TWI773039B

TWI773039B - System and method for lifecycle management of mobile network

Info

Publication number: TWI773039B
Application number: TW109145382A
Authority: TW
Inventors: 吳振奕; 周漢平; 吳世偉
Original assignee: 中華電信股份有限公司
Priority date: 2020-12-22
Filing date: 2020-12-22
Publication date: 2022-08-01
Also published as: TW202226029A

Abstract

A system and a method for lifecycle management of mobile network are provided. In the method, multiple monitoring data related to the mobile network services are obtained, and each monitoring data is related to the resource status of the mobile network service. The monitoring data is encoded, to generate multiple initial vectors. One feature vector are generated through unsupervised learning on the combination of the initial vectors. Multiple decision probability distributions are obtained through multiple inference models performed on the feature vectors, respectively. The inference models use different machine learning algorithms. Each decision probability distribution includes probabilities corresponding to multiple trouble fixing ways. The decision probability distributions are used as the input of the attention layer, to obtain the final probability distribution. The trouble fixing way would be determined based on the final probability distribution. Accordingly, it can be ensured that the network service is operated normally within the lifecycle.

Description

行動網路的生命週期管理系統及方法Life cycle management system and method of mobile network

本發明是有關於一種網路管理技術，且特別是有關於一種行動網路的生命週期管理系統及方法。The present invention relates to a network management technology, and in particular, to a life cycle management system and method of a mobile network.

行動網路服務於生命週期過程障礙發生時，歐洲電信標準化協會(European Telecommunications Standards Institute，ETSI)定義幾種確保網路服務正常運行的調用方法。然而，ETSI並未明確定義這幾種方法所適用的情境。若未即時挑選出最佳的解決方案以確保行動網路服務正常運行，則系統便會在挑選方案的過程中耗費不必要的時間成本。The European Telecommunications Standards Institute (ETSI) defines several invocation methods to ensure the normal operation of network services when obstacles occur during the life cycle process of mobile network services. However, ETSI does not clearly define the context in which these methods are applicable. If the best solution is not selected in real time to ensure the normal operation of the mobile network service, the system will spend unnecessary time and cost in the process of selecting the solution.

有鑑於此，本發明實施例提供一種行動網路的生命週期管理系統及方法，透過基於機器學習的模型決策，以快速且精確地選擇較合適的解決方案。In view of this, embodiments of the present invention provide a system and method for life cycle management of a mobile network, which can quickly and accurately select a more suitable solution through model decision based on machine learning.

本發明實施例的行動網路的生命週期管理方法包括(但不僅於)下列步驟：取得行動網路服務相關的多筆監控資料，且各監控資料相關於行動網路服務的資源狀態。對那些監控資料編碼以產生多個初步向量。將那些初步向量合併後，透過非監督式學習(Unsupervised Learning)產生一個特徵向量。將那特徵向量透過數個推論模型分別得出多個決策機率分布。那些推論模型採用不同機器學習演算法，且各決策機率分布包括對應於數個障礙修復方案的機率。將那些決策機率分布作為注意力層(attention layer)的輸入以得出最終機率分布。依據此最終機率分布決定障礙修復方案。The life cycle management method of the mobile network according to the embodiment of the present invention includes (but is not limited to) the following steps: obtaining multiple pieces of monitoring data related to the mobile network service, and each monitoring data is related to the resource status of the mobile network service. Those monitoring data are encoded to generate a plurality of preliminary vectors. After combining those preliminary vectors, a feature vector is generated through Unsupervised Learning. The eigenvectors are passed through several inference models to obtain multiple decision probability distributions. Those inference models employ different machine learning algorithms, and each decision probability distribution includes probabilities corresponding to several obstacle repair options. Those decision probability distributions are used as input to the attention layer to derive the final probability distribution. Based on this final probability distribution, the obstacle repair plan is determined.

本發明實施例的行動網路的生命週期管理系統包括(但不僅限於)儲存器及處理器。儲存器儲存數個軟體模組。處理器耦接儲存器，載入且執行那些軟體模組。那些軟體模組包括資料蒐集模組、特徵編碼模組、特徵向量產生模組及決策模組。資料蒐集模組用以取得行動網路服務相關的多筆監控資料，且各監控資料相關於行動網路服務的資源狀態。特徵編碼模組對那些監控資料編碼以產生數個初步向量。特徵向量產生模組將那些初步向量合併後，透過非監督式學習產生一個特徵向量。決策模組將那特徵向量透過數個推論模型分別得出數個決策機率分布，將那些決策機率分布作為注意力層的輸入以得出最終機率分布，並依據此最終機率分布決定障礙修復方案。那些推論模型採用不同的機器學習演算法，且各決策機率分布包括對應於數個障礙修復方案的機率。The life cycle management system of the mobile network according to the embodiment of the present invention includes (but is not limited to) a storage and a processor. The memory stores several software modules. The processor, coupled to the memory, loads and executes those software modules. Those software modules include data collection modules, feature encoding modules, feature vector generation modules, and decision-making modules. The data collection module is used to obtain multiple pieces of monitoring data related to the mobile network service, and each monitoring data is related to the resource status of the mobile network service. The feature encoding module encodes those monitoring data to generate several preliminary vectors. The feature vector generation module combines those preliminary vectors to generate a feature vector through unsupervised learning. The decision module uses the feature vector to obtain several decision probability distributions through several inference models, and uses those decision probability distributions as the input of the attention layer to obtain the final probability distribution, and determines the obstacle repair plan according to the final probability distribution. Those inference models employ different machine learning algorithms, and each decision probability distribution includes probabilities corresponding to several obstacle repair options.

