TWI735942B - System and method for predicting and preventing obstacles of network communication equipment based on machine learning - Google Patents

Abstract

The invention discloses system and method for predicting and preventing obstacles of network communication equipment based on machine learning. First, collect performance management (PM) data, configuration management (CM) data and obstacle events of network communication equipment, convert the performance management (PM) data and the configuration management (CM) data into a key performance indicator (KPI), and use the key performance indicator (KPI) as an input factor and the obstacle event as an output result to build a machine learning model as an obstacle diagnosis model. Second, enter a new key performance indicator (KPI) converted from new performance management (PM) data and new configuration management (CM) data into the obstacle diagnosis model to predict or diagnose possibility of obstacles of the network communication equipment, and then adjust the equipment parameters of the network communication equipment to prevent the occurrence of obstacles.

Description

System and method for predicting and preventing obstacles in network communication equipment based on machine learning

The present invention relates to a technology for predicting and preventing obstacles in network communication equipment, in particular to a system and method for predicting and preventing obstacles in network communication equipment based on machine learning.

In the past, when the small cell equipment was obstructed, the small cell equipment itself sent obstruction events and alarms to the network management system to notify the maintenance personnel to investigate the root cause of the small cell equipment obstruction. In addition, before the small base station equipment is obstructed, the network management system will also collect information such as the performance and configuration of the small base station equipment. These information are all time-dependent data, that is, have time-dependent characteristics, and the small base station equipment occurs The reason for the obstacle is also related to the change of the performance and configuration of the small base station equipment to a considerable extent.

In an existing technology, a system and method for the root cause analysis of telecommunications network obstacles is proposed. It mainly applies information sharing technology to the root cause analysis of network obstacles. Network administrators serve as observation points for network problems, and by automatically associating, collecting and sharing network administrators’ dynamics and network obstacle maintenance records to achieve the purpose of root cause analysis of network obstacles. However, this existing technology must be based on the traditional rule-based model, and network management personnel need to report obstacle information and maintenance records regularly, so that the goal of intelligent maintenance (AIOps) cannot be achieved.

Therefore, how to provide a novel or innovative technology for predicting and preventing obstacles in network communication equipment (such as small base station equipment) has become a major research topic for those skilled in the art.

The present invention provides a novel or innovative system and method based on machine learning to predict and prevent obstacles in network communication equipment, which can quickly predict or diagnose whether network communication equipment has obstacles.

The system for predicting and preventing obstacles in network communication equipment based on machine learning of the present invention includes: at least one network communication equipment; State management (Configuration Management; CM) data and failure events; and a diagnostic module, which calculates or converts the performance management (PM) data and configuration management (CM) data of the network communication equipment collected by the network management module The corresponding key performance indicator (KPI) is established, and the diagnosis module uses the key performance indicator (KPI) as the input factor and the obstacle event as the output result to build the machine learning model as the obstacle diagnosis model. Among them, the diagnosis module The set is to input the new key performance indicators (KPI) calculated or converted from the new performance management (PM) data and the new configuration management (CM) data of the network communication equipment into the obstacle diagnosis model, so that the diagnosis module can pass the obstacle Diagnosis model utilizes new key performance Performance indicators (KPI) predict or diagnose whether the network communication equipment may be malfunctioning. When it is predicted or diagnosed that the network communication equipment may be malfunctioning, the diagnostic module adjusts the equipment parameters of the network communication equipment to prevent or prevent the network Obstacles to communication equipment.

The method for predicting and preventing network communication equipment obstacles based on machine learning of the present invention includes: making a network management module collect performance management (PM) data, configuration management (CM) data and obstacle events from at least one network communication equipment ; Make a diagnostic module calculate or convert the performance management (PM) data and configuration management (CM) data of the network communication equipment collected by the network management module into the corresponding key performance indicators (KPI), and make the diagnostic model The group uses key performance indicators (KPI) as input factors and obstacle events as output results to build machine learning models as obstacle diagnosis models; and the diagnosis module will be set by the new performance management (PM) data and new groups of network communication equipment The new key performance indicators (KPI) calculated or converted from the state management (CM) data are input into the obstacle diagnosis model, so that the diagnosis module can use the new key performance indicators (KPI) to predict or diagnose whether the network communication equipment is Obstacles may occur. When it is predicted or diagnosed that the network communication equipment may have an obstacle, the diagnostic module is asked to adjust the equipment parameters of the network communication equipment to prevent or prevent the network communication equipment from malfunctioning.

In order to make the above-mentioned features and advantages of the present invention more comprehensible, embodiments are specifically described below in conjunction with the accompanying drawings. In the following description, the additional features and advantages of the present invention will be partially explained, and these features and advantages will be partly known from the description, or can be learned by practicing the present invention. The features and advantages of the present invention are realized and achieved by means of the elements and combinations specifically pointed out in the scope of the patent application. It should be understood that both the foregoing general description and the following detailed description are only illustrative and explanatory, and are not intended to limit the scope of the present invention.

