EP4183113A1

EP4183113A1 - Network management device and method for mapping network devices from various telecom vendors

Info

Publication number: EP4183113A1
Application number: EP20754206.9A
Authority: EP
Inventors: Milan REDZIC
Original assignee: Huawei Technologies Co Ltd
Current assignee: Huawei Technologies Co Ltd
Priority date: 2020-08-06
Filing date: 2020-08-06
Publication date: 2023-05-24
Also published as: WO2022028696A1

Abstract

The present disclosure relates generally to the field of software-defined networking, in particular to a solution for mapping device configurations of network devices from different telecom vendors in a telecommunication network. To this end, the disclosure proposes a network management device configured to obtain vendor-specific data, input the vendor-specific data into a siamese neural network model in order to compare device configurations, obtain similarity scores, and determine matched device configurations based on the similarity scores. In particular, vendor-specific data are represented into a form of vector or matrix by using word embeddings that are comprised in the siamese neural network model and trained on telecommunication protocols and/or specifications, such as Request for Comments (RFC) documents.

Description

NETWORK MANAGEMENT DEVICE AND METHOD FOR MAPPING NETWORK DEVICES FROM VARIOUS TELECOM VENDORS

TECHNICAL FIELD

The present disclosure relates generally to the field of software-defined networking. More particularly, the present disclosure relates to a network management device and a corresponding method for mapping network device configurations from different telecom vendors in a telecommunication network. In particular, the network management device uses a Siamese neural network model to obtain matched device configurations automatically.

BACKGROUND

Telecommunication (telecom) industries are rapidly moving towards service-oriented approaches with respect to network management. Telecom service providers are transitioning from managing equipment towards managing various aspects of services actively. Configurations and effected equipment for managing new services are among the largest costdrivers in telecom networks. Delivering valued-added services, such as Multiprotocol Label Switching, Virtual Private Networks, Metro Ethernet, and Internet Protocol television, are critical to profitability and growth for telecom service providers. Time-to-market is a critical factor, whereas any delay in configuring network devices may directly affect service deployment and may have a big impact on revenue.

Network devices from different telecom vendors that are on a same protocol stack of a telecom network often share similarities in functionality but are also unique regarding device configurations. For example, internal data models, which represent how a network device internally stores its data for performing its purpose in the telecom network, may be unique, but the internal data models tend to follow a particular style and convention for a particular equipment family.

SUMMARY

Traditionally, configurations, such as parameters, protocols, commands, internal data models and/or application programming interfaces (APIs), of each network device in a telecom network are manually fine-tuned by professional telecom engineers. In particular, for deploying a new service, a set of device configurations needs to be applied to a number of network devices, thus allowing the new service to be established between a number of network points that are connected and supported by the network devices. The device configurations typically comprise an abundant of protocol configurations, security and authorization configurations, policies, and/or user permissions etc. However, those device configurations, while similar at a service level across the network devices, have to be translated/ported individually before they can be applied in a particular network device. The telecom engineers may need to map the device configurations one by one, according to each corresponding telecom vendor of each network device. Therefore, it requires the telecom engineers to be familiar with device configurations such as application program interfaces (APIs) and internal data models of network devices from different telecom vendors.

However, due to the complexity of a modern telecom network, in which innumerous network devices from different telecom vendors or made by different manufacturers are interrelated, massive man-hours may be needed for performing device configuration mapping, in order to deploy the new service.

Furthermore, since different telecom vendors or manufacturers employ different strategies to develop their own products, device configurations are not universal across the network devices. It is thus even more difficult for the telecom engineers to ensure that each single configuration is correct, and thereby result in a diagnosing and debugging phase before the new service can be officially put into commercial operations.

In view of the above-mentioned problems, embodiments of the present disclosure aim to provide an improved scheme for performing device configuration mapping. An objective is, in particular, to enable an automatic mapping of device configurations of network devices from different telecom vendors.

The objective is achieved by the embodiments of the present disclosure as described in the enclosed independent claims. Advantageous implementations of the embodiments of the present disclosure are further defined in the dependent claims.

A first aspect of the present disclosure provides a network management device for performing device configuration mapping between at least two network devices in a telecom network. The at least two network devices are from different telecom vendors and operate on a same protocol stack of the telecom network.

The network management device is configured to obtain first vendor-specific data from a first network device of the network devices and second vendor-specific data from a second network device of the network devices.

Then the network management device is configured to input the first and the second vendor specific data to a Siamese neural network model that comprises a set of word embeddings. The set of word embeddings is trained on one or more text corpora from one or more telecom standards and/or protocols.

By using the set of word embeddings trained on the telecom standards and/or protocols, the vendor-specific data may be translated and analyzed in a machine-learnable format. Semantic features, relationships and/or similarities of the vendor-specific data may be analyzed by further layers in the Siamese neural network model.

Then the network management device is configured to obtain a set of similarity scores indicating semantic similarities between the first vendor-specific data and the second vendorspecific data based on an output of the Siamese neural network model.

Based on the similarity scores, the network management device is configured to obtain at least one matched pair of device configurations comprising a first device configuration of the first network device and a second device configuration of the second network device.

Finally, the network management device is configured to configure the first network device according to the first device configuration and the second network device according to the second device configuration.

