WO2020168676A1

WO2020168676A1 - Method for constructing network fault handling model, fault handling method and system

Info

Publication number: WO2020168676A1
Application number: PCT/CN2019/096623
Authority: WO
Inventors: 匡立伟
Original assignee: 烽火通信科技股份有限公司
Priority date: 2019-02-21
Filing date: 2019-07-19
Publication date: 2020-08-27
Also published as: CN111600734A; CN111600734B

Abstract

The present invention relates to the technical field of communications. Disclosed are a method for constructing a network fault handling model, a fault handling method and a system. The method for constructing a network fault handling model comprises: obtaining or establishing a deep neural network model of a source domain on the basis of a sample set of the source domain in a network; establishing a sample set of a target domain in the network, the sample sets of the target domain and of the source domain having an intersection, and both including quantified alarm data, fault data and configuration data; and if the coincidence rate of the sample sets of the target domain and of the source domain reaches a set threshold, constructing a network fault handling model of the target domain on the basis of the deep neural network model of the source domain. In the present invention, the network fault handling model of the target domain is obtained by means of cross-domain transfer learning on the basis of the deep neural network model of the source domain in an optical network.

Description

Method for constructing network fault processing model, fault processing method and system

Technical field

The invention relates to the field of communication technology, in particular to a method for constructing a network fault processing model, a fault processing method and a system.

Background technique

When the current performance index of the optical network equipment exceeds the limit or some potential performance is deteriorating, a series of alarm data will be generated and reported to the network management platform. When the optical network equipment fails, both alarm data and fault data will be generated and reported. At present, the operation and maintenance experts analyze the alarm data and fault data, locate the fault location, formulate fault repair strategies, and then send the corresponding configuration data to the fault location to repair through the management platform and control platform, and trigger protection switching when necessary to ensure The normal operation of the optical network.

With the ever-increasing scale of optical networks and the increasing number of optical network equipment, the number of alarm data and fault data generated by optical networks is increasing. The location and repair of network faults are becoming more and more complicated and laborious. The traditional fault handling mode faces a huge challenge. , It is difficult to meet actual needs. Especially with the rapid development of communication services and the continuous evolution and change of communication technologies, the transformation of traditional tightly coupled and rigid network architectures into loosely coupled and flexible cloud-based network architectures is the general trend. The bottom layer of the clouded network is forwarded by optical network equipment, and the middle and upper layers realize the management and control of resources and services through the control platform, management platform, and orchestration platform. The system operation and maintenance process is more complicated, and it is necessary to achieve network data integration representation and efficient data extraction operations. And operation to solve the problem of difficulty in timely recovery after a cloud network fails.

Using artificial intelligence technology to analyze and repair network faults is an effective solution to these challenges. However, optical network systems (especially cloud network architectures) include wireless networks, access networks, bearer networks, and data centers with different fault characteristics. On the one hand, the wireless network, access network, bearer network, and data center are separately established. The machine learning model leads to the problem of repeated learning; on the other hand, some target fields have incomplete sample data and it is difficult to establish an effective machine learning model.

Summary of the invention

The purpose of the embodiments of the present invention is to provide a method for constructing a network fault processing model, a fault processing method and a system, based on the deep neural network model of the source field in the optical network, and through cross-domain migration learning, the network fault processing model of the target field is obtained. .

In the first aspect, an embodiment of the present invention provides a method for constructing a network fault handling model, which includes:

Acquiring or establishing a deep neural network model of the source domain based on the sample set of the source domain in the network;

Establishing a sample set of the target field in the network, the sample set of the target field and the source field have an intersection, and both include quantified alarm data, fault data, and configuration data;

When the coincidence rate of the sample sets of the target domain and the source domain reaches the set threshold, the network fault handling model of the target domain is constructed based on the deep neural network model of the source domain.

With reference to the first aspect, in the first optional implementation manner, the deep neural network model of the source domain is used as the network fault handling model of the target domain; or,

Extracting the first input vector and the corresponding first output vector from the intersection, and retraining the deep neural network model of the source domain to obtain the network fault handling model of the target domain.

With reference to the first aspect, in a second optional implementation manner, the difference set between the sample set of the target domain and the sample set of the source domain is obtained, and the network fault handling model of the target domain is optimized based on the difference set .

In a second optional implementation manner, a second input vector and a corresponding second output vector are extracted from the difference set, and the network fault handling model in the target domain is retrained.

In a second optional implementation manner, extract a third input vector from the difference set, input the network fault processing model of the target domain, and obtain a third output vector;

After correcting the third input vector and the third output vector according to the expert evaluation feedback result, the network fault handling model in the target field is retrained.

In a second optional implementation manner, the weight coefficient of the neuron function of the network fault handling model of the target field is corrected based on the difference set to obtain an optimized network fault handling model of the target field.

With reference to the first aspect, in a third optional implementation manner, the input vector of the deep neural network model of the source domain includes the quantized alarm data and fault data, and the output vector is the quantized processing Configuration data.

In a second aspect, an embodiment of the present invention provides a network fault processing method, which includes:

Acquire alarm data and fault data of the target network, and input the network fault processing model after quantitative processing, the network fault processing model being obtained by using the network fault processing model construction method described in the first aspect;

The output vector of the network fault handling model is delivered to the relevant equipment of the target network.