1. A system based on machine learning to predict and prevent obstacles in network communication equipment

10‧‧‧Network communication equipment

20‧‧‧Network Management Module

21‧‧‧Server

22‧‧‧Database

30‧‧‧Diagnostic Module

31‧‧‧Weight evaluation algorithm

40‧‧‧Disorder diagnosis model

50‧‧‧User Equipment

60‧‧‧Security Gateway

70‧‧‧Evolution Node B Gateway

80‧‧‧Core network

90‧‧‧Key Performance Indicator (KPI) Formula Database

100‧‧‧Machine learning model architecture

101‧‧‧Obstacle Event Log

102‧‧‧Barrier label

103‧‧‧PM and CM data

104‧‧‧Key Performance Indicators (KPI)

105‧‧‧New PM and CM data

106‧‧‧New Key Performance Indicators (KPI)

110‧‧‧Data Collection

111‧‧‧Training Data Set

112‧‧‧Verification Data Set

120‧‧‧Machine Learning Model

121‧‧‧softmax algorithm

122‧‧‧Exit layer

CM‧‧‧Configuration Management (Data)

IPSec‧‧‧Internet Security Protocol (connection interface)

KPI‧‧‧Key Performance Indicators

PM‧‧‧Performance Management (Data)

S1‧‧‧Connecting interface

Steps S11 to S15, S21 to S24‧‧‧

WAN‧‧‧Wide Area Network

Figure 1 is a schematic diagram of the system of the present invention based on machine learning to predict and prevent obstacles in network communication equipment and the network architecture of the network communication equipment; Figure 2 is the network communication equipment of the present invention and its network architecture Figure 3 is a schematic diagram of an embodiment of the system for predicting and preventing obstacles in network communication equipment based on machine learning in the present invention; Figure 4 is the machine learning model architecture and machine learning of the present invention The schematic diagram of the module; Figure 5A is the flow chart of the construction of the obstacle diagnosis model in the method of predicting and preventing network communication equipment based on machine learning in the present invention; Figure 5B is the prediction and prevention based on machine learning of the present invention The flow chart for the diagnosis of the obstacles of the network communication equipment in the method for the obstacles of the network communication equipment; Fig. 6A, Fig. 6B and Fig. 6C are respectively the model verification accuracy rate and the model verification accuracy rate of the first test case in the present invention A graph with the model verification recall rate; and Figures 7A, 7B, and 7C are the graphs of the model verification accuracy rate, model verification accuracy rate, and model verification recall rate of the second test case of the present invention, respectively .

The following describes the implementation of the present invention with specific specific embodiments. Those familiar with this technology can understand the other advantages and effects of the present invention from the content disclosed in this specification. It can also be implemented or applied by other different specific equivalent embodiments.

Figure 1 is a schematic diagram of the system 1 and the network architecture of the network communication device 10 based on machine learning to predict and prevent obstacles in network communication equipment of the present invention. At the same time, the main technical content of the system 1 based on machine learning to predict and prevent obstacles in network communication equipment is as follows, and the rest of the technical content is the same as the description in Figures 2 to 7C, and will not be repeated here.

As shown in Figure 1, the system 1 based on machine learning to predict and prevent obstacles in network communication equipment mainly includes at least one (such as plural) network communication equipment 10, a network management module 20, a diagnostic module 30, and a key Performance indicator (KPI) formula database 90 (see Figure 3), may further include a plurality of user equipment (UE) 50, at least one security gateway (Security Gateway; SecGW) 60, and at least one evolved node B gateway (eNode Gateway) 70 and a core network 80. For example, the network communication device 10 may be a small cell (Small Cell) device, or may be a wireless base station (Wi-Fi AP) device or a home gateway (Home Gateway) device and other customer premise equipment (Customer Premise Equipment; CPE). The network management module 20 can be an EMS (Element Management System) network management module or an EMS network management system, and has a server 21 and a database 22 (see Figure 3). The diagnosis module 30 may be a diagnosis server (Diagnosis Server; DS), and has a weight evaluation algorithm 31. However, the present invention is not limited to this.

The network management module 20 can collect performance management (PM) data, configuration management (CM) data, and obstacle events from the network communication device 10. The diagnostic module 30 can calculate or convert the performance management (PM) data and configuration management (CM) data of the network communication equipment 10 collected by the network management module 20 into corresponding key performance indicators (KPI), and the diagnostic model The group 30 can use key performance indicators (KPIs) as input factors and obstacle events as output results to build a machine learning model 120 (see Figure 4) as the obstacle diagnosis model 40. At the same time, the diagnostic module 30 can be used by the network communication equipment 10 The new key performance indicator (KPI) calculated or converted from the new performance management (PM) data and the new configuration management (CM) data is input into the obstacle diagnosis model 40, so that the diagnosis module 30 uses the new key through the obstacle diagnosis model 40 Performance indicators (KPI) predict or diagnose whether the network communication device 10 may be malfunctioning. When it is predicted or diagnosed that the network communication device 10 may be malfunctioning, the diagnostic module 30 adjusts the device parameters of the network communication device 10 to prevent Or to prevent the network communication equipment 10 from malfunctioning.

The network communication device 10 can be connected to a secure gateway (SecGW) 60 through a network such as a wide area network (Wide Area Network; WAN), and the secure gateway (SecGW) 60 can use, for example, an S1 connection interface or an IPSec (Internet Protocol Security (Internet Security Protocol) connection interface to connect to the evolved node B gateway 70, and then the evolved node B gateway 70 extends to, for example, the EPC (Evolved Packet Core) network through the S1 connection interface, for example Road core network 80.

The network management module 20 (such as EMS network management module or EMS network management system) manages the supply, monitoring, parameter setting or (and) key performance indicators (KPIs) of all network communication equipment 10 (such as small base station equipment) through the communication protocol ( Log) to the network management module 20, and then the diagnostic module 30 according to the performance management (PM) data, configuration management (CM) data and the LTE (Long Term Evolution; Long Term Evolution) key performance indicator (KPI) specifications defined by 3GPP Calculate the key performance indicators (KPI) of each network communication device 10.