By using the Siamese neural network model, the network management device may simplify a process of deploying new services in the telecom network, and greatly save time taken to configure or port device configurations for the new service. Notably, a protocol stack may be an implementation of a network protocol suite or protocol family. Thus, the at least two devices operating on the same protocol stack may share the same functionality and/or feature. Moreover, a Siamese neural network, also known as a twin neural network, may be an artificial neural network that uses same weights while working in tandem on two different input vectors to compute comparable output vector. The Siamese neural network model in this implementation may be extended by comprising the set of word embeddings. Each of the word embeddings may be a representation of a vocabulary where words or phrases are mapped to vectors. Moreover, the word embeddings may be trained on the text corpora from the telecom standards and/or protocols, in which each of the text corpora, i.e., a text corpus or a corpus, may be a large and structured set of texts used to do statistical analysis, check occurrences or validate linguistic rules within a specific language territory.

Optionally, the network management may obtain the at least one matched pair of device configurations based on the highest similarity score among the set of similarity scores, or based on similarity scores that are above a per-determined threshold.

In an implementation form of the first aspect, each of the first and the second vendor-specific data may comprise a set of device configurations, in which the set of device configurations comprises one or more device models, and/or one or more device commands, and/or one or more Extensible Markup Language path (XML Path, XPath) expressions of a corresponding network device.

In another implementation form of the first aspect, the network management device may be configured to perform clustering on the device configurations of the first and the second vendorspecific data based on the similarity scores to obtain multiple clusters of device configurations.

Optionally, the network management device may be configured to analyse the obtained multiple clusters of device configurations to obtain a trusted range of the similarity scores. Then the network management device may eliminate one or more untrusted similarity scores that are outside the trusted range from the similarity scores to obtain residual similarity scores. Then the network management device may eliminate device configurations associated with the one or more untrusted similarity scores. Optionally, the network management device may perform re-ranking on the residual similarity scores.

This is beneficial, since the final similarity scores may be clearer, thereby resulting in an improved accuracy.

In another implementation form of the first aspect, the network management device may be configured to obtain user feedback on the at least one matched pair of device configurations, and store the obtained user feedback.

Notably, both positive and negative user feedback may be stored.

Optionally, the network management device may be configured to perform one or more retrainings of the Siamese neural network model based on the stored user feedback to obtain one or more re-trained Siamese neural network models.

Optionally, the network management device may be configured to create one or more further Siamese neural network models based on the stored user feedback.

Optionally, the network management device may be configured to compare a performance of the one or more re-trained Siamese neural network models and/or of the one or more further Siamese neural network models. Then the network management device may select an optimal Siamese neural network model from the one or more re-trained Siamese neural network models and/or the one or more further Siamese neural network models based on the compared performance. Then the network management device may apply the selected optimal Siamese neural network model for performing further device configuration mapping.

By obtaining and restoring the user feedback, the Siamese neural network model may be retrained and enhanced, thereby resulting in an improved system performance.

In another implementation form of the first aspect, the network management device may be configured to process the first and the second vendor specific data into a form of source-target pairs and/or ground-truth pairs. This is beneficial, since one or more training sets may be built based on the source-target pairs and/or ground-truth pairs, optionally along with the user feedback. The training sets may be classified according to different device types or different protocols, which results in a portability of the Siamese neural network model.

Optionally, the network management device may be configured to, encode the first and the second vendor specific data into matrix representations by using the set of word embeddings. Then the network management device may calculate the set of similarity scores based on a similarity function as: similarity(h_a, h_b~) = exp(— \\h_a — h_b ||_x) G [0,1] in which h_a is a first matrix representation of the first vendor-specific data, hb is a second matrix representation of the second vendor-specific data, and \\h_a — h_b || s a distance function.

Optionally, the distance function may be a Euclidean distance function with a norm of 1. Alternatively, the distance function may be a Manhattan distance function with a norm of 1.

In another implementation form of the first aspect, the Siamese neural network model may comprises a pair of Long Short-Term Memory (LSTM) network models. The network management device may extract features of the first and the second vendor specific data based on the set of word embeddings by using the LSTM network models, and obtain the set of similarity scores based on the extracted features.

Notably, the extracted features may be seen as the matrix representations of the vendor-specific data.

Optionally, each of the LSTM network models may be adapted to output each of the extracted features in a form of a feature vector.

In another implementation form of the first aspect, wherein the Siamese neural network may comprise a last layer, and the last layer may be adapted to output a feature map corresponding to the first and the second inputted vendor specific data. In another implementation form of the first aspect, the network management device may be configured to generate the set of word embeddings based on a FastText algorithm.

A second aspect of the present disclosure provides a method for performing device configuration mapping between at least two network devices in a telecom network by a network management device. The at least two network devices are from different telecom vendors and operate on a same protocol stack of the telecom network. The method comprises the following steps: obtaining first vendor-specific data from a first network device of the at least two network devices and second vendor-specific data from a second network device of the at least two network devices; inputting the first and the second vendor specific data to a Siamese neural network model comprising a set of word embeddings trained on one or more text corpora from one or more telecommunication standards and/or protocols; obtaining a set of similarity scores indicating semantic similarities between the first vendor-specific data and the second vendor-specific data based on an output of the Siamese neural network model; obtaining at least one matched pair of device configurations comprising a first device configuration of the first network device and a second device configuration of the second network device based on the similarity scores; and configuring the first network device according to the first device configuration and the second network device according to the second device configuration.