In a third aspect, an embodiment of the present invention provides a construction system for a network fault handling model, which includes:

An acquisition module, which is used to acquire or establish a deep neural network model of the source domain based on the sample set of the source domain in the network;

A processing module, which is used to establish a sample set of the target field in the network, the sample set of the target field and the source field have an intersection, and both include quantified alarm data, fault data, and configuration data; and calculate the target field and the source The coincidence rate of the sample set of the field;

The construction module is used to construct a network fault handling model of the target domain based on the deep neural network model of the source domain when the coincidence rate of the sample set of the target domain and the source domain reaches a set threshold.

With reference to the third aspect, in a first optional implementation manner, the construction module is used to use the deep neural network model of the source domain as the network fault handling model of the target domain; The first input vector and the corresponding first output vector are concentratedly extracted, and the deep neural network model of the source domain is retrained to obtain the network fault handling model of the target domain.

With reference to the third aspect, in a second optional implementation manner, the processing module is further configured to obtain the difference between the sample set of the target field and the sample set of the source field;

The construction module is also used for optimizing the network fault processing model of the target domain based on the difference set.

In a second optional implementation manner, the construction module is configured to extract a second input vector and a corresponding second output vector from the difference set, and retrain the network fault handling model of the target domain.

In a second optional implementation manner, the construction module is used to extract a third input vector from the difference set, and input it into the network fault processing model of the target field to obtain a third output vector; After the feedback result corrects the third input vector and the third output vector, the network fault handling model in the target field is retrained.

In a second optional implementation manner, the construction module is used to modify the weight coefficients of the neuron function of the network fault handling model of the target field based on the difference set to obtain an optimized value of the target field Network fault handling model.

With reference to the third aspect, in a third optional implementation manner, the input vector of the deep neural network model of the source domain includes the quantized alarm data and fault data, and the output vector is the quantized processing Configuration data.

In a fourth aspect, an embodiment of the present invention provides a network fault processing system, which includes:

Input control module, which is used to obtain alarm data and fault data of the target network and perform quantitative processing;

The model processing module is used to store the network fault processing model constructed by the network fault processing model construction system described in the third aspect, and input the quantitatively processed alarm data and fault data into the network fault processing model to obtain State the output vector of the network fault handling model;

The output control module is used to deliver the output vector of the network fault handling model to the relevant equipment of the target network.

Compared with the prior art, the embodiment of the present invention uses a method for constructing a network fault handling model to obtain or establish a deep neural network model of the source field based on a sample set of the source field in the network; establish a sample set of the target field in the network, The sample sets of the target field and the source field both include quantified alarm data, fault data, and configuration data, and have an intersection; when the coincidence rate of the sample sets of the target field and the source field reaches the set threshold, the source field-based Deep neural network model to build a network fault handling model in the target field. Based on the deep neural network model of the source domain in the optical network, through cross-domain migration learning, the network fault handling model of the target domain is obtained.

Description of the drawings

In order to more clearly describe the technical solutions in the embodiments of the present invention, the following will briefly introduce the accompanying drawings used in the description of the embodiments. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

Figure 1 is a schematic diagram of a cloud network architecture;

2 is a flowchart of a method for constructing a network fault handling model according to an embodiment of the present invention;

Figure 3 is a schematic diagram of obtaining data from a database and performing vectorization and matrixization;

4 is a flowchart of a method for constructing a network fault handling model according to another embodiment of the present invention;

Figure 5 is an example of a multi-layer high-dimensional space;

Figure 6 is an example of the construction and optimization of the network fault handling model in the target field;

FIG. 7 is a schematic diagram of a construction system of a network fault handling model according to an embodiment of the present invention;

Fig. 8 is a schematic diagram of a network fault handling system according to an embodiment of the present invention.

detailed description

The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work shall fall within the protection scope of the present invention.

The network involved in the embodiment of the present invention may be a traditional optical transport network such as (Optical Transport Network, OTN), Packet Transport Network (PTN), and Packet Optical Transport Network (Packet Optical Transport Network, POTN), or It is a cloud network.

As an example, Figure 1 is a schematic diagram of a cloudized network architecture. The bottom left part of Figure 1 is a cloudized network base station, including Active Antenna Unit (AAU), Centralized Unit (CU), and Distributed Unit (Distributed Unit, DU), where CU supports non-real-time wireless high-level protocols and some core network sink functions and edge application functions, and DU supports physical layer functions and real-time functions. The lower part of Figure 1 is the cloudized network access ring, aggregation ring and core ring. The alarm data, fault data and configuration data of the network devices in these ring networks are reported to the edge data in the upper part of Figure 1 through the network management platform or the controller platform. Center, regional data center and core data center, base station and edge application alarm data, fault data and configuration data are reported to edge data center through local network. The core network functions of the 5G core network are divided into user plane (UP) functions and control plane (CP) functions. On the one hand, these data centers are responsible for the management, orchestration, and control of cloud-based networks. On the other hand, they deploy intelligent platforms for cloud-based networks, and build a cloud-based network operation and maintenance management knowledge base based on massive network data and powerful computing capabilities. , As the brain of the cloud network.

Because the massive amount of optical network alarm data, fault data, and configuration data contains a large amount of redundant, incomplete and inconsistent data, the data center first cleans the data to remove redundant, low-quality data and obtain high-quality alarm data Sets, fault data sets, and configuration data sets are stored in the database.