Furthermore, the 5th generation (5G) network was initially built to enhance service coverage and provide a stable and efficient wireless network environment, using network communication equipment such as long-term evolution technology (LTE) small base station equipment. 10 plays an important role. Also, in view of the global deployment of networks Communication equipment 10 (such as small base station equipment) has a large increase trend. When network communication equipment 10 is obstructed, telecom operators need to immediately identify the obstacle and dispatch workers to eliminate the problem to ensure the normal operation of its service. However, there are many factors that cause obstacles to the network communication equipment 10, and the obstacles may come from the network communication equipment 10 itself, the evolved node B gateway 70 or the core network 80, or between the network communication equipment 10 and each other. Communication network barriers. Therefore, finding out the cause of the obstacle of the network communication device 10 in the shortest time will effectively help the telecommunication operator to improve the user experience.

Therefore, the present invention can interface with the information of the network communication equipment 10 in the initial stage, and collect a large amount of performance management (PM) data, configuration management (CM) data, equipment records (Log) and failure events of the network communication equipment 10 And other information to use the machine learning model 120 shown in Figure 4 (for example, a recurrent neural network model including a complex recurrent neural network (Recurrent Neural Network; RNN) layer and a complex number of neurons) to construct a machine learning model architecture (such as Recursive neural network architecture), and then calculate the most suitable category weight or time sequence weight, and then quickly diagnose the probability of failure of the network communication device 10.

In addition, with the evolution of wireless network architecture, data routing and exchange between network communication devices 10 (such as small base station devices) will become more and more complicated. Therefore, if the fault diagnosis of the network communication device 10 is performed in a traditional rule-based mode, it will take a lot of time to trace the status of each node. Therefore, the present invention adopts the emerging machine learning model 120 (see Figure 4) or machine learning technology, which will help, for example, telecom operators in the 5th generation (5G) network or more advanced network, by diagnosis The module establishes a knowledge-based obstacle diagnosis method to achieve the goal of Artificial Intelligence for IT Operations (AIOps).

In addition, the present invention provides or designs the weight evaluation algorithm 31 related to obstacle events, which can improve the accuracy of model learning in the data set of the unbalanced network communication device 10. therefore, The present invention can be used for obstacle detection and prevention of various network communication equipment 10 to reduce the maintenance cost and time of the network communication equipment 10 by the telecommunication operator.

Figure 2 is a schematic diagram of the network communication equipment 10 (such as a small base station equipment) of the present invention and related network obstacle characteristics in its network architecture. As shown in the figure, the network communication device 10 may include, for example, the following 11 obstacle events: [1] PCI (Peripheral Component Interconnect) conflict of the device, [2] The device does not obtain PCI (Peripheral Component Interconnect) , [3] The CPU (Central Processing Unit) load of the device is too high, [4] The memory load of the device is too high, [5] The temperature of the device exceeds the critical value, [6] The time synchronization of the device fails, [7] Device IP (Internet Protocol) address conflict, [8] Device MAC (Media Access Control) address conflict, [9] Device firmware update failed or missing , [10] The X2 connection interface of the device handover (Handover) was disconnected, [11] The device failed to upload PM (performance management) data and configuration management (CM) data. The 11 obstacle events mentioned above are all the obstacle event log 101 or event log (see Figure 4) generated by the network communication equipment 10 (such as small base station equipment). These obstacle events will seriously affect the network communication equipment 10 to provide The quality of network services (such as wireless network services) even makes the network communication device 10 unable to provide network services.

The network management module 20 (such as the EMS network management module or the EMS network management system) to manage the network communication equipment 10 (such as the small base station equipment) may include, for example, the following 4 obstacles: [1] Failure to install the equipment, [2] Failed to obtain the parameters of the equipment, [3] Failed to set the parameters of the equipment, [4] Failed to set the parameters of the equipment's disaster prevention warning system (Public Warning System; PWS). The above 4 obstacle events are all the obstacle event log 101 or event log (see Figure 4) generated by the network management module 20 (such as EMS network management module or EMS network management system) for equipment management, although it does not affect the network communication equipment 10 Provide network services (such as wireless network services), but cannot monitor and optimize network communication in real time Parameters of the performance management (PM) data of the telecommunications equipment 10.

The security gateway (SecGW) 60 of a network (such as a wide area network or a local network) can include, for example, the following 3 obstacles: [1] IPSec (Internet Security Protocol) tunnel establishment failure , [2] S1-MME (Mobility Management Entity) connection interface is disconnected, [3] IPSec channel is unexpectedly disconnected. The above three obstacle events are all the obstacle event log 101 or event log (see Figure 4) generated by the network communication equipment 10 (such as small base station equipment) connected to the secure gateway (SecGW) 60. These obstacle events will be Seriously affect the network communication equipment 10 to provide network services.

Figure 3 is a schematic diagram of an embodiment of the system 1 for predicting and preventing obstacles in network communication equipment based on machine learning of the present invention. As shown in the figure, the network communication equipment 10 (such as small base station equipment) regularly transmits performance management (PM) data and configuration management (CM) data to the network management module 20 (such as EMS network management module or EMS) through the communication protocol. The server 21 of the network management system (such as Auto Configuration Server (ACS)) to store performance management (PM) data and configuration management (CM) data in the network management module 20 (or server 21) ) In the database 22 (such as the ACS database), and the network communication device 10 forwards the obstacle event log 101 (event record log) related to the obstacle event that has occurred to the network management module 20 for recording on the server 21 or Database 22.