In an implementation form of the second aspect, each of the first and the second vendor-specific data may comprise a set of device configurations, in which the set of device configurations comprises one or more device models, and/or one or more device commands, and/or one or more XML Path, XPath, expressions of a corresponding network device.

In another implementation form of the second aspect, the method may further comprise performing clustering on the device configurations of the first and the second vendor-specific data based on the similarity scores to obtain multiple clusters of device configurations.

Then the method may further comprise: analysing the obtained multiple clusters of device configurations to obtain a trusted range of the similarity scores; eliminating one or more untrusted similarity scores from the similarity scores to obtain residual similarity scores, in which the one or more untrusted similarity scorers are outside the trusted range; and eliminating device configurations associated with the one or more untrusted similarity scores from the obtained multiple clusters of device configurations.

Optionally, the method may further comprise performing re-ranking on the residual similarity scores.

In another implementation form of the second aspect, the method may further comprise obtaining user feedback on the at least one matched pair of device configurations, and storing the obtained user feedback.

Optionally, the method may further comprise performing one or more re-trainings of the Siamese neural network model based on the stored user feedback to obtain one or more retrained Siamese neural network models.

Optionally, the method may further comprise creating one or more further Siamese neural network models based on the stored user feedback.

Optionally, the method may further comprise: comparing performance of the one or more retrained Siamese neural network models and/or the one or more further Siamese neural network models; selecting an optimal Siamese neural network model from the one or more re-trained Siamese neural network models and/or the one or more further Siamese neural network models based on the compared performance; and applying the selected optimal Siamese neural network model for performing device configuration mapping.

In another implementation form of the second aspect, prior to the step of obtaining the set of similarity scores indicating semantic similarities between the first vendor-specific data and the second vendor-specific data, the method may further comprise processing the first and the second vendor specific data into a form of source-target pairs and/or ground-truth pairs.

Optionally, the method may further comprise encoding the first and the second vendor specific data into matrix representations by using the set of word embeddings and calculating the set of similarity scores based on a similarity function as: similarity(Ji_a, h_b~) = exp(— \\h_a — h_b ||_x) G [0,1] in which h_a is a first matrix representation of the first vendor-specific data, hb is a second matrix representation of the second vendor-specific data, and \\h_a — h_b || s a distance function.

In another implementation form of the second aspect, the Siamese neural network model may comprises a pair of LSTM network models, and the method may further comprise extracting features of the first and the second vendor specific data based on the set of word embeddings by using the LSTM network models; and obtaining the set of similarity scores based on the extracted features.

Optionally, the method may further comprise outputting the extracted features in a form of a feature vector by each of the LSTM network models.

In another implementation form of the second aspect, wherein the Siamese neural network may comprise a last layer, and the method may further comprise outputting a feature map corresponding to the first and the second inputted vendor specific data by the last layer.

In another implementation form of the second aspect, the method may further comprise generating the set of word embeddings based on a FastText algorithm.

A third aspect of the present disclosure provides a computer program comprising a program code for performing the method according to the second aspect, or any of its implementation forms, when executed on a computer.

A fourth aspect of the present disclosure provides a non-transitory storage medium storing executable program code which, when executed by a processor, causes the method according to the second aspect or any of its implementation forms to be performed. It has to be noted that all devices, elements, and methods described in the present application could be implemented in the software or hardware elements or any kind of combination thereof. All steps which are performed by the various entities described in the present application as well as the functionalities described to be performed by the various entities are intended to mean that the respective entity is adapted to or configured to perform the respective steps and functionalities. Even if, in the following description of specific embodiments, a specific functionality or step to be performed by external entities is not reflected in the description of a specific detailed element of that entity which performs that specific step or functionality, it should be clear for a skilled person that these methods and functionalities can be implemented in respective software or hardware elements, or any kind of combination thereof.

BRIEF DESCRIPTION OF DRAWINGS

The above described aspects and implementation forms will be explained in the following description of specific embodiments in relation to the enclosed drawings, in which

FIG. 1 shows a schematic view of a network management device for performing device configuration mapping, according to an embodiment of the present invention, and two network devices to be configured by the network management device;

FIG. 2 shows a schematic view of a Siamese neural network model comprised in the network management device, according to an embodiment of the present invention;

FIG. 3 shows a schematic view of another Siamese neural network model comprised in the network management device, according to an embodiment of the present invention;

FIG. 4 shows a diagram illustrating the network management device processing vendorspecific data as an input;

FIG. 5 shows a flowchart of a method performed by the network management device, in order to improve similarity scores, according to an embodiment of the present invention; FIG. 6 shows a flowchart of another method performed by the network management device for augmenting user feedback, according to an embodiment of the present invention;

FIG. 7 shows a flowchart of another method performed by the network management device for selecting an optimal Siamese neural network model based on user feedback, according to an embodiment of the present invention; and

FIG. 8 shows a flowchart of another method performed by the network management device for performing device configuration mapping, according to an embodiment of the present invention.

DETAILED DESCRIPTION OF EMBODIMENTS

Illustrative embodiments of a network management device and a method are described with reference to the figures. Although this description provides a detailed example of possible implementations, it should be noted that the details are intended to be exemplary and in no way limit the scope of the application.

Moreover, an embodiment/example may refer to other embodiments/examples. For example, any description including but not limited to terminology, element, process, explanation, and/or technical advantage mentioned in one embodiment/example is applicative to the other embodiments/examples.

A concept of the present disclosure is based on similarities between device configurations, such as APIs and internal data models between network devices from different vendors. Network devices in a telecom network may need to support standardized networking protocols and may offer similar or even the same functionality and features. There may be a lexicon associated with such protocol s/standards, features, and device configurations. For this reason, it may be possible to identify equivalence or relationships between vendors-specific data that allows configurations from one network device to be mapped to another network device. For example, for a particular protocol, there may be common concepts and configuration items, which are supported and often have similar names, data types and relationships with other configuration items. Thus, it may be possible to identify equivalent or matched configurations for other devices. FIG. 1 shows a network management device 1100 adapted for mapping network devices, i.e., for performing device configuration mapping between at least two network devices 1200, 1300 in a telecom network according to an embodiment of the disclosure. The network management device 1100 may comprise processing circuitry (not shown) configured to perform, conduct or initiate the various operations of the network management device 1100 described herein. The processing circuitry may comprise hardware and software. The hardware may comprise analog circuitry or digital circuitry, or both analog and digital circuitry. The digital circuitry may comprise components such as application-specific integrated circuits (ASICs), field- programmable arrays (FPGAs), digital signal processors (DSPs), or multi-purpose processors. The network management device 1100 may further comprise memory circuitry (not shown), which stores one or more instruction(s) that can be executed by the processor or by the processing circuitry, in particular under control of the software. For instance, the memory circuitry may comprise a non-transitory storage medium storing executable software code which, when executed by the processor or the processing circuitry, causes the various operations of the network management device 1100 to be performed. In one embodiment, the processing circuitry comprises one or more processors and a non-transitory memory connected to the one or more processors. The non-transitory memory may carry executable program code which, when executed by the one or more processors, causes the network management device 1100 to perform, conduct or initiate the operations or methods described herein.

The telecom network may be based on software-defined networking, in which the network devices 1200, 1300 may be configurable for various services with suitable configurations. The network management device 1100 may be deployed in the telecom network to manage the network devices 1200, 1300, i.e., to provide the network devices 1200, 1300 with appropriate device configurations to ensure proper functioning of a service.

The network management device 1100 may obtain first vendor-specific data from a first network device 1200 and second vendor-specific data from a second network device 1300. The network devices 1200, 1300 may operate on a same protocol stack of the telecom network and may be from different telecom vendors. As a result, the network devices 1200, 1300 may have different device configurations for executing the same function in the same protocol stack. The first and the second vendor-specific data may be provided by each telecom vendor, such as user manuals, wherein device configurations are described in detail. The device configurations may comprise one or more of the following: device parameters, device models, device commands and XML Path (XPath) expressions.

Notably, the device configurations may be represented by a Yet Another Next Generation (YANG) language. The YANG language is a data modeling language and is defined by Request for Comments (RFC) 6020. Values of the device types, the parameters, and the interfaces may be defined by using the YANG language. Moreover, the XPath expressions may represent a sequence of commands in a joint fashion, created to represent device types and/or parameters and/or interfaces as a “sentence”.

After obtaining the first and the second vendor-specific data, the network management device 1100 may input the first and the second vendor specific data into a Siamese neural network model 1101 for conversion, analysis, and comparison.

FIG. 2 shows an enlarged schematic view of the Siamese neural network model 1101 that is comprised in the network management device 1100 according to an embodiment of the present disclosure.

As depicted in FIG. 2, the network management device 1100 may input the first vendor-specific data 2101 and the second vendor-specific data 2201 into the Siamese neural network model 1101. The Siamese neural network model 1101 may comprise a set of word embeddings 2102, 2202.

Notably, the word embeddings 2102, 2202 may be a set of learned word representations for natural language processing, wherein each word may be mapped to a vector. The word embeddings 2102, 2202 may be trained on text corpora from one or more telecom standards and/or protocols. Optionally, corresponding word embeddings may be used for particular types of the network devices. For example, if the network devices are routers for executing Broader Gateway Protocol (BGP), then RFC 4271 and/or RFC 1771 that standardize the BGP may be used to obtain/train the word embeddings. For another example, if the network devices are routers for executing Open Shortest Path First (OSPF) routing protocol, then RFC 2328 and/or RFC 5340 that standardize the OSPF protocol may be used to obtain/train the word embeddings. Optionally, the word embeddings may be obtained based on a conventional word-to-vector (Word2Vec) method, such as a Skip-Gram method, in a manner known to those skilled in the art, and thus not be described further herein.

After the first and the second vendor-specific data 2101, 2201 are input into the Siamese neural network model 1101, each vendor-specific data may be split word by word to obtain a set of words. The obtained words may be then multiplied by each corresponding matrix in the wording embeddings, and may further be analyzed by further layers in the Siamese neural network, to obtain a first matrix representation for the first vendor-specific data and a second matrix representation for the second vendor-specific data.

Then the network management device 1100 may obtain a set of similarity scores indicating semantic similarities between the first vendor-specific data and the second vendor-specific data based on an output of the Siamese neural network model 1101.

Notably, the Siamese neural network model 1101 may further comprise a pair of convolutional neural network (CNN) models 2103, 2203 that are configured to extract features.

In particular, the CNN models may be Long Short-Term Memory (LSTM) models. This is beneficial, since a semantically structured representation space may be learned such that simple metrics may suffice to capture similarity.