基於上述，本發明實施例的行動網路的生命週期管理系統及方法，對將行動網路服務特徵進行非監督式學習模型編碼，以取得可表示行動網路服務彼此關係的特徵向量。此外，透過機器學習模型決策，使系統能更精確地選擇適當的障礙排除方案，以確保行動網路服務正常運作，進而降低時間成本，且提升決策效能。Based on the above, the mobile network life cycle management system and method according to the embodiments of the present invention encode the characteristics of mobile network services in an unsupervised learning model to obtain feature vectors that can represent the relationship between mobile network services. In addition, through machine learning model decision-making, the system can more accurately select an appropriate obstacle removal solution to ensure the normal operation of mobile network services, thereby reducing time costs and improving decision-making efficiency.

為讓本發明的上述特徵和優點能更明顯易懂，下文特舉實施例，並配合所附圖式作詳細說明如下。In order to make the above-mentioned features and advantages of the present invention more obvious and easy to understand, the following embodiments are given and described in detail with the accompanying drawings as follows.

圖1是依據本發明一實施例的行動網路的生命週期管理系統100的元件方塊圖。請參照圖1，生命週期管理系統100包括但不僅限於儲存器110及處理器130。生命週期管理系統100可以是一台或更多台伺服器或其他電腦系統。FIG. 1 is a block diagram of components of a life cycle management system 100 of a mobile network according to an embodiment of the present invention. Referring to FIG. 1 , the life cycle management system 100 includes but is not limited to a storage 110 and a processor 130 . Life cycle management system 100 may be one or more servers or other computer systems.

儲存器110可以是任何型態的固定或可移動隨機存取記憶體(Radom Access Memory，RAM)、唯讀記憶體(Read Only Memory，ROM)、快閃記憶體(flash memory)、傳統硬碟(Hard Disk Drive，HDD)、固態硬碟(Solid-State Drive，SSD)或類似元件。在一實施例中，記憶體110用以記錄程式碼、軟體模組(例如，資料蒐集模組111、特徵編碼模組112、特徵向量產生模組113、決策模組114、第一推論模組115、第二推論模組116、決策權重分配模組117、及排序模組118)、組態配置、資料或檔案(例如，監控資料、向量、障礙修復方案等)。而軟體模組的運作待後續實施例詳述。The storage 110 may be any type of fixed or removable random access memory (RAM), read only memory (ROM), flash memory, conventional hard disks (Hard Disk Drive, HDD), Solid-State Drive (Solid-State Drive, SSD) or similar components. In one embodiment, the memory 110 is used to record code, software modules (eg, the data collection module 111, the feature encoding module 112, the feature vector generation module 113, the decision module 114, the first inference module 115. Second inference module 116, decision weight assignment module 117, and ranking module 118), configuration configuration, data or files (eg, monitoring data, vectors, obstacle repair solutions, etc.). The operation of the software module will be described in detail in subsequent embodiments.

處理器130耦接記憶體110，處理器130並可以是中央處理單元(Central Processing Unit，CPU)、圖形處理單元(Graphic Processing unit，GPU)，或是其他可程式化之一般用途或特殊用途的微處理器(Microprocessor)、數位信號處理器(Digital Signal Processor，DSP)、可程式化控制器、現場可程式化邏輯閘陣列(Field Programmable Gate Array，FPGA)、特殊應用積體電路(Application-Specific Integrated Circuit，ASIC)、神經網路加速器或其他類似元件或上述元件的組合。在一實施例中，處理器130用以執行生命週期管理系統100的所有或部份作業，且可載入並執行記憶體110所記錄的各程式碼、軟體模組、檔案及資料。The processor 130 is coupled to the memory 110, and the processor 130 may be a central processing unit (CPU), a graphics processing unit (GPU), or other programmable general-purpose or special-purpose Microprocessor (Microprocessor), Digital Signal Processor (DSP), Programmable Controller, Field Programmable Gate Array (FPGA), Application-Specific Integrated Circuit Integrated Circuit, ASIC), neural network accelerator or other similar elements or a combination of the above elements. In one embodiment, the processor 130 is used to execute all or part of the operations of the life cycle management system 100 , and can load and execute each code, software module, file and data recorded in the memory 110 .

下文中，將搭配生命週期管理系統100中的各項裝置、元件及模組說明本發明實施例所述之方法。本方法的各個流程可依照實施情形而隨之調整，且並不僅限於此。Hereinafter, the method according to the embodiment of the present invention will be described in conjunction with various devices, components and modules in the life cycle management system 100 . Each process of the method can be adjusted according to the implementation situation, and is not limited to this.

圖2是依據本發明一實施例的行動網路的生命週期管理方法的流程圖。請參照圖2，資料蒐集模組111取得監控資料(步驟S210)。具體而言，資料蒐集模組111可經由網路、儲存載具或儲存器110取得所有與行動網路服務(Network Service，NS)實例(Instance)關的多筆監控資料。在一實施例中，各監控資料相關於行動網路服務的資源狀態。以ETSI定義的參數為例，nsState(行動網路服務的目前狀態)、vnfInstanceIds(與行動網路服務相關網路功能虛擬化(Network Functions Virtualization，VNFs))、pnfInfoIds(與行動網路服務相關的實體網路功能(Physical Network Functions，PNFs))、virtualLinkInfoIds (與行動網路服務相關的虛擬連結(Virtual Links，VLs))、vnfgInfoIds(與行動網路服務相關的網路功能虛擬化轉送圖(VNF Forwarding Graphs，VNFGs))、sapInfoIds(與行動網路服務相關的服務存取點(Service Access Points，SAPs))、topology (VNFs與VLs所組成的拓樸訊息)、severity (網路告警的嚴重程度)、eventType (網路告警的類型)、nsOperation(與行動網路服務相關的操作)、label(標籤)或其他特徵(feature)。表(1)是一範例說明監控資料：表(1) id nsOperation vnfInstanceIds pnfInfoIds virtualLinkInfoIds features severity eventType label 0 1 [1,2,15] [0,6] [5,2] … 0 0 0 1 1 [5,8,11] [2,7,8] [8,4] … 3 3 1 2 0 [1,3] [1,2,3,5,7,10] [7] … 1 1 1 3 2 [2,5] [1] [3] … 1 0 1 … … … … … … … … … 100 0 [0,5,11] [5,7,9] [8,13] … 2 2 101 1 [3,4,22] [21,23] [4,7] … 0 3 0 102 2 [1,7,10] [6] [13,7] … 2 0 2 FIG. 2 is a flowchart of a life cycle management method of a mobile network according to an embodiment of the present invention. Referring to FIG. 2 , the data collection module 111 obtains monitoring data (step S210 ). Specifically, the data collection module 111 can obtain all the multiple monitoring data related to the mobile network service (Network Service, NS) instance (Instance) through the network, the storage vehicle or the storage 110 . In one embodiment, each monitoring data is related to the resource status of the mobile network service. Taking the parameters defined by ETSI as an example, nsState (the current state of the mobile network service), vnfInstanceIds (network function virtualization (VNFs) related to the mobile network service), pnfInfoIds (the mobile network service related Physical Network Functions (PNFs)), virtualLinkInfoIds (Virtual Links (VLs) related to mobile network services), vnfgInfoIds (Virtualization transfer map (VNF) of network functions related to mobile network services Forwarding Graphs, VNFGs)), sapInfoIds (Service Access Points (SAPs) related to mobile network services), topology (topology information composed of VNFs and VLs), severity (severity of network alarms) ), eventType (type of network alert), nsOperation (operation related to mobile network service), label (label), or other features. Table (1) is an example to illustrate the monitoring data: Table (1) id nsOperation vnfInstanceIds pnfInfoIds virtualLinkInfoIds features severity eventType label 0 1 [1,2,15] [0,6] [5,2] … 0 0 0 1 1 [5,8,11] [2,7,8] [8,4] … 3 3 1 2 0 [1,3] [1,2,3,5,7,10] [7] … 1 1 1 3 2 [2,5] [1] [3] … 1 0 1 … … … … … … … … … 100 0 [0,5,11] [5,7,9] [8,13] … 2 2 101 1 [3,4,22] [21,23] [4,7] … 0 3 0 102 2 [1,7,10] [6] [13,7] … 2 0 2

須說明的是，依據不同設計需求，監控資料的類型及數量可再變更，且本發明實施例不加以限制。It should be noted that, according to different design requirements, the type and quantity of monitoring data can be changed, and the embodiment of the present invention is not limited.

特徵編碼模組112對那些監控資料編碼以產生多個初步向量(即，資料轉換)(步驟S220)。在一實施例中，特徵編碼模組112將那些監控資料中的非數值資料透單熱(One-Hot)編碼轉換成一個或更多個初步向量。One-Hot編碼是分類變數作為二進位向量的表示。首先，特徵編碼模組112將分類值映射到整數值。然後，特徵編碼模組112將各整數值被表示為二進位向量。The feature encoding module 112 encodes those monitoring data to generate a plurality of preliminary vectors (ie, data transformations) (step S220). In one embodiment, the feature encoding module 112 converts the non-numeric data in those monitored data into one or more preliminary vectors by one-hot encoding. One-Hot encoding is the representation of categorical variables as binary vectors. First, the feature encoding module 112 maps categorical values to integer values. Then, the feature encoding module 112 represents each integer value as a binary vector.

為了配合非監督學習(Unsupervised Learning)模型需求，部分監控資料(例如，VNFs、VLs、PNFs及SAP)的編碼方式與其他特徵有所差異。首先，特徵編碼模組112使用四維單熱編碼向量區別各監控資料/資源。接著，特徵編碼模組112計算各自種類的數量(例如，SUM _VNF、SUM _VL、SUM _PNF及SUM _SAP)，並使用一個有SUM _ELEMENT+ 1個元素且初始值為0的向量表示之，其中ELEMENT(元素)

{ VNF, VL, PNF, SAP }。特徵編碼模組112將有此種ELEMENT對應的索引(index)填1。若其所引不存在，則最後一個索引填1。最後，特徵編碼模組112將區別資源資訊的四維向量與最大(MAX(SUM _VNF+ 1、SUM _VL+ 1、SUM _PNF+ 1及SUM _SAP+ 1))維向量合併(不滿維度的向量補0至相同維度)，以得到表達這些監控資料(即，VNFs、VLs、PNFs和SAP)的初步向量。即，這些初步向量的維度相同。 In order to meet the requirements of the Unsupervised Learning model, some monitoring data (eg, VNFs, VLs, PNFs, and SAP) are encoded differently from other features. First, the feature encoding module 112 uses the four-dimensional one-hot encoding vector to distinguish each monitoring data/resource. Next, the feature encoding module 112 calculates the number of each category (eg, SUM _VNF , SUM _VL , SUM _PNF , and SUM _SAP ), and uses a vector with SUM _ELEMENT + 1 elements and an initial value of 0 to represent it, where ELEMENT (element)

{ VNF, VL, PNF, SAP }. The feature encoding module 112 fills in 1 with the index corresponding to this ELEMENT. If its reference does not exist, the last index is filled with 1. Finally, the feature encoding module 112 combines the four-dimensional vector of the distinguishing resource information with the maximum (MAX(SUM _VNF + 1, SUM _VL + 1, SUM _PNF + 1, and SUM _SAP + 1))-dimensional vector (the vectors with insufficient dimensions are filled with 0s) to the same dimensions) to obtain preliminary vectors expressing these monitoring data (ie, VNFs, VLs, PNFs, and SAPs). That is, the dimensions of these preliminary vectors are the same.

舉例而言，假設數量SUM _VNF、SUM _VL、SUM _PNF及SUM _SAP分別為5、7、3及5。[1, 0, 0, 0]為VNF，[0, 1, 0, 0]為VL，[0, 0, 1, 0]為PNF，且[0, 0, 0, 1] 為SAP。MAX(SUM _VNF+ 1 SUM _VL+ 1, SUM _PNF+ 1, SUM _SAP+ 1) = 8。即，初步向量的維度為8。若VNF的監控資料為[0,1,3]，則其初步向量是[1, 0, 0, 0, 1, 1, 0, 1, 0, 0, 0, 0]。若VL的監控資料為[4,9]，則其初步向量是[0, 1, 0, 0, 0, 0, 0, 0, 1, 0, 0, 1]。若PNF的監控資料為[0,1,4]，則其初步向量是[0, 0, 1, 0, 1, 1, 0, 1, 0, 0, 0, 0]。若SAP的監控資料為[1]，則其初步向量是[0, 0, 0, 1, 0, 1, 0, 0, 0, 0, 0, 0]。 For example, assume that the numbers SUM _VNF , SUM _VL , SUM _PNF and SUM _SAP are 5, 7, 3 and 5, respectively. [1, 0, 0, 0] is VNF, [0, 1, 0, 0] is VL, [0, 0, 1, 0] is PNF, and [0, 0, 0, 1] is SAP. MAX(SUM _VNF + 1 SUM _VL + 1, SUM _PNF + 1, SUM _SAP + 1) = 8. That is, the dimension of the preliminary vector is 8. If the monitoring data of the VNF is [0,1,3], its initial vector is [1, 0, 0, 0, 1, 1, 0, 1, 0, 0, 0, 0]. If the monitoring data of VL is [4,9], its initial vector is [0, 1, 0, 0, 0, 0, 0, 0, 1, 0, 0, 1]. If the monitoring data of PNF is [0,1,4], its initial vector is [0, 0, 1, 0, 1, 1, 0, 1, 0, 0, 0, 0]. If the monitoring data of SAP is [1], its initial vector is [0, 0, 0, 1, 0, 1, 0, 0, 0, 0, 0, 0].

須說明的是，在其他實施例中，特徵編碼模組112也可能採用其他文字轉向量的編碼。It should be noted that, in other embodiments, the feature encoding module 112 may also use the encoding of other character turning quantities.

特徵向量產生模組113將那些初步向量合併後，透過非監督式學習產生附關聯性一個特徵向量(步驟S230)。具體而言，圖3是依據本發明一實施例的特徵向量產生的示意圖。請參照圖3，此非監督學習模型301(例如，分群法、神經網路或主成份分析等)是用行動網路服務的監控資料預測其所包含的資源(例如，VNFs、VLs、PNFs及SAP)。其中，非監督學習模型301可不給定事先標記過的訓練樣本，且自動對輸入的資料進行分類或分群。藉此，資源相似的行動網路服務的特徵向量在高維空間是相近的。After the feature vector generating module 113 combines those preliminary vectors, a feature vector with associated correlation is generated through unsupervised learning (step S230 ). Specifically, FIG. 3 is a schematic diagram of generating a feature vector according to an embodiment of the present invention. Referring to FIG. 3, the unsupervised learning model 301 (eg, clustering method, neural network or principal component analysis, etc.) is used to predict the resources (eg, VNFs, VLs, PNFs and SAP). The unsupervised learning model 301 can automatically classify or group the input data without pre-labeling training samples. In this way, the feature vectors of mobile network services with similar resources are similar in high-dimensional space.

舉例而言，假設表(2)是一範例說明監控資料。表(2) id nsOperation vnfInstanceIds pnfInfoIds virtualLinkInfoIds features severity eventType label 0 1 [1,2,15] [0,6] [5,2] … 0 0 0 1 1 [5,8,11] [2,7,8] [8,4] … 3 3 1 2 0 [1,3] [1,2,3,5,7,10] [7] … 1 1 1 這三筆監控資料的特徵向量分別是： V(NS ₀) = [ 0.547, 0.0454, …, 0.56 , 0.1027 ] V(NS ₁) = [ 0.332, 0.258, …, 0.337 , 0.0003 ] V(NS ₂) = [ 0.302, 0.376, …, 0.117, 0.179 ] 而其餘弦相似性(Cosine similarity)分別是： Cosine Similarity ( V(NS ₀), V(NS ₁)) = 0.223 Cosine Similarity ( V(NS ₀), V(NS ₂)) = 0.254 Cosine Similarity ( V(NS ₁), V(NS ₂)) = 0.310 For example, assume that Table (2) is an example of monitoring data. Table 2) id nsOperation vnfInstanceIds pnfInfoIds virtualLinkInfoIds features severity eventType label 0 1 [1,2,15] [0,6] [5,2] … 0 0 0 1 1 [5,8,11] [2,7,8] [8,4] … 3 3 1 2 0 [1,3] [1,2,3,5,7,10] [7] … 1 1 1 The feature vectors of the three monitoring data are: V(NS ₀ ) = [ 0.547, 0.0454, …, 0.56 , 0.1027 ] V(NS ₁ ) = [ 0.332, 0.258, …, 0.337 , 0.0003 ] V(NS ₂ ) = [ 0.302, 0.376, …, 0.117, 0.179 ] and Cosine similarity are: Cosine Similarity ( V(NS ₀ ), V(NS ₁ )) = 0.223 Cosine Similarity ( V(NS ₀ ), V(NS ₂ )) = 0.254 Cosine Similarity ( V(NS ₁ ), V(NS ₂ )) = 0.310

決策模組114可進行確保生命週期之決策(步驟S240)。圖4是依據本發明一實施例的方案決策的流程圖。請參照圖4，決策模組114透過應用程式介面(Application Programming Interface，API)介接以取得發生障礙之行動網路服務樣本數據(步驟S241)。決策模組114可將那些特徵向量透過數個推論模型分別得出數個決策機率分布(步驟S243)。那些推論模型採用不同的機器學習演算法。例如，推論模型分別是基於隨機森林(Random Forest)、XGBoost(Extreme Gradient Boosting)及深度神經網路(Neural Network)(例如，卷積神經網絡(Convolutional Neural Network，CNN)、遞迴神經網路(Recurrent Neural Network，RNN）等)。假設第一推論模組115使用隨機森林及XGBoost。而第二推論模組116使用深度神經網路。推論模型可分析訓練樣本以自中獲得規律，從而透過規律對未知資料預測。各推論模型所輸出的決策機率分布包括對應於數個障礙修復方案的機率。這些障礙修復方案可以是ETSI或其他協會或規範所提供針對網路障礙的排除/修復的方案。而決策機率分布即是相關於使用這些障礙修復方案可確實解決障礙的機率。The decision module 114 may make a decision to ensure the life cycle (step S240). FIG. 4 is a flow chart of solution decision according to an embodiment of the present invention. Referring to FIG. 4 , the decision module 114 is interfaced through an application programming interface (API) to obtain sample data of the mobile network service in which the obstacle occurs (step S241 ). The decision module 114 may obtain a plurality of decision probability distributions through a plurality of inference models from those feature vectors (step S243 ). Those inference models employ different machine learning algorithms. For example, the inference models are based on Random Forest (Random Forest), XGBoost (Extreme Gradient Boosting) and deep neural network (Neural Network) (for example, Convolutional Neural Network (CNN), recurrent neural network ( Recurrent Neural Network, RNN), etc.). Assume that the first inference module 115 uses random forests and XGBoost. And the second inference module 116 uses a deep neural network. The inference model can analyze the training samples to obtain the rules, so as to predict the unknown data through the rules. The decision probability distribution output by each inference model includes probabilities corresponding to several obstacle repair options. These obstacle repair solutions may be those provided by ETSI or other associations or specifications for the removal/fixing of network obstacles. And the decision probability distribution is related to the probability that the obstacle will actually be solved using these obstacle repair solutions.

須說明的是，在一些實施例中，推論模型也可能是基於多層感知器 (Multi-Layer Perceptron，MLP)、支持向量機(Support Vector Machine，SVM)或其他演算法。It should be noted that, in some embodiments, the inference model may also be based on a Multi-Layer Perceptron (MLP), a Support Vector Machine (SVM) or other algorithms.