In addition, the server 21 of the network management module 20 receives, for example, more than two hundred performance management (PM) data and configuration management (CM) parameters, and uses the Representational State Transfer Application Programming Interface (Representational State Transfer Application Programming Interface). ; Rest API) The parameters of performance management (PM) data and configuration management (CM) data are sent to the diagnostic module 30 (such as a diagnostic server), and the diagnostic module 30 uses the key performance indicator (KPI) formula database 90 Calculate or convert the parameters of performance management (PM) data and configuration management (CM) data into relative According to the key performance indicators (KPI), the wireless network quality information of the network communication equipment 10 (such as small base station equipment) is presented based on the key performance indicators (KPI). Moreover, the diagnostic module 30 (such as a diagnostic server) records 89 key performance indicators (KPIs) and event logs of all equipment for a period of time (such as every 15 minutes), and at the same time performs data pre-processing, merging, and Model building and obstacle prediction.

Table 1 below shows the key performance indicators (KPI) and formulas of the present invention. In the data processing of the key performance indicators (KPI) of the network communication equipment 10 (such as small base station equipment) and the obstacle event log 101 (event record log) shown in Figure 3, the wireless network quality of the network communication equipment 10 The information can be presented by, for example, 89 key performance indicators (KPIs), and can also include time and device identification numbers (ID). Each key performance indicator (KPI) has a corresponding formula, and the parameters in the formula are the parameters of performance management (PM) data and configuration management (CM) data regularly brought up by the network communication device 10.

Table 2 below is the content of the obstacle event log of the present invention. For example, the obstacle event log of the equipment can be divided into 18 types of obstacles. Each obstacle event log is recorded in the database 22 (such as the ACS database) shown in Figure 3, and the diagnostic module 30 (such as the diagnostic servo Device) will periodically read the latest log of the database 22. From the content of the obstacle event log, you can know the date of the obstacle occurrence Time (e.g. 2018-06-23 09:05:18), device identification code (e.g. B22539-LTE) and event content (e.g. s1 connection interface disconnection from ip...). From the information of the above-mentioned key performance indicators (KPI) and the content of the obstacle event log, the date and time of the occurrence of the obstacle can be combined with the equipment identification code (ID). In addition, if two obstacles occur on the same device within a period of time (such as 15 minutes), there will be two identical key performance indicator (KPI) information corresponding to the content of the two obstacle event logs during the merging process.

In the data pre-processing part, the parameters of performance management (PM) data and configuration management (CM) data brought by the network communication device 10 shown in Figure 3 occasionally cause a small number of parameters to be used in the calculation of key performance indicators (KPI). Make the denominator of the key performance indicator (KPI) 0, so that the calculation result is wrong. Therefore, the diagnostic module 30 can use the average, mode or zero of each key performance indicator (KPI) of each obstacle event category to fill the empty value of the key performance indicator (KPI) according to the respective obstacle event category. In addition, the diagnostic module 30 can use normalization to optimize the value of key performance indicators (KPI), because some key performance indicators (KPI) have extremely large values, such as Uplink throughput and Downlink throughput Quantity (for example, in Kbit); or, the diagnostic module 30 For some key performance indicators (KPI) values are simply true and false values (such as True:1 or Flase:0). The value of each key performance indicator (KPI) can be scaled to [0,1]. To speed up the subsequent model convergence. Then, the diagnostic module 30 can use Label Encoding and One Hot Encoding to convert the parsed network communication equipment 10 (such as category 18+1) obstacle event category and normal category into ( Such as N*19 dimension) multi-dimensional matrix.

The present invention finds from the numerical changes between multiple key performance indicators (KPI) that a single obstacle event can be linked to the numerical changes of multiple key performance indicators (KPI) at the same time. For example, if a network communication device 10 (such as a small base station device) shown in Figure 2 is connected to a secure gateway (SecGW) 60, the S1-MME (S1-Mobile Management Entity) connection interface is suddenly disconnected, At this time, the traffic of Bytes received of S1 link interface of the device will decrease, and the availability rate of the device will also decrease. However, due to the 10 ( If there are still related user equipment 50 (see Figure 1) on the small base station equipment), the number of connections of the user equipment 50 remains unchanged. Therefore, it is known from the above information that the key performance indicators (KPI) are between The value change of is related to the obstacle event. Therefore, the diagnostic module 30 of the present invention uses, for example, a long short-term memory (Long Short-Term Memory, LSTM) model as a basis to find out the correlation between key performance indicators (KPIs), and then builds as shown in Figure 4 A machine learning model framework 100 (such as a recurrent neural network model) with a machine learning model 120 (such as a recurrent neural network model).

FIG. 4 is a schematic diagram of the machine learning model architecture 100 (such as the recurrent neural network model architecture) and its machine learning model 120 (such as the recurrent neural network model) of the present invention. As shown in the figure, in the machine learning model architecture 100, the obstacle event log 101 (event record log) can be processed (such as pre-processing or merging) to generate obstacle tags 102 and then sent to the data set 110, and Performance management (PM) data and configuration management (CM) data 103 are processed (such as pre-processing or merging) to generate key performance indicators (KPI) 104 and then sent to the data set 110.

In addition, the data set 110 after data processing is divided into a training data set 111 and a verification data set 112. The training data set 111 can be used to pre-train a recurrent neural network (RNN) such as a two-layer long short-term memory (LSTM) , For example, a four-layer neural network (Neural Network; NN) is added, and then the softmax algorithm 121 analyzes, for example, 18 obstacle categories and normal categories. During the training process of the machine learning model 120, the accuracy of the machine learning model 120 can be verified by the verification data set 112, and after multiple rounds of iterative training, each hidden layer (such as a two-layer recurrent neural network) Road (RNN) plus four layers of neural network (NN) (total six hidden layers) are added to the dropout layer 122 to prevent overfitting of the machine learning model 120. Then, the diagnostic module 30 verifies the accuracy of the machine learning model 120 based on the information of the new key performance indicator (KPI) 106 generated by the new performance management (PM) data and the new configuration management (CM) data 105.

Table 3 below is the weight evaluation algorithm 31 of the obstacle event in Figure 1 of the present invention. For example, there are 18 types of equipment obstacles, and the number of occurrences of each obstacle is not the same. Therefore, in the training process of reconciling the machine learning model 120 (such as the recurrent neural network model) shown in Figure 4, it can be The diagnostic module 30 uses the weight evaluation algorithm 31 to optimize or adjust the category weight of each obstacle event occurrence, so as to improve the classification accuracy.

)。每個障礙事件類別之類別權重為

，且類別權重之數值最小為1，並以log作為權重評估的正規化。權重評估演算法31主要表示當某一類障礙事件發生次數愈多時，機器學習模型120之訓練過程中，障礙事件之類別權重愈小；反之，當某一類障礙事件發生次數愈少時，機器學習模型120之訓練過程中，障礙事件之類別權重愈大。例如，class_sampleSize={c0：2813,c1：78,c2：1014,c3：510,c4：7914,c5：348}，total_sample=12677，mu=12677/7914=1.6018，

，並依此類推c1至c5的權重，其中c0至c5表示類別。 For example, in the above Table 3 and Figure 1 in the weight evaluation algorithm for obstacle events 31, you can define class_sampleSize to represent the number of occurrences of each obstacle category, total_sample to represent the total number of times, and mu to represent the number of occurrences in class_sampleSize that account for the largest number of total_sample Ratio (i.e.

). The category weight of each obstacle event category is

, And the minimum value of the category weight is 1, and log is used as the normalization of the weight evaluation. The weight evaluation algorithm 31 mainly means that when the number of occurrences of a certain type of obstacle event increases, the weight of the type of obstacle event becomes smaller during the training process of the machine learning model 120; conversely, when the number of occurrences of a certain type obstacle event decreases, the machine learning In the training process of the model 120, the larger the weight of the obstacle event category. For example, class_sampleSize={c0:2813,c1:78,c2:1014,c3:510,c4:7914,c5:348}, total_sample=12677, mu=12677/7914=1.6018,

, And so on, the weights of c1 to c5, where c0 to c5 represent categories.

A specific example is given below to illustrate the operation of the present invention. First of all, as the first Figure 1 shows the network architecture of the network communication equipment 10, taking the supervision of the network communication equipment 10 (such as small base station equipment) as an example, suppose a certain telecom operator currently deploys about 750 network communication equipment 10 in the whole station Areas, including public places such as supermarkets, banquet halls, high-speed rail, hospitals and police stations throughout Taiwan. At the same time, network communication equipment 10 (such as small base station equipment) is connected to a secure gateway (SecGW) 60 through a network (such as a wide area network), and is managed by a network management module 20 (such as an EMS network management module or an EMS network management system). ) Manage functions such as installation, monitoring, parameter setting, or (and) key performance indicator (KPI) management of the network communication equipment 10 through the communication protocol.

Next, as shown in Figure 3, the network architecture of the network communication device 10 (such as a small base station device). The network communication device 10 transmits performance management (PM) data and configuration management (CM) data to the server 21 of the network management module 20 (such as the automatic configuration server (ACS) )) to store in the database 22, and forward the obstacle event log of the equipment related obstacle events to the server 21 for recording in the database 22 (such as the ACS database). In addition, the server 21 receives, for example, more than two hundred performance management (PM) data and configuration management (CM) data parameters, and transmits performance management (PM) data and parameters through the presentation layer state transition application program interface (Rest API) The configuration management (CM) data parameters are sent to the diagnostic module 30 (such as a diagnostic server), and the performance management (PM) data and configuration management (CM) data parameters are passed through the key performance indicator (KPI) by the diagnostic module 30 ) The formula database 90 is converted into a key performance indicator (KPI) to present the wireless network quality information of the network communication device 10 (such as a small base station device) according to the key performance indicator (KPI). Then, the diagnostic module 30 records 89 key performance indicators (KPIs) and event logs of all equipment for a period of time (such as every 15 minutes), and at the same time performs data pre-processing, merging, and machine learning model 120 (see Figure 4) Construction and obstacle prediction.

This specific example collects from May 1 to August 30, 107 (a total of 123 Day), related equipment performance management (PM) data, configuration management (CM) data, equipment record (Log), equipment incident log (obstacle event record) and other data. These data are calculated, processed and merged by key performance indicators (KPI) in the diagnostic module 30 (such as the diagnostic server) shown in Figure 4. For example, a total of 5,050,010 key performance indicators (KPI) information, each key performance KPI information has 89 key performance indicators (KPIs), of which 10,986 key performance indicators (KPI) information have a record of obstacles.