The Siamese neural network model 1101 may further comprise a similarity function 2004. The similarity function 2004 may be adapted to perform a pairwise comparison on the extracted features of the first and the second vendor-specific data in order to output the similarity scores 2005.

After obtaining the similarity scores, the network management device 1100 may obtain one or more pairs of matched device configurations and may configure the network devices 1200, 1300 accordingly.

FIG. 3 shows, as an example, an extended Siamese LSTM (ESLSTM) network model 3000 comprised in the network management device 1100, according to an embodiment of the present disclosure. The ESLSTM network model 3000 is built based on the Siamese network model 1101 shown in FIG. 2, and may have the same features as the Siamese network model 1101 describe above.

Using the ESLSTM network model 3000, without losing generality, both the first and the second network device may be BGP routers in this example. The first vendor specific data 3101 may comprise 4 XPaths as follows:

• XPath Al :

"vendorA-I0S-XR-ipv4-bgp-cfg:bgp/instance/instance-as/four-byte-as/neighbor- prefix/neighbor-afs/neighbor-af/af-name";

• XPath A2:

"vendorA-I0S-XR-ipv4-bgp-cfg:bgp/instance/instance-as/four-byte-as/vrfs/vrf/vrf- global/router-id";

• XPath A3 : "vendorA-I0S-XR-ipv4-bgp-cfg:bgp/instance/instance-as/four-byte-as/default-vrf/bgp- entity/global/router-id"; and

• XPath A4:

"vendorA-I0S-XR-ipv4-bgp-cfg:bgp/instance/instance-as/four-byte-as/default-vrf/bgp- entity/neighbors/hold-time" .

While the second vendor-specific data 3201 may comprise 4 XPaths as follows:

• XPath B 1 :

" vendorB-bgp:bgp/bgpcomm/bgpVrfs/bgpVrf/bgpVrfAFs/bgpVrfAF/afType";

• XPath B2:

" vendorB-bgp:bgp/bgpcomm/bgpVrfs/bgpVrf/bgpVrfAFs/bgpVrfAF/routerl";

• XPath B3 : "vendorB-bgp:bgp/bgpcomm/bgpVrfs/bgpVri7bgpVrfAFs/bgpVrfAF/routerl"; and

• XPath B4:

" vendorB-bgp:bgp/bgpcomm/bgpVrfs/bgpVrf/holdTime".

The network management device 1100 may represent/convert each XPath into a matrix representation by using word embeddings 3102, 3202 trained on an RFC text corpus.

FIG. 4 shows a method for representing a device configuration into a matrix representation. Optionally, the network management device 1100 may pre-process the input in order to obtain a set of words. For example, the XPath B4 has 5 words separated by “vendorB-bgp:bgp”, “bgpcomm”, “bgpVrfs”, “bgpVrf’ and “holdTime”. Each word may be represented by a vector based on the word embeddings 3102, 3202 trained on the RFC text corpus. In this example, each vector may have a dimension of 200. Thus, Tb in FIG. 3 is equal to 5 for the XPath B4, each is a 200-dimensional vector and a matrix of (x^\ x^\ x^\ x^ , x®)⁷ represents the XPath B4 like the matrix shown in FIG. 4. It is noted that values in the matrix of FIG. 4 are only given as an example. Similarly, T_a in FIG. 3 is equal to 8 for the XPath A4 and each x-^ is also a 200-dimensional vector. Then those vectors are provided as inputs to the corresponding LSTM network model.

Notably, a plurality of vectors with the same dimension may form a matrix, and the LSTM network model may be configured to encode each vector-based or matrix-based representation device configuration, in this example each XPath, into a final vector of a fixed dimensionality. A plurality of final vectors may form a final matrix that represents a plurality of device configurations.

Optionally, a first LSTM network model and a second LSTM network model, e.g., the LSTM a 3103 and LSTM b 3203 shown in FIG. 3, may have tied weights. That is to say, the first and the second LSTM network model 3103, 3203 may share common parameters.

For example, the first and the second LSTM network models 3103, 3203 may have 4 layers with 1000 cells at each layer. Parameters of the first and the second LSTM network models 3103, 3203 may be initialized with a uniform distribution between -0.09 and 0.09. In this example, a state of a final hidden layer of the LSTM network model is employed as the final vector of the fixed dimensionality.

Optionally, the network management device 1100 may extract features of the first and the second vendor specific data by using the LSTM network models 3103, 3203, and obtain the set of similarity scores based on the extracted features.

Optionally, each of the LSTM network models may be adapted to output the extracted features of each device configuration in a form of a feature vector. Optionally, the Siamese neural network may comprise a last layer, and the last layer may output a feature map corresponding to the first and the second inputted vendor specific data.

By outputting the feature vector and/or the feature map, the Siamese neural network model 1101 may be tunable and portable.

After obtaining the final matrix as an output of each LSTM network model, the network management device 1100 may compare the final matrix (from the left-side branch in FIG. 3) pairwise with another final matrix (from the right-side branch in FIG. 3), in order to obtain similarity scores.

This is achievable, since both final matrices have the same dimension.

To obtain similarity scores, the network management device 1100 may use a similarity function 3004 as shown exemplary in FIG. 3, or as follows: similarity(h_a, h_b~) = exp(— \\h_a — h_b ||_x) G [0,1] (1) in which h_a is a first matrix representation of the first vendor-specific data 3101, hb is a second matrix representation of the second vendor-specific data 3201, and \\h_a — h_b Hiis a distance function.