決策模組114將那些決策機率分布作為注意力層(attention layer)的輸入以得出最終機率分布。具體而言，神經網路是注意力層的重要元件。決策權重分配模組117可透過神經網路決定第一推論模組115及第二推論模組116所輸出之決策機率分布的對應權重(步驟S245)。決策模組114可對那些決策機率分布分別乘上對應的權重，以得出最終機率分布。The decision module 114 takes those decision probability distributions as input to the attention layer to derive the final probability distributions. Specifically, the neural network is an important element of the attention layer. The decision weight allocation module 117 may determine the corresponding weights of the decision probability distributions output by the first inference module 115 and the second inference module 116 through the neural network (step S245 ). The decision module 114 can multiply those decision probability distributions by corresponding weights to obtain the final probability distribution.

舉例而言，圖5是依據本發明一實施例的方案決策的示意圖。請參照圖5，隨機森林模型501輸出的決策機率分布為[0.4, 0.5, 0.1]，XGBoost模型502輸出的決策機率分布為[0.1, 0.4, 0.8]，及DNN模型503輸出的決策機率分布為[0.7, 0.2, 0.1]。這些決策機率分布經由注意力層504的神經網路(NN)405可得出對應的權重w1~w3(例如，是0.1、0.3、0.6)。最後，最終機率分布為[0.49, 0.29, 0.22]。For example, FIG. 5 is a schematic diagram of a solution decision according to an embodiment of the present invention. Please refer to FIG. 5 , the decision probability distribution output by the random forest model 501 is [0.4, 0.5, 0.1], the decision probability distribution output by the XGBoost model 502 is [0.1, 0.4, 0.8], and the decision probability distribution output by the DNN model 503 is [0.7, 0.2, 0.1]. Corresponding weights w1 to w3 (eg, 0.1, 0.3, 0.6) can be derived from these decision probability distributions through the neural network (NN) 405 of the attention layer 504 . Finally, the final probability distribution is [0.49, 0.29, 0.22].

接著，決策模組114可依據此最終機率分布決定那些障礙修復方案中的一者作為排除當前障礙的解決方案。具體而言，排序模組118可對此最終機率分布中的數個機率降冪排序(步驟S247)，且決策模組114可依據排序結果決定嘗試那些障礙修復方案的排序。換句而言，處理器130可依據順序來分別嘗試不同解決方案，且機率越高者越先嘗試。Then, the decision module 114 may determine one of those obstacle repair solutions as a solution to eliminate the current obstacle according to this final probability distribution. Specifically, the sorting module 118 may sort the probabilities in the final probability distribution in descending order (step S247 ), and the decision module 114 may decide to try the sorting of those obstacle repair solutions according to the sorting result. In other words, the processor 130 may try different solutions according to the order, and the one with a higher probability will be tried first.

舉例而言，最終機率分布為[0.3, 0.6, 0.1]，則機率為0.6的方案最先嘗試，接著是機率為0.3的方案，最後是0.1的方案。藉此，可確保行動網路服務在生命週期內可正常運作。For example, if the final probability distribution is [0.3, 0.6, 0.1], the plan with probability 0.6 is tried first, then the plan with probability 0.3, and finally the plan with probability 0.1. In this way, the normal operation of the mobile network service during the life cycle can be ensured.

綜上所述，在本發明實施例的行動網路的生命週期管理系統及方法中，將網路服務特徵透過非監督式學習模型進行編碼，使表達行動網路服務的向量帶有行動網路服務之間關係的資訊。接著，透過訓練人工智慧模型(例如，機器學習、深度學習、注意力層)計算出確保行動網路服務正常運作方法的機率分布，並以機率高的修復方案為優先，從而得到最佳方法嘗試順序。藉此，加速決策修復障礙之行為，並使服務快速且有效恢復能正常運作，以降低挑選適當方法的時間成本，減少行動網路服務處於錯誤狀態的時間，提升行動網路服務的品質。To sum up, in the mobile network life cycle management system and method according to the embodiments of the present invention, the network service features are encoded through an unsupervised learning model, so that the vector expressing the mobile network service has the mobile network Information about the relationship between services. Then, by training an artificial intelligence model (eg, machine learning, deep learning, attention layer) to calculate the probability distribution of the method to ensure the normal operation of the mobile network service, and give priority to the repair plan with high probability, so as to obtain the best method to try order. In this way, the decision-making to repair the obstacle is accelerated, and the service can be quickly and effectively restored to normal operation, so as to reduce the time cost of selecting an appropriate method, reduce the time when the mobile network service is in an error state, and improve the quality of the mobile network service.

雖然本發明已以實施例揭露如上，然其並非用以限定本發明，任何所屬技術領域中具有通常知識者，在不脫離本發明的精神和範圍內，當可作些許的更動與潤飾，故本發明的保護範圍當視後附的申請專利範圍所界定者為準。Although the present invention has been disclosed above by the embodiments, it is not intended to limit the present invention. Anyone with ordinary knowledge in the technical field can make some changes and modifications without departing from the spirit and scope of the present invention. Therefore, The protection scope of the present invention shall be determined by the scope of the appended patent application.