The machine learning model architecture 100 shown in Figure 4 (such as the recurrent neural network model architecture). For example, in the training data set 111 from May 1 to August 15, 107, there were 4,099,340 key performance indicator (KPI) information and 9,834 obstacle event records. From August 16th to August 30th, 2007, the verification data set 112 has a total of 891,815 key performance indicator (KPI) information and 1,053 obstacle event records. The test data set was on August 31, 107, with a total of 58,855 key performance indicator (KPI) information and 99 obstacle event records. The Epoch of the machine learning model 120 is about 1000 times, and the batch size is about 128 to 256. The activation function of the hidden layer of the machine learning model 120 uses ReLU (Rectified Linear Unit; modified linear unit). Compared with the softmax algorithm 121, each hidden layer has about 64 to 128 neurons, and finally is divided into 19 categories (including 18 obstacle categories and 1 normal category), of which 10 are network communication equipment (such as small base station equipment) The obstacles are shown in Figure 2.

FIG. 5A is a flow chart of the establishment of the obstacle diagnosis model 40 in the method for predicting and preventing the occurrence of obstacles in network communication equipment based on machine learning of the present invention, and refer to FIG. 1 as well. In the construction process of the obstacle diagnosis model 40 in Fig. 5A, the network management module 20 shown in Fig. 1 collects performance management (PM) data, configuration management (CM) data, and equipment records of the network communication equipment 10 Equipment information such as logs and obstacle events is calculated or converted into LTE by the diagnostic module 30 (such as a diagnostic server) shown in Figure 1 based on performance management (PM) data and configuration management (CM) data. Key performance indicators (KPIs) are key performance indicators (KPIs) to further build obstacle diagnosis models40.

For example, in step S11 in Figure 5A, the network management module 20 (such as an EMS network management module or an EMS network management system) shown in Figure 1 collects the performance from the network communication equipment 10 (such as small base station equipment) Management (PM) data, configuration management (CM) data, equipment records (Log) and equipment information such as obstacle events. Then, in step S12 of FIG. 5A, the diagnostic module 30 (such as a diagnostic server) shown in FIG. 1 is connected to or the communication network management module 20 to transfer the performance management (PM) data from the network communication device 10 Equipment information such as configuration management (CM) data is calculated or converted into corresponding key performance indicators (KPIs) through, for example, the LTE key performance indicators (KPI) specified by 3GPP. Then, in step S13, step S14, and step S15 in Figure 5A, the diagnostic module 30 establishes a weight evaluation algorithm 31 related to the obstacle event to calculate or adjust the weight of the obstacle event, and the diagnostic module 30 determines the key performance The index (KPI) is used as an input factor and the obstacle event is used as an output result to build a machine learning model as the obstacle diagnosis model 40, and then the diagnosis module 30 outputs the obstacle diagnosis model 40.

FIG. 5B is a flowchart of the obstacle diagnosis of the network communication device 10 in the method for predicting and preventing network communication device obstacles based on machine learning of the present invention, and also refer to FIG. 1. In the failure diagnosis process of the network communication equipment 10 in Fig. 5B, the network management module 20 collects the new performance management (PM) data and the new configuration management (CM) data from the network communication equipment 10 for the diagnosis model The group 30 calculates or converts the new performance management (PM) data and the new configuration management (CM) data into corresponding new key performance indicators (KPI), and then inputs the new key performance indicators (KPI) into the obstacle diagnosis model 40. Then, the diagnostic module 30 utilizes the new key performance through the obstacle diagnosis model 40 Indicators (KPI) predict or diagnose whether the network communication device 10 may occur or appear to be faulty. When it is predicted or diagnosed that the network communication device 10 may occur or appear to be faulty, the diagnostic module 30 adjusts the equipment of the network communication device 10 Parameters, such as adjusting the upper limit of the number of connections of the user equipment 50 to reduce the load of the network communication device 10, to prevent or prevent the network communication device 10 from malfunctioning.

For example, in step S21 in Figure 5B, the network management module 20 (such as EMS network management module or EMS network management system) shown in Figure 1 collects the performance from the network communication equipment 10 (such as small base station equipment) Management (PM) data, configuration management (CM) data, equipment records (Log) and equipment information such as obstacle events. Then, in step S22 of FIG. 5B, the diagnostic module 30 (such as a diagnostic server) shown in FIG. 1 is connected to or the communication network management module 20 to transfer the performance management (PM) data from the network communication device 10 Equipment information such as configuration management (CM) data is calculated or converted into corresponding key performance indicators (KPIs) through, for example, the LTE key performance indicators (KPI) specified by 3GPP. Then, in step S23 and step S24 in Figure 5B, the diagnosis module 30 inputs the network obstacle characteristics to the obstacle diagnosis model 40, and when an obstacle occurs, the diagnosis module 30 prevents network communication according to the obstacle diagnosis model 40 Obstacles of the device 10 and adjustment of the device parameters of the network communication device 10.

Fig. 6A, Fig. 6B and Fig. 6C are graphs of model verification accuracy, model verification accuracy and model verification recall rate (recall) of the first test case of the present invention, respectively. Fig. 7A, Fig. 7B and Fig. 7C are respectively the graphs of the model verification accuracy rate, the model verification accuracy rate and the model verification recall rate of the second test case of the present invention. At the same time, this specific embodiment can be divided into two types of test cases. The first test case shown in Fig. 6A to Fig. 6C does not use the obstacle event weight evaluation algorithm of the present invention (that is, the weight of the obstacle event is not adjusted). The second test case shown in Fig. 7A to Fig. 7C has the related obstacles in the present invention. The algorithm for evaluating the weight of obstacle events (that is, the weight of events with adjustment obstacles).