Notably, the distance function may a Euclidean distance function with a norm of 1. By using a norm of 1, incorrect judgements due to vanishing gradients of the Euclidean distance may be reduced.

Alternatively, the distance function may be a Manhattan distance function with a norm of 1.

Eventually, as an output of the Siamese network model 3000, the network management device 1101 may obtain the set of similarity scores 3005 in a form of a N X M matrix, where the first vendor-specific data comprises N device configurations and the second vendor-specific data comprises M device configurations. N and M are both positive integers, and at least one of them is larger than 1. Based on the N X M matrix, the network management device 1100 may determine, for every device configuration of the first network device 1200, a list of M similarity scores and associated device configurations of the second device, or vice versa. Thus, the network management device 1100 may determine the most similar device configurations as the match pair of device configurations.

For example, a set of similarity scores obtained as an output of the Siamese network model 1101 for the XPaths A1-A4 and B1-B4 mentioned above may be as follows:

Bl B2 B3 B4

Al 0.871 0.025 0.183 0.262

A2 0.149 0.673 0.439 0.086

A3 0.073 0.334 0.987 0.136

A4 .0.138 0.210 0.231 0.776

Then, the network management device 1100 may obtain match pairs of device configurations of (Al, Bl), (A2, B2), (A3, B3) and (A4, B4).

Finally, the network management device 1100 may configure the first and the second network devices 1200, 1300 correspondingly based on the match pairs of device configurations.

FIG. 5 shows a method 500 performed by the network management device 1100 in order to improve a credibility of the obtained similarity scores.

The method 500 may comprise the following steps:

• Step 501 : inputting vendor-specific data; and

• Step 502: applying the Siamese network model to obtain the set of similarity scores.

It is noted that the steps 501 and 502 may share the same features as mentioned in FIG. 1-4.

Furthermore, the method 500 may comprise one or more of the following steps:

• Step 503: performing clustering on the obtained set of similarity scores;

• Step 504: applying filtering to eliminate scores that are not reliable and to obtain a set of residual similarity scores;

• Step 505: re-ranking the set of residual similarity scores in order to increase a clarity of the residual similarity scores; and

• Step 506: obtaining a final set of similarity scores. In the step 503, various machine learning-based clustering algorithms, such as K-means, density-based spatial clustering of applications with noise (also known as: DBSCAN), may be used to obtain clusters, in which similar feature representations are gathered. For example, the network management device 1100 may calculate difference values between two adjacent scores, such as difference values between the first and the second scores, the second and the third scores etc. Then the network management device 1100 may associate the difference values with predictions based on ground-truth. Further, the network management device 1100 may observe for what difference values matched pairs of device configurations are obtained. Thus, the network management device 1100 may use the observed difference values as thresholds to cluster the obtained similarity scores.

In the step 504, the network management device 1100 may analyse the obtained clusters, and may determine ranges of similarity scores that are reliable. Then, the network management device 1100 may filter similarity scores that are not in the ranges, i.e., not reliable, along with associated device configurations. Thus, a smaller set of residual similarity scores is left.

In the step 505, the network management device 1100 may keep a limited number of top ranking similarity scores. For example, similarity scores and their associated device configurations that are ranked top 7 may remain. Thus, noise introduced by obviously unmatched pairs of device configurations may be reduced.

By performing clustering, filtering and/or re-ranking, accuracy of the similarity scores may be enhanced, thereby resulting in an improved system performance of the network management device 1100.

Taking the XPaths A1-A4 and XPaths B1-B4 described above as an example, with respect to the similarity scores of XPath A2, a difference value between the 1^st ranked XPath B2 scoring 0.673 and the 2^nd ranked XPath B3 scoring 0.439 is equal to 0.673-0.439=0.234. This difference value, in some circumstances, may be too small to be regarded as trusted or reliable. Therefore, by performing clustering, filtering, and/or re-ranking, clarity of similarity scores may be improved, i.e., difference values between ranked similarity scores may be larger and thus resulting in a more reliable outcome. Moreover, after the step 506, the method 500 may further comprise:

• Step 507: obtaining user feedback on the final set of similarity scores and storing the user feedback.

Notably, the network management device 1100 may employ the stored user feedback to re-train the Siamese network model 1101 or create a new Siamese network model.

FIG. 6 shows a method 600 for augmenting user feedback performed by the network management device 1100 according to an embodiment of the present disclosure.

The method 600 may comprise the following steps:

• Step 601 : obtaining user feedback;

• Step 602: processing user feedback;

• Step 603 : training a new Siamese neural network based on the processed user feedback;

• Step 604: re-training the Siamese neural network based on the processed user feedback;

• Step 605: processing vendor-specific data that are verified either positive or negative based on the user feedback

• Step 606: selecting an optimal Siamese neural network model and applying the optimal Siamese neural network model; and

• Step 607: obtaining similarity scores by using the optimal Siamese neural network.

Notably, the step 603 and step 604 may be repeated multiple times.

Moreover, in the step 605, the network management device 1100 may process the vendorspecific data based on the user feedback into a form of source-target and/or ground-truth pairs. Each source-target and ground-truth pair may comprise a source device configuration of the first network device 1200, a target device configuration of the second network device 1300, and/or a positive or negative label indicating whether the source device configuration and the target device configuration a match.