100:生命週期管理系統 110:儲存器 111:資料蒐集模組 112:特徵編碼模組 113:特徵向量產生模組 114:決策模組 115:第一推論模組 116:第二推論模組 117:決策權重分配模組 118:排序模組 S210~S240、S241~S247:步驟 301:非監督式學習模組 501:隨機森林模型 502: XGBoost模型 503: DNN模型 504:注意力層 505:神經網路(NN) w1~w3:權重 100: Lifecycle Management System 110: Storage 111: Data collection module 112: Feature coding module 113: Feature vector generation module 114: Decision Module 115: The first inference module 116: Second inference module 117: Decision Weight Distribution Module 118: Sorting Module S210~S240, S241~S247: Steps 301: Unsupervised Learning Module 501: Random Forest Model 502: XGBoost Model 503: DNN Model 504: Attention Layer 505: Neural Networks (NN) w1~w3: weight

圖1是依據本發明一實施例的行動網路的生命週期管理系統的元件方塊圖。圖2是依據本發明一實施例的行動網路的生命週期管理方法的流程圖。圖3是依據本發明一實施例的特徵向量產生的示意圖。圖4是依據本發明一實施例的方案決策的流程圖。圖5是依據本發明一實施例的方案決策的示意圖。 FIG. 1 is a block diagram of components of a life cycle management system of a mobile network according to an embodiment of the present invention. FIG. 2 is a flowchart of a life cycle management method of a mobile network according to an embodiment of the present invention. FIG. 3 is a schematic diagram of feature vector generation according to an embodiment of the present invention. FIG. 4 is a flow chart of solution decision according to an embodiment of the present invention. FIG. 5 is a schematic diagram of a solution decision according to an embodiment of the present invention.

S210~S240:步驟S210~S240: Steps

Claims

一種行動網路的生命週期管理方法，包括：取得一行動網路服務相關的多筆監控資料，其中每一該監控資料相關於該行動網路服務的資源狀態；對該些監控資料編碼以產生多個初步向量，將該些監控資料中的非數值資料透單熱(One-Hot)編碼轉換成該些初步向量中的至少一者，並使用四維單熱編碼向量區別該些監控資料，再將該些四維單熱編碼向量與最大維向量合併，以使該些初步向量的維度相同；將該些初步向量合併後，透過非監督式學習(Unsupervised Learning)產生一特徵向量；將該特徵向量透過多個推論模型分別得出多個決策機率分布，其中該特徵向量是透過應用程式介面(Application Programming Interface，API)介接以取得發生障礙之行動網路服務樣本數據，且該些推論模型採用不同的機器學習演算法，並分析用以訓練該些行動網路服務樣本數據以獲得規律，並透過該些規律對未知資料預測，且每一該決策機率分布包括對應於多個障礙修復方案的機率，其中該些障礙修復方案是ETSI或其他協會或規範提供針對網路障礙的修復方案，且每一該決策機率分布是相關於使用該些障礙修復方案解決障礙的機率；將該些決策機率分布作為一注意力層(attention layer)的輸入以得出一最終機率分布，其中，該注意力層包含至少一神經網路；以及依據該最終機率分布決定該些障礙修復方案中的一者，並透過該神經網路決定該些決策機率分布的對應權重。 A life cycle management method of a mobile network, comprising: obtaining a plurality of monitoring data related to a mobile network service, wherein each monitoring data is related to the resource status of the mobile network service; encoding the monitoring data to generate A plurality of preliminary vectors, the non-numeric data in the monitoring data are converted into at least one of the preliminary vectors through one-hot (One-Hot) encoding, and the four-dimensional one-hot encoding vector is used to distinguish the monitoring data, and then Combine these four-dimensional one-hot encoded vectors with the largest dimension vector, so that the dimensions of these preliminary vectors are the same; after combining these preliminary vectors, a feature vector is generated through unsupervised learning; the feature vector is A plurality of decision probability distributions are respectively obtained through a plurality of inference models, wherein the feature vector is obtained through an application programming interface (Application Programming Interface, API) interface to obtain the sample data of mobile network services with obstacles, and the inference models use Different machine learning algorithms are used to analyze and train the mobile network service sample data to obtain patterns, and to predict unknown data through these patterns, and each of the decision probability distributions includes a probability distribution corresponding to a plurality of obstacle repair solutions. probabilities, wherein the barrier fixes are fixes for network barriers provided by ETSI or other associations or specifications, and each of the decision probability distributions is related to the probability of using the barrier fixes to resolve the barriers; the decision probabilities distribution as input to an attention layer to derive a final probability distribution, wherein the attention layer includes at least one neural network; and determining one of the obstacle repair solutions according to the final probability distribution, and determining the corresponding weights of the decision probability distributions through the neural network.

如請求項1所述的行動網路的生命週期管理方法，其中該些推論模型分別是基於隨機森林(Random Forest)、XGBoost(Extreme Gradient Boosting)及深度神經網路(Neural Network)。 The life cycle management method of a mobile network as claimed in claim 1, wherein the inference models are based on Random Forest, XGBoost (Extreme Gradient Boosting) and Deep Neural Network respectively.