The first test case (the weight evaluation algorithm of the present invention is not used to adjust the weight of the obstacle event) when the number of times is 1000, the model verification accuracy shown in Figure 6A reaches 98.3% (ie 0.983) at point A1, Figure 6B The model verification accuracy rate shown in point A2 reached 27.5% (ie 0.275), the model verification recall rate shown in Figure 6C reached 98% (ie 0.98) at point A3, and the F1 score (F1-Scroe) was 42.95 . On the contrary, in the second test case (the weight evaluation algorithm of the present invention is used to adjust the weight of the obstacle event) when the number of times is 1000, the model verification accuracy shown in Figure 7A reaches 96% (ie 0.96) at point B1. The model verification accuracy rate shown in Figure 7B reached 53.5% at point B2, the model verification recall rate shown in Figure 7C reached 92.2% at point B3, and the F1 score was 67.66. From the above test results, when the second test case shown in Fig. 7A to Fig. 7C adopts the weight evaluation algorithm of the present invention to adjust the weight of obstacle events, the model verification accuracy rate is greatly improved by 26%, and the model verification is accurate. Although the accuracy rate and model verification recall rate decreased slightly from 2% to 6% compared with the first test case, they still remained above 90%, and the F1 score increased by 24.71. Therefore, the test results of the first test case and the second test case can be seen in Table 4 below.

In summary, the system and method for predicting and preventing obstacles in network communication equipment based on machine learning of the present invention can at least have the following characteristics, advantages, or technical effects.

1. The present invention adopts emerging machine learning technology, which will help, for example, telecom operators establish a knowledge-based model of obstacle diagnosis method in the 5th generation (5G) network or more advanced network, so as to achieve intelligent maintenance (AIOps). The goal of) is to predict and prevent obstacles to network communication equipment.

2. The present invention can interface with the information of network communication equipment in the initial stage, and collect a large amount of information such as performance management (PM) data, configuration management (CM) data, equipment log (Log) and obstacle events of the network communication equipment , In order to construct a machine learning model (such as a recurrent neural network model) and calculate the most suitable category weight, and then quickly predict or diagnose whether the network communication equipment is malfunctioning or its probability.

3. The present invention provides or designs a weight evaluation algorithm related to obstacle events, which can improve the accuracy of model learning in the data set of unbalanced network communication equipment. Moreover, in the training process of reconciling machine learning models (such as recurrent neural network models), the use of weight evaluation algorithms can optimize or adjust the category weights of each obstacle event to improve the classification accuracy.

4. The present invention can be used for obstacle detection and prevention of various network communication equipment, so as to reduce the maintenance cost and time of, for example, the telecommunication operator in the network communication equipment.

5. The possible applications of the present invention are the telecommunications industry, communication industry, etc., and the possible applications of the products are network management services and systems, communication equipment management services and systems, etc.

The above embodiments are only illustrative of the principles, features and effects of the present invention, and are not intended to limit the scope of implementation of the present invention. Anyone familiar with the art can comment on the above without departing from the spirit and scope of the present invention. Modifications and changes to the implementation form. Any equivalent changes and modifications made by using the content disclosed in the present invention should still be covered by the scope of the patent application. Therefore, the protection scope of the present invention should be as listed in the scope of the patent application.

1. A system based on machine learning to predict and prevent obstacles in network communication equipment

10‧‧‧Network communication equipment

20‧‧‧Network Management Module

30‧‧‧Diagnostic Module

31‧‧‧Weight evaluation algorithm

40‧‧‧Disorder diagnosis model

50‧‧‧User Equipment

60‧‧‧Security Gateway

70‧‧‧Evolution Node B Gateway

80‧‧‧Core network

CM‧‧‧Configuration Management (Data)

IPSec‧‧‧Internet Security Protocol (connection interface)

KPI‧‧‧Key Performance Indicators

PM‧‧‧Performance Management (Data)

S1‧‧‧Connecting interface

WAN‧‧‧Wide Area Network

Claims (20)