An example of two pairs of processed vendor-specific data based on the user feedback may be as follows: Positive pair: <confidence>:

“positive”

<target-path>:

"vendorA-I0S-XR-ipv4-bgp-cfg:bgp/instance/instance-as/four-byte-as/default- vrf/bgp-entity/neighbors/neighbor/remote-as/as-xx"

<source-path>:

"vendorB-bgp:bgp/ bgpcomm/ bgpVrfs/ bgpVrf/ bgpVrfAFs/ bgpVrfAF/ vrfAsNum"

Negative pair:

<confidence>:

"negative"

<target-path>:

"vendorA-IOS-XR-ipv4-bgp-cfg:bgp/instance/instance-as/four-byte-as/default- vrf/bgp-entity/neighbors/neighbor/neighbor-afs/neighbor-af/advertise-local-v6' <source-path>:

"vendorB-bgp:bgp/ bgpcomm/ bgpVrfs/ bgpVrf/ bgpVrfAFs/ bgpVrfAF/ acti veRoute Adverti se "

This is beneficial, since the network management device 1100 may employ both the positive and the negative pairs to refine network trainings or re-trainings, and remove any potential biasness originated from newly acquired vendor-specific data.

It is noted that the steps 606 and 607 may share the same functions and details from the perspective of FIG. 1-4 described above.

FIG. 7 shows a method for selecting an optimal Siamese network model performed by the network management device 1100 according to an embodiment of the present disclosure.

The method 700 comprises the following steps:

• Step 701 : accumulating user feedback;

• Step 702: training a current Siamese neural network model based on the accumulated user feedback to obtain a new Siamese neural network model; • Step 703 : assessing whether the new Siamese neural network model performs better than a previous (the current) Siamese neural network model. If yes, go to step 704, if not, go to step 706.

• Step 704: applying the new Siamese neural network model;

• Step 705: replacing the current Siamese neural network model with the new Siamese neural network model;

• Step 706: obtaining similarity scores by using an existing (either the new one or the current one, depending on the step 703) Siamese neural network.

The method 700 may be executed by the network management device 1100 continuously until a performance of the Siamese neural network model cease improving.

It is noted that the step 706 may share the same functions and details from the perspective of FIG. 1-4 described above.

FIG. 8 shows a method 800 according to an embodiment of the present invention. The method is performed by the network management device 1100 for performing the device configuration mapping between the at least two network devices.

The method 800 comprises the following steps:

• Step 801 : obtaining the first vendor-specific data from the first network device of the at least two network devices and the second vendor-specific data from the second network device of the at least two network devices;

• Step 802: inputting the first and the second vendor specific data to the Siamese neural network model comprising a set of word embeddings trained on one or more text corpora from one or more telecom standards and/or protocols;

• Step 803 : obtaining the set of similarity scores indicating semantic similarities between the first vendor-specific data and the second vendor-specific data based on the output of the Siamese neural network model;

• Step 804: obtaining the at least one matched pair of device configurations comprising the first device configuration of the first network device and the second device configuration of the second network device based on the similarity scores; and Step 805: configuring the first network device according to the first device configuration and the second network device according to the second device configuration.

Each step of the method 800 may share the same functions and details from the perspective of the network management device 1100 described above. Therefore, the corresponding method implementations 800 are not described again at this point.

In the disclosure, a network management device and a method for mapping device configurations are proposed. Embodiments of the disclosure provide a solution based on a Siamese neural network model. In particular, the network management device may employ the Siamese neural network comprising word embeddings to convert vendor-specific data into machine-learnable matrix representation. Further, the word embeddings are trained on RFC data corpus for the conversion. Thus, the network management device may obtain similarity scores from outputs of the Siamese neural network and may obtain match pairs of device configurations. The whole process may be automated and no human intervention is required.

Moreover, if user feedback is obtained, the network management device may optionally accumulate the user feedback and may optionally employ the accumulated user feedback as a training set for training a new Siamese neural network or re-training the existing Siamese neural network.

To summarize, the proposed method finds an appropriate solution based on machine learning approach. It leverages a network management device to automatically map device configurations, thereby resulting in a fast deployment speed to deploy a new service in a telecommunication network.

The present disclosure has been described in conjunction with various embodiments as examples as well as implementations. However, other variations can be understood and effected by those persons skilled in the art and practicing the claimed invention, from the studies of the drawings, this disclosure and the independent claims. In the claims as well as in the description the word “comprising” does not exclude other elements or steps and the indefinite article “a” or “an” does not exclude a plurality. A single element or other unit may fulfill the functions of several entities or items recited in the claims. The mere fact that certain measures are recited in the mutual different dependent claims does not indicate that a combination of these measures cannot be used in an advantageous implementation.

Furthermore, any method according to embodiments of the invention may be implemented in a computer program, having code means, which when run by processing means causes the processing means to execute the steps of the method. The computer program is included in a computer readable medium of a computer program product. The computer readable medium may comprise essentially any memory, such as a ROM (Read-Only Memory), a PROM (Programmable Read-Only Memory), an EPROM (Erasable PROM), a Flash memory, an EEPROM (Electrically Erasable PROM), or a hard disk drive.