如請求項1所述的行動網路的生命週期管理方法，其中得出該最終機率分布的步驟包括：對該些決策機率分布分別乘上對應的該些權重，以得出該最終機率分布。 The life cycle management method of a mobile network according to claim 1, wherein the step of obtaining the final probability distribution includes: multiplying the decision probability distributions by the corresponding weights respectively to obtain the final probability distribution.

如請求項1所述的行動網路的生命週期管理方法，其中依據該最終機率分布決定該些障礙修復方案中的一者的步驟包括：對該最終機率分布中的多個機率降冪排序；以及依據一排序結果決定嘗試該些障礙修復方案的順序。 The life cycle management method of a mobile network according to claim 1, wherein the step of determining one of the obstacle repair solutions according to the final probability distribution comprises: descending order of a plurality of probabilities in the final probability distribution; and determining the order of trying the obstacle repair solutions according to a sorting result.

一種行動網路的生命週期管理系統，包括：一儲存器，儲存多個軟體模組；以及一處理器，耦接該儲存器，載入且執行該些軟體模組，其中該些軟體模組包括：一資料蒐集模組，取得一行動網路服務相關的多筆監控資料，其中每一該監控資料相關於該行動網路服務的資源狀態；一特徵編碼模組，對該些監控資料編碼以產生多個初步向量，將該些監控資料中的非數值資料透單熱(One-Hot)編碼轉換成該些初步向量中的至少一者，並使用四維單熱編碼向量區別該些監控資料，再將該些四維單熱編碼向量與最大維向量合併，以使該些初步向量的維度相同；一特徵向量產生模組，將該些初步向量合併後，透過非監督式學習產生一特徵向量；以及一決策模組，將該特徵向量透過多個推論模型分別得出多個決策機率分布，其中該特徵向量是透過應用程式介面(Application Programming Interface，API)介接以取得發生障礙之行動網路服務樣本數據，且該些推論模型採用不同的機器學習演算法，並分析用以訓練該些行動網路服務樣本數據樣本以獲得規律，並透過該些規律對未知資料預測，將該些決策機率分布作為一注意力層的輸入以得出一最終機率分布，其中，該注意力層包含至少一神經網路，並依據該最終機率分布決定該些障礙修復方案中的一者，其中每一該決策機率分布包括對應於多個障礙修復方案的機率，其中該些障礙修復方案是ETSI或其他協會或規範提供針對網路障礙的修復方案，且每一該決策機率分布是相關於使用該些障礙修復方案解決障礙的機率；一決策權重分配模組，透過該神經網路決定該些決策機率分布的對應權重。 A life cycle management system of a mobile network, comprising: a storage, storing a plurality of software modules; and a processor, coupled to the storage, loading and executing the software modules, wherein the software modules Including: a data collection module to obtain a plurality of monitoring data related to a mobile network service, wherein each monitoring data is related to the resource status of the mobile network service; a feature encoding module for encoding the monitoring data to generate a plurality of preliminary vectors, converting the non-numerical data in the monitoring data into at least one of the preliminary vectors by one-hot encoding, And use the four-dimensional one-hot encoding vector to distinguish the monitoring data, and then combine these four-dimensional one-hot encoding vectors with the largest dimension vector, so that the dimensions of the preliminary vectors are the same; a feature vector generation module, these preliminary vectors After merging, a feature vector is generated through unsupervised learning; and a decision module is used to obtain a plurality of decision probability distributions from the feature vector through a plurality of inference models, wherein the feature vector is obtained through an application programming interface (Application Programming Interface). , API) to obtain the sample data of the mobile network service where the obstacle occurs, and the inference models use different machine learning algorithms, and analyze and train the sample data samples of the mobile network service to obtain the rules, and through the The laws predict unknown data, and the decision probability distributions are used as the input of an attention layer to obtain a final probability distribution, wherein the attention layer includes at least one neural network and determines the final probability distribution according to the final probability distribution. one of the barrier fixes, wherein each of the decision probability distributions includes a probability corresponding to a plurality of barrier fixes, wherein the barrier fixes are fixes for network barriers provided by ETSI or other associations or specifications, and Each of the decision probability distributions is related to the probability of solving obstacles using the obstacle repair solutions; a decision weight assignment module determines the corresponding weights of the decision probability distributions through the neural network.

如請求項5所述的行動網路的生命週期管理系統，其中該些推論模型分別是基於隨機森林、XGBoost及深度神經網路。 The life cycle management system of a mobile network as claimed in claim 5, wherein the inference models are respectively based on random forest, XGBoost and deep neural network.

如請求項5所述的行動網路的生命週期管理系統，其中該決策模組對該些決策機率分布分別乘上對應的該些權重，以得出該最終機率分布。 The life cycle management system of a mobile network according to claim 5, wherein the decision module multiplies the decision probability distributions by the corresponding weights to obtain the final probability distribution.

如請求項5所述的行動網路的生命週期管理系統，其中該決策模組對該最終機率分布中的多個機率降冪排序，並依據一排序結果決定嘗試該些障礙修復方案的順序。 The life cycle management system of a mobile network according to claim 5, wherein the decision module sorts the probabilities in the final probability distribution in descending order, and determines the order of trying the obstacle repair solutions according to a sorting result.