In the embodiment of the present invention, taking FIG. 1 as an example, the source domain may be defined as an access network, the target domain may be defined as an aggregation network, or the source domain may be defined as a core network, and the target domain may be defined as a data center network, which is not limited. Devices on different networks of clouded networks may have their own professional network management or dedicated control platforms. In other embodiments, the source domain and the target domain may also be an access network, an aggregation network, and a core network in a traditional optical network (OTN, PTN, and POTN), respectively.

The network equipment reports the alarm data and related fault data to the network management platform, and the network management platform submits it to the data center. Alarms generated by network equipment include root cause alarms and derivative alarms, and there is a correlation between root cause alarms and derivative alarms. When a network device fails, alarm data and fault data are generated and reported at the same time, and the fault needs to be repaired through the issued configuration data.

The embodiment of the present invention is based on the deep neural network model of the source domain in the network, and obtains the network fault processing model of the target domain through cross-domain migration learning. Therefore, when an alarm or failure occurs in the target field, the network fault handling model in the target field automatically generates configuration data, and distributes the equipment in the target field through the management control platform to complete the equipment recovery, switching, parameter adjustment and rerouting of the target field, etc. Operate to achieve self-healing of network failures in the target area.

The embodiment of the present invention solves the problem of repeated learning when establishing network fault handling models in different fields in the network, and the problem of incomplete sample data in some target fields, which makes it difficult to establish an effective machine learning model, and is conducive to unifying different fields in the network management.

Figure 2 shows a flowchart of a method for constructing a network fault handling model according to an embodiment of the present invention. The network includes a source field and a target field. The method for constructing a network fault handling model includes:

S110 obtains the sample set of the source domain and its deep neural network model.

S120 establishes a sample set of the target field. The sample sets of the target field and the source field have an intersection, and both include quantified alarm data, fault data, and configuration data.

S130: When the coincidence rate of the sample sets of the target domain and the source domain reaches the set threshold, construct a network fault handling model of the target domain based on the deep neural network model of the source domain.

In step S110, the deep neural network model of the source domain is created in advance based on the sample set of the source domain. Common deep neural network models include Stacked Auto-Encoder, Convolutional Neural Network (CNN), Deep Belief Network, etc.

The sample set in the source domain includes quantitatively processed alarm data, fault data, and configuration data, see the specific description in step S120.

The input and output sample data of the deep neural network model in the source field usually takes the form of vectors, that is, the sample set includes the alarm data vector group, the fault data vector group and the configuration data respectively obtained according to the alarm data, fault data and configuration data of the source field Vector set.

As an example, the input vector of the deep neural network model of the source domain includes quantized alarm data and fault data, and the output vector is quantized configuration data.

Using the artificial intelligence deep learning method, the quantified alarm data and fault data are used as input, and the quantized configuration data is used as the output. A deep neural network model is generated and trained. Through large-scale high-quality sample data training, the depth The neural network model learns the fault intelligent self-healing knowledge in the source field, and the relevant knowledge is stored in a series of neurons in the deep neural network in an abstract form. Through the deep neural network model in the source domain, the association rules between the optical network-derived alarms and the root-cause alarms are excavated, and the precise relationship between the root-cause alarms and the fault location is generated. The network configuration plan can be given according to the alarm and fault information, and the network management and The controller platform realizes the automatic repair of faults in the optical network source field.

In step S120, the alarm data, fault data and configuration data at multiple time points are obtained from the database of the target field, and the sample set of the target field is obtained after quantization processing. Among them, in order to obtain the intersection of the sample sets of the target field and the source field, the field definitions of the alarm data, fault data, and configuration data of the source field and the target field are the same, but the ordering is not required to be the same.

Based on the correlation between the alarm data, fault data and configuration data in the generation time, all the alarm data, fault data and configuration data of the target field in the set time period can be obtained from the database, or by day, week or month Periodically obtain all alarm data, fault data and configuration data of the target field from the database. The set time period or cycle includes alarm data at multiple time points, fault data at multiple time points, and configuration data at multiple time points.

Alarm data, fault data, and configuration data are not only heterogeneous data, but these data include various types of fields, and different fields have different dimensions. Therefore, the quantization process includes the vectorized representation of heterogeneous data of different dimensions, including:

In S121, each piece of alarm data, fault data or configuration data is converted into a basic vector V _b , and each element of the basic vector V _b is the value of a field in each piece of alarm data, fault data or configuration data.

For example, the sample data set acquired all alarms are constituted M _a in a trap data, wherein the alerting data generated at a time point may be one or more pieces, each alarm data field has N _a.

As an example, a piece of alarm data shown in Figure 3 includes eight fields, namely: the sequence number of the alarm data Seq.No., address Addr., line number Line, alarm type AlarmType, alarm start time BeginTime, alarm end Time EndTime, board type BoardType, and network element type NetType, where the alarm start time BeginTime and alarm end time EndTime are accurate to seconds, the address Addr. and alarm type AlarmType are character numbers, and the network element type NetType is an integer value.

The values of all fields of the alarm data shown in Fig. 3 are converted into real numbers, and thus expressed as elements of a vector. In the vectorization process of alarm data, the integer values of these fields are represented in the vector as element values. The minimum value of all alarm start time BeginTime and alarm end time EndTime can be corresponding to the value 1, and the number of seconds between other times and the minimum time can be added to the value 1, and the corresponding values of the alarm start time BeginTime and the alarm end time EndTime can be obtained respectively . For example, if the alarm start time BeginTime is 10 seconds longer than the minimum time, the alarm start time BeginTime corresponds to the value 11. The two fields are arranged in lexicographic order, and then numbered from 1, and the string is converted to a value as an element of the vector .

S122 performs dimension conversion on the basis vector V _b , and the converted vector V is the hadamard product of the basis vector V _b and the dimension expansion vector V _s , namely

The element of the dimension expansion vector V _s is the expansion or reduction multiple of the corresponding element of the basic vector V _b . For example, if the bandwidth unit M is expanded to giga G, the element of the dimension expansion vector V _s is 1024.

According to the training requirements of the machine learning model, the basic vector can be multiplied by the corresponding elements of the dimension expansion vector to generate sample data suitable for the training requirements. In the same way, the configuration data and fault data in the lower left part of Fig. 3 are also converted into corresponding vectors. The configuration data includes Num_CPUs: 4, which is the number of CPU cores. As an example, the vector group in the lower part of Fig. 3 shows two vectors. The alarm data and configuration data are converted.

For the fault data and configuration data stored in the semi-structured XML document in the optical network, the above method can also be used to construct the data basis vector and the dimension expansion vector. The number of key/value pairs in XML corresponds to the number of the vector Dimension, the value of the vector element corresponds to the Value value in the XML document.

Construct three pairs of vector groups for the target field, which are the basic vector group of alarm data and the expanded vector group of dimensions, the basic vector group of fault data and the expanded vector group of dimensions, and the basic vector group of configuration data and the expanded vector group of dimensions. including alarm data vectors obtained by the data converting M _a M _a in a trap number alarm data vectors, each data vector having N _a warning elements; vector data group includes fault fault M _f M _f obtained from the data conversion section fault data vectors, each data vector having N _f fault elements; vector set of configuration data including configuration data obtained by conversion of configuration data vector M _c M _c from the bar, each configuration data vector with N _c elements.

Further, the alarm data vector group, the fault data vector group, and the configuration data vector group can also be expressed in matrix. For example, the alarm data vector group is stored in a two-dimensional empty matrix in the form of row vectors to form an alarm matrix. for example, two-dimensional matrix of the lower right portion of FIG. 3, M _a = 7000 if there is data in a trap, trap matrix is formed 7000 rows and 8 columns. Similarly, fault matrix and configuration matrix can be constructed.

S123 finds the intersection of the sample set of the target field and the source field.

Specifically, according to the vector elements in the alarm data vector group, the fault data vector group and the configuration data vector group of the target domain and the source domain, the intersection of the sample sets of the target domain and the source domain is obtained.

In step S130, when the coincidence rate of the sample set of the source field and the sample set of the target field reaches the set threshold, the network fault handling model of the target field can be constructed in different implementation manners, for example, one of the following implementation manners can be adopted One:

Implementation mode 1: The deep neural network model in the source domain is used as the network fault processing model in the target domain.

Embodiment 2: Extract the first input vector and the corresponding first output vector from the intersection, retrain the deep neural network model of the source domain, and obtain the network fault handling model of the target domain. The network fault handling model of the target domain is the same as that of the source domain. The deep neural network model is similar to the deep neural network model.

After the source domain's deep neural network model obtains the source domain fault self-healing knowledge base, the intersection of the sample data of the source domain and the target domain is obtained. Because the intersection fault self-healing knowledge is already included in the source domain fault self-healing knowledge base Therefore, based on the intersection of the sample data of the source domain and the target domain, the knowledge base for self-healing faults in the source domain is migrated to the target domain to realize cross-domain migration learning. In the process of transfer learning, if the intersection of the sample data in the source field and the target field is relatively large (that is, the coincidence rate is higher), the transfer learning effect will be better.

In practical applications, the size of the threshold can be adjusted according to specific scenarios. The threshold is a percentage value. For example, the overlap rate of the data intersection of the source domain and the target domain is between 0% and 100%. For example, a coincidence rate of 60% means that the sample data of the source field and the target field are 60% the same, and 40% are different.

If the threshold is small, the migration process of the knowledge base of the source domain fault self-healing is faster, and the subsequent correction and optimization process of the weight parameter is longer. Conversely, if the threshold is larger, the migration process of the knowledge base for self-healing failures in the source domain is slower, but the subsequent correction and optimization process of the weight parameters is shorter.

If the coincidence rate is lower than the set threshold, you need to add new data to the sample set of the target field, or you can select a batch of data samples again to supplement the intersection data of the source field and the target field, until the coincidence rate exceeds the set Threshold.

In this embodiment, steps S110 and S120 are executed in sequence, and in another embodiment of the present invention, steps S110 and S120 can also be executed in other ways, for example, obtaining alarm data and fault data in the source domain and target domain respectively After quantitative processing, the sample sets of the source field and the target field are established respectively, and then the deep neural network model of the source field is constructed.

Fig. 4 is a flowchart of a method for constructing a network fault handling model according to another embodiment of the present invention. The method for constructing a network fault handling model includes:

S200 data collection and preprocessing. It specifically includes:

S201 data collection and preprocessing in the source field.

S202 Data collection and preprocessing in the target field.

The data collection and preprocessing process of the source field and target field are basically the same.

The alarm data, fault data and configuration data of the optical network are uploaded to the three types of data centers by the network management platform or the controller platform. Because the alarm data, fault data, and configuration data of the massive optical network contain a large amount of redundant, incomplete, and inconsistent data, the three types of data centers will first clean the data, remove the redundant, low-quality data, and get the highest Quality alarms, faults, and configuration data sets are stored in the source domain database and target domain database respectively.