Claims

1. A network management device (1100) for performing device configuration mapping between at least two network devices in a telecommunication network, wherein the at least two network devices are from different telecommunication vendors and operate on a same protocol stack of the telecommunication network, the network management device (1100) being configured to: obtain first vendor-specific data (2101) from a first network device (1200) of the at least two network devices and second vendor-specific data (2201) from a second network device (1300) of the at least two network devices; input the first and the second vendor specific data (2101, 2201) to a Siamese neural network model (1101) comprising a set of word embeddings (2102, 2202), wherein the set of word embeddings (2102, 2202) is trained on one or more text corpora from one or more telecommunication standards and/or protocols; obtain a set of similarity scores indicating semantic similarities between the first vendorspecific data (2101) and the second vendor-specific data (2201) based on an output (2005) of the Siamese neural network model (1101); obtain at least one matched pair of device configurations comprising a first device configuration of the first network device (1200) and a second device configuration of the second network device (1300) based on the similarity scores; and configure the first network device (1200) according to the first device configuration and the second network device (1300) according to the second device configuration.

2. The network management device (1100) according to claim 1, wherein each of the first and the second vendor-specific data (2201) comprises a set of device configurations, wherein the set of device configurations comprises one or more device models, and/or one or more device commands, and/or one or more XML Path, XPath, expressions of a corresponding network device.

3. The network management device (1100) according to claim 2, configured to: perform clustering on the device configurations of the first and the second vendorspecific data (2101, 2201) based on the similarity scores to obtain multiple clusters of device configurations.

25

4. The network management device (1100) according to claim 3, configured to: analyze the obtained multiple clusters of device configurations to obtain a trusted range of the similarity scores; eliminate one or more untrusted similarity scores from the similarity scores to obtain residual similarity scores, wherein the one or more untrusted similarity scorers are outside the trusted range; and eliminate device configurations associated with the one or more untrusted similarity scores from the obtained multiple clusters of device configurations.

5. The network management device (1100) according to claim 4, configured to: perform re-ranking on the residual similarity scores.

6. The network management device (1100) according to one of the claims 1 to 5, configured to: obtain user feedback on the at least one matched pair of device configurations; and store the obtained user feedback.

7. The network management device (1100) according to claim 6, configured to: perform one or more re-trainings of the Siamese neural network model (1101) based on the stored user feedback to obtain one or more re-trained Siamese neural network models.

8. The network management device (1100) according to claim 6 or 7, configured to: create one or more further Siamese neural network models based on the stored user feedback.

9. The network management device (1100) according to claim 7 and/or 8, configured to: compare performance of the one or more re-trained Siamese neural network models and/or the one or more further Siamese neural network models; select an optimal Siamese neural network model (1101) from the one or more re-trained Siamese neural network models and/or the one or more further Siamese neural network models based on the compared performance; and apply the selected optimal Siamese neural network model for performing device configuration mapping. The network management device (1100) according to one of the claims 1 to 9, configured to: process the first and the second vendor specific data into a form of source-target pairs and/or ground-truth pairs. The network management device (1100) according to one of the claims 1 to 10, configured to: encode the first and the second vendor specific data (2101, 2201) into matrix representations by using the set of word embeddings (2102, 2202). The network management device (1100) according to claim 11, configured to: calculate the set of similarity scores based on a similarity function as: similarity(h_a, h_b~) = exp(— \\h_a — h_b ||_x) G [0,1] wherein h_a is a first matrix representation of the first vendor-specific data (2101), hb is a second matrix representation of the second vendor-specific data (2201), and || h_a — h_b Hiis a distance function. The network management device (1100) according to claim 12, wherein the distance function is a Euclidean distance function with a norm of 1. The network management device (1100) according to claim 12, wherein the distance function is a Manhattan distance function with a norm of 1. The network management device (1100) according to one of the claims 1 to 14, wherein the Siamese neural network model (1101) comprises a pair of Long Short-Term Memory, LSTM, network models (3103, 3203), and the network management device (1100) is configured to extract features of the first and the second vendor specific data based on the set of word embeddings (2102, 2202) by using the LSTM network models; and obtain the set of similarity scores based on the extracted features.

16. The network management device (1100) according to claim 15, wherein each of the LSTM network models is adapted to output the extracted features in a form of a feature vector.

17. The network management device (1100) according to one of the claims 1 to 16, wherein the Siamese neural network comprises a last layer, wherein the last layer is adapted to output a feature map corresponding to the first and the second inputted vendor specific data.

18. The network management device (1100) according to one of the claims 1 to 17, configured to generate the set of word embeddings (2102, 2202) based on a FastText algorithm.

19. A method (800) for performing device configuration mapping between at least two network devices in a telecommunication network by a network management device, wherein the at least two network devices are from different telecommunication vendors and operate on a same protocol stack of the telecommunication network, the method comprising: obtaining (801) first vendor-specific data from a first network device of the at least two network devices and second vendor-specific data from a second network device of the at least two network devices; inputting (802) the first and the second vendor specific data to a Siamese neural network model comprising a set of word embeddings trained on one or more text corpora from one or more telecommunication standards and/or protocols; obtaining (803) a set of similarity scores indicating semantic similarities between the first vendor-specific data and the second vendor-specific data based on an output of the Siamese neural network model; obtaining (804) at least one matched pair of device configurations comprising a first device configuration of the first network device and a second device configuration of the second network device based on the similarity scores; and configuring (805) the first network device according to the first device configuration and the second network device according to the second device configuration.

28

20. Computer program comprising a program code for performing the method according to claim 19, when executed on a computer.

29