A system based on machine learning to predict and prevent obstacles in network communication equipment, including: at least one network communication device; a network management module, which collects performance management (PM) data and configuration management ( CM) data and obstacle events; and a diagnostic module, which calculates or converts the performance management (PM) data and configuration management (CM) data of the network communication equipment collected by the network management module into the corresponding key Performance indicators (KPI), and the diagnostic module uses the KPI as an input factor and the obstacle event as an output result to build a machine learning model with a hidden layer and an exit layer as the obstacle diagnosis model, where: During the training process of the machine learning model, verify the accuracy of the machine learning model with the hidden layer and the exit layer by verifying the data set, and add the hidden layer of the machine learning model to the exit layer, and Among them, the diagnostic module inputs new key performance indicators (KPIs) calculated or converted from new performance management (PM) data and new configuration management (CM) data of the network communication equipment as having the hidden layer The obstacle diagnosis model of the machine learning model with the exit layer, so that the diagnosis module uses the new key performance indicator (KPI) through the obstacle diagnosis model of the machine learning model with the hidden layer and the exit layer Predict or diagnose whether the network communication equipment may be malfunctioning. When it is predicted or diagnosed that the network communication equipment may be malfunctioning, the diagnostic module adjusts the equipment parameters of the network communication equipment to prevent or prevent the network Obstacles to communication equipment. For example, the system described in item 1 of the scope of patent application, wherein the diagnostic module uses the machine learning model to establish a knowledge-based obstacle diagnosis method, and the machine learning model includes a complex recurrent neural network (RNN) layer and complex numbers Recurrent neural network model of neurons. Such as the system described in item 1 of the scope of patent application, wherein the network communication device uploads the performance management (PM) data, the configuration management (CM) data and the device log (Log) to the network management module every period of time , And the network management module manages the installation, monitoring, parameter setting of the network communication equipment or the management of the key performance indicator (KPI). For example, the system described in item 1 of the scope of patent application, wherein the network management module has a server or database for the network communication equipment to perform the performance management (PM) data and the configuration management (CM) data Send to the network management module for storage in the server or database, and the network communication equipment forwards the obstacle event log related to the obstacle event to the network management module for recording in the server or database middle. For example, the system described in item 1 of the scope of patent application further includes a key performance indicator (KPI) formula database, in which the network management module uses the performance management (PM) ) Data and parameters of the configuration management (CM) data are sent to the diagnostic module, so that the diagnostic module uses the key performance indicator (KPI) formula database to manage the performance management (PM) data and configuration ( CM) data parameters are calculated or converted into corresponding key performance indicators (KPI). For example, the system described in item 1 of the scope of patent application, wherein the diagnostic module further establishes a weight evaluation algorithm related to the obstacle event, so that the diagnostic module uses the weight evaluation during the training process of the machine learning model Algorithm to optimize or adjust the category weight of the obstacle event. For example, the system described in item 1 of the scope of patent application, wherein the diagnostic module calculates the key performance indicator (KPI) by using the average, mode, or zero fill of the key performance indicators (KPI) of each obstacle category The null value of the key performance indicator (KPI). Such as the system described in item 1 of the scope of patent application, wherein the diagnostic module adopts normalization to optimize the value of the key performance indicator (KPI), or the value of the key performance indicator (KPI) is expressed as a true or false value . For example, in the system described in item 1 of the scope of patent application, the diagnostic module converts the obstacle event category and the normal category of the network communication equipment into a multi-dimensional matrix through a tag code and a one-hot code. For example, the system described in item 1 of the scope of the patent application, wherein the diagnostic module uses a long and short-term memory (LSTM) model as a basis to find the correlation between the key performance indicators (KPI), and then build the Machine learning model. A method for predicting and preventing obstacles in network communication equipment based on machine learning includes: ordering a network management module to collect performance management (PM) data, configuration management (CM) data and obstacle events from at least one network communication equipment; Make a diagnostic module calculate or convert the performance management (PM) data and configuration management (CM) data of the network communication equipment collected by the network management module into corresponding key performance indicators (KPI), and make the The diagnosis module uses the key performance indicator (KPI) as an input factor and the obstacle event as an output result to build a machine learning model with a hidden layer and an exit layer as an obstacle diagnosis model. In the training process of the machine learning model In the process, the accuracy of the machine learning model with the hidden layer and the exit layer is verified by the verification data set, and the hidden layer of the machine learning model is added to the exit layer; and the diagnostic module will be used by the network The input of the new key performance indicator (KPI) calculated or converted from the new performance management (PM) data and the new configuration management (CM) data of the communication equipment is the machine learning model with the hidden layer and the exit layer. The obstacle diagnosis model so that the diagnosis module can pass The obstacle diagnosis model, which is the machine learning model with the hidden layer and the exit layer, uses the new key performance indicator (KPI) to predict or diagnose whether the network communication device may have an obstacle, so as to predict or diagnose the network When the communication equipment may be obstructed, the diagnostic module is made to adjust the equipment parameters of the network communication equipment to prevent or prevent the network communication equipment from malfunctioning. For example, the method described in item 11 of the scope of patent application, wherein the diagnostic module uses the machine learning model to establish a knowledge-based obstacle diagnosis method, and the machine learning model includes a complex recurrent neural network (RNN) layer and A recurrent neural network model of complex neurons. For example, the method described in item 11 of the scope of patent application further includes enabling the network communication device to upload the performance management (PM) data, the configuration management (CM) data and the device log (Log) to the network management module every period of time. Group, and make the network management module take over the supply, monitoring, parameter setting of the network communication equipment or the management of the key performance indicator (KPI). For example, the method described in item 11 of the scope of patent application further includes making the network communication device transmit the performance management (PM) data and the configuration management (CM) data to the network management module for storage in the network management module In a server or database, the network communication device is allowed to forward the obstacle event log related to the obstacle event to the network management module for recording in the server or database. For example, the method described in item 11 of the scope of patent application further includes the parameters of the performance management (PM) data and the configuration management (CM) data through the presentation layer state transition application program interface (Rest API) of the network management module Send to the diagnostic module so that the diagnostic module calculates or converts the parameters of the performance management (PM) data and configuration management (CM) data into the corresponding parameters through a key performance indicator (KPI) formula database. Key performance indicators (KPI). For example, the method described in item 11 of the scope of patent application further includes making the diagnostic module establish a weight evaluation algorithm related to the obstacle event, so as to make the diagnostic module use the weight evaluation during the training process of the machine learning model Algorithm to optimize or adjust the category weight of the obstacle event. For example, the method described in item 11 of the scope of the patent application also includes the use of the average, mode or zero filling of the key performance indicators (KPI) of each obstacle category when the diagnostic module calculates the key performance indicator (KPI) The null value of the key performance indicator (KPI). For example, the method described in item 11 of the scope of patent application further includes making the diagnostic module use normalization to optimize the value of the key performance indicator (KPI), or express the value of the key performance indicator (KPI) with a true or false value . For example, the method described in item 11 of the scope of patent application further includes making the diagnostic module convert the obstacle event category and the normal category of the network communication equipment into a multi-dimensional matrix through a tag code and a one-hot code. For example, the method described in item 11 of the scope of the patent application further includes making the diagnostic module use the long-term short-term memory (LSTM) model as the basis to find the correlation between the key performance indicators (KPI), and then build the Machine learning model.

Images