S210 constructs a deep neural network model of the source domain.

The method of constructing deep neural network models in the source domain is not limited. For example, common deep neural network models include Stacked Auto-Encoder, Convolutional Neural Network (CNN), and Deep Belief Network ( Deep Belief Network) and so on.

S220 When the coincidence rate of the sample set of the source domain and the sample set of the target domain reaches the set threshold, construct a network fault handling model of the target domain based on the deep neural network model of the source domain.

Step S220 specifically includes:

S221 Unified representation of the sample set of the source field and the sample set of the target field.

Specifically, a multi-layer high-dimensional space is constructed to realize the unified representation of alarm data, fault data, and configuration data in the source and target fields.

The vectorization and matrix representation methods of heterogeneous data of different dimensions are successively used to convert the alarm data, fault data and configuration data of the source and target fields into one-dimensional vectors, and then respectively express them into corresponding two-dimensional matrices. It specifically includes: the construction process of one-dimensional vector and the construction process of two-dimensional matrix.

Specifically, a two-dimensional alarm matrix, a fault matrix, and a configuration matrix are constructed according to the alarm data, fault data, and configuration data of the source field, and a two-dimensional alarm matrix, fault, and configuration data are constructed according to the alarm data, fault data, and configuration data of the target field. The construction methods of the matrix and the configuration matrix, the one-dimensional vector and the two-dimensional matrix are similar to the foregoing embodiments, and will not be repeated here.

As an example, if the alarm data, fault data, and configuration data of the source field and the target field are expressed in a matrix, the number of rows and columns of the two-dimensional matrix may be different, as shown in Table 1:

Table 1 Examples of the number of rows and columns of the two-dimensional matrix of the source and target fields

矩阵类型Matrix type	告警矩阵的行列数Number of rows and columns of the alarm matrix	故障矩阵的行列数Number of rows and columns of the fault matrix	配置矩阵的行列数Configure the number of rows and columns of the matrix
源领域Source field	5000×125000×12	7000×187000×18	3000×323000×32
目标领域Target field	3000×83000×8	5000×125000×12	2000×352000×35

Obtain the maximum number of rows and the maximum number of columns of all alarm matrix, fault matrix, and configuration matrix, and use the maximum number of rows and the maximum number of columns as the number of rows and columns of each layer of the two-dimensional matrix in the multi-layer high-dimensional space. Taking Table 1 as an example, the number of rows and columns of each layer of a two-dimensional matrix in a multi-layer high-dimensional space are 7000 and 35, respectively. Among them, the number of rows 7000 means that the largest number of rows in the six matrices is the number of rows of the source field fault matrix, and the number of columns 35 means that the largest number of columns in the six matrices is the number of columns of the target field configuration matrix.

After obtaining the maximum number of rows of 7000 and the maximum number of columns of 35, a 6-layer high-dimensional space representation model is constructed based on the six matrices in Table 1 above, and 6 empty matrices with 7000 rows and 35 columns are generated, and these 6 The data in the matrix is copied to the newly generated empty matrix, and the matrix elements without stored data are filled with zero elements.

Specifically, the multi-layer high-dimensional space constructed for the source field and the target field is shown in Figure 5. The six-layer multi-layer high-dimensional space D=R(K ₁ , K ₂ , K ₃ ), the first to the third The layers are the alarm data layer, fault data layer and configuration data layer of the source field, corresponding to the alarm matrix, fault matrix and configuration matrix of the source field respectively. The fourth to sixth layers are the alarm data layer, fault data layer and The configuration data layer corresponds to the alarm matrix, fault matrix and configuration matrix of the target field. Among them, the three-layer high-dimensional space in the source field can also be expressed as D _s = R(I ₁ , I ₂ , I ₃ ), and the three-layer high-dimensional space in the target field can also be expressed as D _t =R(J ₁ , J ₂ , J ₃ ).

Using the method in the above embodiment, it is also possible to construct a multi-layer high-dimensional space for multiple fields, such as an access network, a convergence network, a core network, and a data center network, which is not limited.

Through the embodiment of the present invention, the vectorization and matrix representation methods for heterogeneous data of different dimensions can convert structured and semi-structured optical network data of different dimensions into vectors and matrices, because there are a large number of zero elements Filling, multi-layer high-dimensional space is a sparse matrix. In the process of saving, the classic sparse matrix storage method can be used to save data to save storage space. At the same time, constructing a multi-layer high-dimensional space not only realizes the unified representation of sample data in the source domain and the target domain, but also realizes the intercommunication and sharing of cross-domain sample data from different vendors, removing information island barriers for subsequent machine learning.

The sample set of the source domain can be a three-layer high-dimensional space D _s = R(I ₁ , I ₂ , I ₃ ) in the source domain, or a sub-sub of D _s = R(I ₁ , I ₂ , I ₃ ) space. Similarly, the sample set of the target field can be the three-layer high-dimensional space D _t =R(J ₁ , J ₂ , J ₃ ) of the target field, or D _t =R(J ₁ , J ₂ , J ₃ ) Of a subspace.

The subspace includes at least one submatrix of an alarm data layer, a fault data layer, and a configuration data layer. The sub-matrix can be a sub-matrix in one layer of the multi-layer high-dimensional space; the sub-matrix can also be two or more layers of the multi-layer high-dimensional space, where each layer of the sub-matrix is a layer of the multi-layer high-dimensional space A sub-matrix of.

As an example, the matrix S and the matrix T in FIG. 6 respectively represent the sample set of the source field and the sample set of the target field in the transfer learning process, and these two matrices are both matrices with 3 rows and 3 columns.

S222 finds the intersection of the sample set of the target domain and the source domain.

Specifically, the intersection is also a subspace of a multi-layer high-dimensional space, and the subspace includes at least one submatrix of an alarm data layer, a fault data layer, and a configuration data layer.

The sub-matrix can be a sub-matrix in one layer of the multi-layer high-dimensional space; the sub-matrix can also be two or more layers of the multi-layer high-dimensional space, where each layer of the sub-matrix is a layer of the multi-layer high-dimensional space A sub-matrix of.

Take the matrix S and the matrix T in Fig. 6 as examples. The matrix S and the matrix T respectively represent the sample set of the source field and the sample set of the target field in the transfer learning process. Both matrices are three-row and three-column matrices. Find the data intersection of the matrix S and T, get the intersection matrix I, get a matrix with 2 rows and 3 columns, that is, the first row vector of the matrix S in Figure 6 is equal to the first row vector of the matrix T, that is, S ₁₁ = T ₁₁ , S ₁₂ = T ₁₂ , S ₁₃ = T ₁₃ , and the second row vector of the matrix S is equal to the second row vector of the matrix T, that is, S ₂₁ = T ₂₁ , S ₂₂ = T ₂₂ , S ₂₃ = T ₂₃ , it means that the first two row vectors of the matrix S and the matrix T are equal, and the two equal row vectors are taken out to obtain the intersection I.

S223 builds a deep neural network model of the target field.

Step S223 is basically the same as step S130 in the foregoing embodiment.

Specifically, the overlap between the sample set of the source field and the sample set of the target field is the intersection. The deep neural network model of the source field can be directly used as the network fault processing model of the target field, or the first input vector and the corresponding can be extracted from the intersection. Retrain the deep neural network model of the source domain to obtain the network fault handling model of the target domain, thereby migrating the fault handling knowledge base of the source domain to the fault handling knowledge base of the target domain.

Still take the example shown in Figure 6 for explanation. In Figure 6, the matrices S and T are 9-element matrices, and the intersection matrix I is a 6-element matrix. If the threshold is set to 60%, the proportion of the intersection data exceeds the set Set a threshold of 60%, you can directly use the deep neural network model in the source field as the network fault handling model in the target field, or extract the first input vector and the corresponding first output vector from the intersection to retrain the deep neural network in the source field Network model to obtain the network fault handling model of the target field.

As shown in FIG. 4, the method for constructing a network fault handling model further includes: S300 obtains the difference set between the sample set of the target domain and the sample set of the source domain, and optimizes the network fault handling model of the target domain based on the difference set.

In one embodiment, the second input vector and the corresponding second output vector can be extracted from the difference set, and the network fault handling model of the target domain can be retrained.

In another embodiment, the third input vector can also be extracted from the difference set, and the third output vector can be obtained by inputting the network fault processing model of the target field; the third input vector and the third output vector are corrected according to the expert evaluation feedback result , Retrain the network fault handling model in the target domain.

Among them, retraining the network fault handling model of the target field includes: correcting the weight coefficient of the neuron function of the network fault handling model of the target field based on the difference set to obtain an optimized network fault handling model of the target field.

The intersection matrix I in Figure 6 is used to directly generate the weight parameters of the fitting function in the deep neural network model of the target field. The lower part of Figure 6 is the difference matrix D between the source field and the target field. The difference matrix D is a 2 row 3 The column matrix and the difference matrix D are used to optimize the weight parameters of the fitting function in the deep neural network model of the target field.

Take FIG. 6 as an example to illustrate the process of modifying the weight coefficient w ₂₂ of the neuron function f ₂₂ through the difference set data x ₂₂ and y ₂₂ . This example selects the alarm time, alarm category, and fault category, and constructs the input vector after quantified representation. The configuration data is the configuration plan quantified representation data to construct the output vector. In this embodiment, the quantification of the configuration plan indicates that the value 1 indicates that the first configuration plan is adopted, and the number set to -1 indicates that the second configuration plan is adopted. Table 1 Number 1 row of data corresponds to input vector x=(2,5,7), output y=1, indicating that the quantified values of alarm time, alarm category and fault category are 2, 5 and 7, respectively, and the configuration plan is quantified The expression value is 1. This input vector and output vector are fitted by the deep learning neural network model neuron function f ₂₂ through the formula y=f(x)=sgn(wx ^T ).

As an example, the weight coefficient w of the neuron function of the deep neural network model of the target field is obtained through a large amount of sample data similar to the serial number 1 and the serial number 2 in Table 2. The intersection data in Table 2 represents the sample data in the intersection, and the difference data represents the sample data in the difference. The weight coefficient w = (1,0,1) corresponding to sequence number 3 in Table 2 satisfies sgn[(1,0,1)*(2,5,7) ^T ]=sgn(9)=1, sgn[( 1,0,1)*(3,2,8) ^T ]=sgn(11)=1. In Table 2, the sequence numbers 4 and 5 correspond to the difference set data. The input vectors (5,7,3) and (8,3,7) are constructed based on the difference set data, and the output y value is -1. Inject the input vector and output vector constructed by the difference data into the deep neural network model of the target field, readjust the weight of the neuron function of the deep neural network model of the target field, and obtain the corrected neuron function corresponding to the number 6 in Table 2. The weight coefficient w=(1,-1,-1), this weight coefficient satisfies sgn[(1,-1,-1)*(5,7,3) ^T ]=sgn(-5)=-1, sgn[[(1,-1,-1)*(8,3,7) ^T ]=sgn(-2)=-1.

Table 2 is an example of correcting neuron weight coefficient based on difference set

Based on the difference set data, the weight parameters of the neuron function of the deep neural network model in the optimized target field are continuously revised, and finally the optimized deep neural network model in the target field is obtained, which realizes automatic recovery and automatic elimination of optical network faults. The weight parameters of the corrected and optimized neuron function are stored in each neuron node of the deep neural network model of the target field, as shown in the right part of Figure 6.

In the foregoing description, step S300 is based on steps S200 to S220 of the foregoing embodiment, and further optimizes the network fault processing model of the target field based on the difference set.

Similar to the foregoing process, step S300 can also be based on steps S110 to S130 of the foregoing embodiment to further optimize the network fault processing model of the target field based on the difference set, which will not be repeated here.

The embodiment of the present invention also provides a network fault processing method. Based on the foregoing embodiments, the network fault processing method includes:

S410 acquires alarm data and fault data of the target network, and inputs the network fault processing model after quantitative processing. The network fault processing model is obtained by using the aforementioned method for constructing the network fault processing model.

The output vector of the S420 network fault handling model is delivered to the relevant equipment of the target network.

The embodiment of the present invention is based on the deep neural network model of the source domain in the optical network, and obtains the network fault handling model of the target domain through cross-domain migration learning. When an alarm or fault occurs in the target domain, the network fault handling model automatically generates configuration data, and Through the management and control platform, the equipment in the target field is issued to complete operations such as equipment restoration, switching, parameter adjustment and rerouting in the target field, so as to realize the self-healing of network failures in the target field.

As shown in FIG. 7, the embodiment of the present invention also provides a construction system of a network fault handling model, which is used to implement the construction method of the network fault handling model of the foregoing embodiment. The construction system of the network fault handling model includes an acquisition module 100 and a processing module. 200 and building block 300.

The acquiring module 100 is used to acquire or establish a deep neural network model of the source domain based on the sample set 102 of the source domain in the network.

In a possible implementation, the acquisition module 100 includes the acquired source domain sample set 102 and a source domain deep neural network model established based on the source domain sample set 102.

In another possible implementation, the acquisition module 100 includes a source domain data collection unit 101, a source domain sample set 102, and a source domain deep neural network model construction unit 103.

The source domain data collection unit 101 collects sample data and saves it in the source domain sample set 102. The deep neural network model of the source domain is constructed by the source domain deep neural network model construction unit 103 based on the source domain sample set 102.

The processing module 200 is used to establish a sample set 202 of the target domain in the network. The sample sets of the target domain and the source domain have an intersection 203, and both include quantitatively processed alarm data, fault data, and configuration data. Wherein, the target field data collection unit 201 in the processing module 200 collects sample data and saves it in the target field sample set 202. The processing module 200 is also used to calculate the coincidence rate of the sample sets of the target field and the source field.

The construction module 300 is used for constructing a network fault processing model of the target domain based on the deep neural network model of the source domain when the coincidence rate of the sample sets of the target domain and the source domain reaches a set threshold.

Further, the construction module 300 is used to use the deep neural network model of the source domain as the network fault processing model of the target domain; it is also used to extract the first input vector and the corresponding first output vector from the intersection, and retrain the depth of the source domain. Neural network model to obtain the network fault handling model of the target field.

Further, the processing module 200 is also used to obtain the difference set 204 between the sample set in the target field and the sample set in the source field. The construction module 300 is used to optimize the network fault handling model of the target field based on the difference set 204.

Further, the construction module 300 is also used for extracting the second input vector and the corresponding second output vector from the difference set 204, and retraining the network fault processing model of the target domain.

Further, the construction module 300 is also used to extract a third input vector from the difference set 204, and the deep neural network model of the input source domain is used to obtain the third output vector; it is also used to compare the third input vector and the third output vector according to the expert evaluation feedback result. After the output vector is corrected, the network fault handling model in the target field is retrained.

Specifically, the construction module 300 is used to modify the weight coefficient of the neuron function of the network fault processing model of the target field based on the difference set 204 to obtain an optimized network fault processing model of the target field.

Specifically, the input vector of the deep neural network model of the source domain includes quantized alarm data and fault data, and the output vector is quantized configuration data.

Referring to FIG. 8, an embodiment of the present invention provides a network fault processing system, which includes an input control module 400, a model processing module 500, and an output control module 600.

The input control module 400 is used to obtain alarm data and fault data of the target network, and perform quantitative processing.

The model processing module 500 is used to store the network fault processing model constructed by the aforementioned network fault processing model construction system, and input the quantitatively processed alarm data and fault data into the network fault processing model to obtain the output vector of the network fault processing model.

The output control module 600 is used to deliver the output vector of the network fault handling model to related devices of the target network.

In the above-mentioned embodiments, it may be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented by software, it can be implemented in the form of a computer program product in whole or in part. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on the computer, the processes or functions according to the embodiments of the present application are generated in whole or in part. The computer can be a general-purpose computer, a dedicated computer, a computer network, or other programmable devices. Computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, computer instructions can be transmitted from a website, computer, server, or data center through a cable (such as Coaxial cable, optical fiber, Digital Subscriber Line (DSL)) or wireless (such as infrared, wireless, microwave, etc.) transmission to another website, computer, server or data center. The computer-readable storage medium may be any available medium that can be read by a computer or a data storage device such as a server or data center integrated with one or more available media. Available media can be magnetic media (for example, floppy disks, hard drives, tapes), optical media (for example, Digital Video Disc (DVD)) or semiconductor media (for example, Solid State Disk (SSD)), etc. .

The present invention is not limited to the above-mentioned embodiments. For those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and modifications can be made, and these improvements and modifications are also regarded as the protection of the present invention. Within range. The content not described in detail in this specification belongs to the prior art known to those skilled in the art.

Claims

A method for constructing a network fault handling model, which is characterized in that it includes:

Acquiring or establishing a deep neural network model of the source domain based on the sample set of the source domain in the network;

Establishing a sample set of the target field in the network, the sample set of the target field and the source field have an intersection, and both include quantified alarm data, fault data, and configuration data;

When the coincidence rate of the sample sets of the target domain and the source domain reaches the set threshold, the network fault handling model of the target domain is constructed based on the deep neural network model of the source domain.
The method for constructing a network fault handling model according to claim 1, wherein the deep neural network model of the source domain is used as the network fault handling model of the target domain; or,

Extracting the first input vector and the corresponding first output vector from the intersection, and retraining the deep neural network model of the source domain to obtain the network fault handling model of the target domain.
The method for constructing a network fault handling model according to claim 1, wherein the method further comprises: obtaining the difference between the sample set of the target domain and the sample set of the source domain, and optimizing based on the difference set The network fault handling model of the target field.
3. The method for constructing a network fault processing model according to claim 3, wherein the second input vector and the corresponding second output vector are extracted from the difference set, and the network fault processing model of the target domain is retrained.
3. The method for constructing a network fault processing model according to claim 3, characterized in that: extracting a third input vector from the difference set, inputting the network fault processing model of the target domain, and obtaining a third output vector;

After correcting the third input vector and the third output vector according to the expert evaluation feedback result, the network fault handling model in the target field is retrained.
The method for constructing a network fault processing model according to claim 3, wherein the weight coefficient of the neuron function of the network fault processing model of the target field is corrected based on the difference set to obtain the optimized target The network fault handling model of the domain.
The method for constructing a network fault processing model according to claim 1, wherein the input vector of the deep neural network model of the source field includes the quantized alarm data and fault data, and the output vector is the Quantified configuration data.
A network fault processing method, characterized in that it includes:

Obtain the alarm data and fault data of the target network, and input the network fault processing model after quantitative processing, the network fault processing model obtained by using the method for constructing a network fault processing model according to any one of claims 1 to 7;

The output vector of the network fault handling model is delivered to the relevant equipment of the target network.
A construction system for a network fault handling model, which is characterized in that it includes:

An acquisition module, which is used to acquire or establish a deep neural network model of the source domain based on the sample set of the source domain in the network;

A processing module, which is used to establish a sample set of the target field in the network, the sample set of the target field and the source field have an intersection, and both include quantified alarm data, fault data, and configuration data; and calculate the target field and the source The coincidence rate of the sample set of the field;

The construction module is used to construct a network fault handling model of the target domain based on the deep neural network model of the source domain when the coincidence rate of the sample sets of the target domain and the source domain reaches a set threshold.
The network fault processing model construction system according to claim 9, wherein the construction module is used to use the deep neural network model of the source domain as the network fault handling model of the target domain; The first input vector and the corresponding first output vector are extracted from the intersection, and the deep neural network model of the source domain is retrained to obtain the network fault handling model of the target domain.
10. The network fault processing model construction system according to claim 9, wherein the processing module is further used to obtain the difference between the sample set of the target domain and the sample set of the source domain;

The construction module is also used for optimizing the network fault processing model of the target domain based on the difference set.
The network fault handling model construction system of claim 11, wherein the construction module is used to extract a second input vector and a corresponding second output vector from the difference set, and retrain the target domain Network fault handling model.
The network fault processing model construction system of claim 11, wherein the building module is used to extract a third input vector from the difference set, and input the network fault processing model of the target domain to obtain a third output Vector; also used to modify the third input vector and the third output vector according to the expert evaluation feedback result, and then retrain the network fault handling model in the target field.
The network fault processing model construction system according to claim 11, wherein the building module is used to modify the weight coefficients of the neuron function of the network fault processing model of the target field based on the difference set, An optimized network fault handling model of the target field is obtained.
The network fault processing model construction system of claim 9, wherein the input vector of the deep neural network model of the source field includes the quantized alarm data and fault data, and the output vector is the Quantified configuration data.
A network fault processing system, characterized in that it includes:

Input control module, which is used to obtain alarm data and fault data of the target network and perform quantitative processing;

A model processing module, which is used to store the network fault processing model constructed by the network fault processing model construction system of any one of claims 9 to 15, and input the quantitatively processed alarm data and fault data into the network fault processing Model to obtain the output vector of the network fault handling model;

The output control module is used to deliver the output vector of the network fault handling model to the relevant equipment of the target network.