al. | Online coordinated NFV resource allocation via novel machine learning techniques | |
Han et al. | Multi-SP Network Slicing Parallel Relieving Edge Network Conflict |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination |