Setting error detection system, setting error detection method, and setting error detection program

Abstract

To automatically detect a configuration error of an optical transmission network (OTN).SOLUTION: A setting error detection system for detecting setting errors of optical communication devices configuring an optical transport network comprises: a configuration acquisition unit for acquiring setting information from optical communication devices C6 to C10; an error detection unit C3 for detecting an error from the acquired setting information by a neural network using a previously learned error identification model; and an error identification unit C4 for identifying an optical communication device, at which the error is detected, using a previously learned error analysis model before performing labeling. The setting error detection system outputs information showing the existence of an error and the identified error.SELECTED DRAWING: Figure 4

Description

[Technical Field]

[0001] The present invention relates to a configuration error detection system, a configuration error detection method, and a configuration error detection program for detecting configuration errors in optical communication devices that constitute an optical transport network. [Background technology]

[0002] Methods for detecting faults in an OTN (Optical Transport Network) have been studied for some time. As shown in Non-Patent Documents 1 to 3, active research is being conducted on diagnosing faults caused by abnormalities in the physical layer in the OTN and identifying the location of the faults.

[0003] Furthermore, as shown in Non-Patent Documents 4 to 6, techniques using NLP (Natural Language Processing) in networks are known. Natural language processing (NLP) based on LM (Language Model) has recently been utilized in tasks of understanding network configurations. For example, Non-Patent Document 4 discloses the use of NLP to solve integrated control problems in networks. [Prior art documents] [Non-patent literature]

[0004] [Non-Patent Document 1] H. Date, T. Kubo, T. Kawasaki and H. Maeda, "Silent Failure Localization on Optical Transport System," in IEEE Photonics Technology Letters, vol. 33, no. 13, pp. 649-651, 1 July1, 2021, doi: 10.1109 / LPT.2021.3084686. [Non-patent document 2] J. Luo, S. Huang, J. Zhang, X. Li and W. Gu, "A novel multi-fault localization mechanism in PCE-based multi-domain large capacity optical transport networks," OFC / NFOEC, 2012, pp. 1-3. [Non-Patent Document 3] F. Inuzuka et al., "Demonstration of a Novel Framework for Proactive Maintenance Using Failure Prediction and Bit Lossless Protection With Autonomous Network Diagnosis System," in Journal of Lightwave Technology, vol. 38, no. 9, pp. 2695-2702, 1 May1, 2020, doi: 10.1109 / JLT.2020.2967510. [Non-Patent Document 4] Huangxun Chen, Yukai Miao, Li Chen, Haifeng Sun, Hong Xu, Libin Liu, Gong Zhang, and Wei Wang. 2022. Software-defined network assimilation: bridging the last mile towards centralized network configuration management with NAssim. In Proceedings of the ACM SIGCOMM 2022 Conference (SIGCOMM '22). Association for Computing Machinery, New York, NY, USA, 281-297. https: / / doi.org / 10.1145 / 3544216.3544244 [Non-Patent Document 5] Houidi, Zied Ben, and Dario Rossi. "Neural language models for network configuration: Opportunities and reality check." arXiv preprint arXiv:2205.01398 (2022). [Non-patent document 6] M. -T. -A. Nguyen, SB Souihi, H. -A. Tran and S. Souihi, "When NLP meets SDN : an application to Global Internet eXchange Network," ICC 2022 - IEEE International Conference on Communications, 2022, pp. 2972-2977, doi: 10.1109 / ICC45855.2022.9838633. Summary of the Invention [Problem to be solved by the invention]

[0005] In optical transport networks (OTNs), centralized control of optical communication equipment is possible, but configuration errors can occur when configuring optical communication equipment due to differences in the control management systems between the equipment. For example, when configuring lightpaths, differences in the control plane structure and control automation level can lead to configuration errors due to inconsistencies within nodes, mismatches between management ports and physical connections, inconsistencies in control logic, or unpredictable conditions during dynamic adjustments. In commercial optical communication equipment, XML (Extensible Markup Language) is widely used for configuration. However, checking XML configuration files to find errors is difficult because it requires a great deal of effort, knowledge, and experience.

[0006] Another problem is that the possibility of configuration errors is much higher than the possibility of physical layer failures. Although the studies in Non-Patent Documents 1 to 3 use machine learning, they only use machine learning to perform prediction tasks and do not use language models to perform fault diagnosis of large-scale networks. Furthermore, the studies in Non-Patent Documents 4 to 6 are not targeted at OTN, but are related to electrical packet switching networks, and therefore cannot be applied to OTN.

[0007] The present invention has been made in consideration of the above circumstances, and aims to provide a configuration error detection system, a configuration error detection method, and a configuration error detection program that can automatically detect configuration errors in optical transport networks (OTN) by utilizing natural language processing (NLP) technology based on a language model (LM) using an attention-style deep neural network with pre-training. [Means for solving the problem]

[0008] (1) In order to achieve the above object, the present invention provides the following means: That is, the setting error detection system of the present invention is a setting error detection system that detects setting errors in optical communication devices that constitute an optical transport network, and Configuration information an acquisition unit that acquires the The configuration information acquired by the acquisition unit Pre-trained error discrimination model By inputting the above, an error detection unit that detects an error; The configuration information in which an error has been detected by the error detection unit Pre-trained error analysis model By using the input , The configuration information including the error is set Optical communication equipment Types of Identify do Error Identification Unit and The type of the optical communication device identified by the error identification unit Output the information shown an output unit, wherein the error identification model is generated by first pre-learning that learns vocabulary related to the optical transport network, second pre-learning that is executed after the first pre-learning and that learns sentences related to the optical transport network, and pre-learning that is executed after the second pre-learning and that uses the configuration information as an input, and the error analysis model is generated by the first pre-learning, the second pre-learning that is executed after the first pre-learning, and pre-learning that is executed after the second pre-learning and that uses the configuration information containing an error as an input. . [Effects of the Invention]

[0014] According to the present invention, it is possible to automate the detection of configuration errors in optical communication devices. Furthermore, since the location of the detected error can be identified, the process of detecting and identifying the error can be significantly accelerated compared to manual methods. Furthermore, it is possible to reduce the possibility of configuration errors occurring and reduce the burden on network administrators involved in identifying errors. [Brief explanation of the drawings]

[0015] [Figure 1] This is a diagram showing the learning process of "BERT". [Figure 2] 1 shows examples of pre-training and datasets. [Figure 3] 10 is a flowchart showing the operation of pre-training and error detection / identification stage A. [Figure 4] 1 is a diagram showing a schematic configuration of a setting error detection system according to the present invention; [Figure 5] 10 is a flowchart showing the operation of Stage B of the configuration error detection system according to the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0016] The inventors discovered that when creating a language model (LM), the accuracy of applying the language model can be improved by not only pre-training vocabulary but also by fine-tuning the vocabulary by learning sentences, and that the probability of identifying errors can be increased by following two steps: detecting whether there is an error and identifying the error, which led to the present invention.

[0017] In other words, the configuration error detection system of the present invention is a configuration error detection system that detects configuration errors in optical communication devices that constitute an optical transport network, and is characterized by comprising an acquisition unit that acquires configuration information from the optical communication devices, an error detection unit that uses a pre-trained error identification model to detect errors from the acquired configuration information using a neural network, and an error identification unit that uses a pre-trained error analysis model to identify and label the optical communication device in which the error was detected, and outputting information indicating the presence or absence of an error and the identified error.

[0018] As a result, the inventors have made it possible to automate the detection of configuration errors in optical communication devices. Furthermore, because detected errors can be identified, the process of error detection and identification is significantly accelerated compared to manual methods. Furthermore, it is possible to reduce the likelihood of configuration errors occurring and to reduce the burden on network administrators involved in identifying errors. Hereinafter, each embodiment of the present invention will be described in detail with reference to the drawings.

[0019] The present invention is characterized by implementing two steps for error detection and identification: first, a (task-specific) LM is used to identify correct and incorrect configurations, and then another (task-specific) LM is used to find the devices in the node where the error occurred. Based on this, an administrator can further identify the details of the error using manual or automated logic.

[0020] [Creating a language model] In this specification, a language model (LM) is based on a neural network (NN), and the NN specifically has a structure called a "Transformer" (see reference paper: Attention is all you need). The present invention utilizes a "Transformer-based NN." There are many types of "Transformer-based" LMs, but this specification describes an example using an LM called "BERT" (see reference paper: BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding). Unlike conventional NLP (natural language processing) models, "BERT" is designed to pre-train sentences from both the beginning and the end of the sentence. Furthermore, it has the advantage of having a large amount of data available for training and being able to flexibly adapt to various tasks. Note that the present invention is not limited to "BERT."

[0021] Figure 1 shows the training process of BERT. BERT is a pre-trained model. The left side of Figure 1 shows the first pre-training process. Here, masking and next sentence prediction are performed using vocabularies and sentences in the tens of millions, i.e., various large-scale corpora (general phrases such as news). LM can handle natural language (generally English), but does not currently support OTN vocabulary. The right side of Figure 1 shows the second pre-training process. Here, thousands to tens of thousands of sentences, especially sentences related to OTN, are trained.

[0022] FIG. 2 shows an example of pre-training and a dataset, and FIG. 3 is a flowchart showing the operation of Stage A of pre-training and error detection / identification. In FIGS. 2 and 3, first, the OTN vocabulary is used to expand the original language model LM(C1) (Step 01). Here, for example, "edfa, roadm, odu4" are trained as the "OTN vocabulary (terms related to optical communication equipment)" as a language model. This training is the first pre-training.

[0023] Next, to pre-train the original LM (C1) by expanding the dictionary, training is performed using an OTN corpus (Step 02). The OTN corpus (sentences) includes, for example, product manuals for OTN, transponder, wavelength selective switch (WSS), reconfigurable optical add-drop multiplexer (ROADM), optical switch, etc., as well as Google search results. This training results in fine-tuning as a second pre-training step. The above procedure creates a trained language model, "OTN-LM," which is capable of recognizing and understanding OTN terms.

[0024] Next, the status changes to acquire information from the network that manages the actual operation. That is, the OTN agent (C5) acquires configuration information from each optical communication device (Step 03). Here, the Agent, which is a computer that acquires configuration information from each optical communication device and outputs it to the specified destination, obtains (obtains) the status from the network element (network equipment / device), and an operator (human) generates a dataset. This "Dataset" is generally created by humans, and is a standard technique in the field of machine learning.

[0025] An example of an OTN dataset is shown in Figure 2 and has the following format (the following four items are required): <No., text, class, label> No.: the ID of each entry of the dataset; Text: by conjugating configurations on each device of a lightpath; Class: to identify the text (i.e., lightpath configurations) <correct> or <error>; Label: to indicate the error locations, eg, in transponder, ROADM, or optical switch etc.

[0026] The next "Step 4" and "Step 5" are the process of obtaining a "Task-specific LM", which corresponds to the fine-tuning shown on the right side of Figure 1. Here, in this specification, the error discrimination model (Error Discrimination Language Model) is called "EDiMo". To obtain this EDiMo(C3), we use a binary discrimination ( <class>All data sets for executing the task (regarding) are used to determine the Class (Step 04). The "Dataset" is generated at the discretion of the operator, but in this "Step 4", a decision is also made by a neural network (NN). Then, for optical communication equipment (devices) with errors, the errors are identified. In other words, a partial data set (Class = Error) is used to classify multiple <label>In this specification, the error analysis model is called "ENaMo". Then, this ENaMo (C4) is obtained. "Multi labeling" is one of the tasks, and "Multiple classification" is also one of the tasks. Here, for those with "Class=Error", we determine which ones to assign a "Label".

[0027] [Challenges specific to optical networks] Error identification in optical transport networks involves identifying the type of equipment in which an error occurred along a given optical path. Optical transport networks are composed of equipment with different functions, such as transponders (optical transceivers) and wavelength selective switches (WSSs).

[0028] On the other hand, identifying errors in "L2 / L3 networks (packet networks)" involves identifying errors in each network layer as follows: Example 1: Broadcast storm at the MAC layer Example 2: IP address mistakenly divided at the IP layer Example 3: VLAN or XVLAN rule violations in L2

[0029] Therefore, in the two-step method proposed in this specification, step 1 checks whether the configuration of the optical path is correct, and step 2 checks whether the configuration error occurs in this optical path. <label>(refer to the

[0030] In optical transport networks, the causes of errors are different from those in IP networks, and the response to errors and failures at each network layer is different. In IP networks, the cause of "configuration errors" is mainly violations of the "TCP / IP protocol" rules. In contrast, in optical transport networks, the cause of "configuration errors" is "physical layer problems" such as "port mapping, port on / off settings, wavelength settings, and optical amplifier settings," which are different from packet layer problems. In this invention, a method is adopted to obtain an LM that can learn errors, so the optical-specific issues and error causes mentioned above are reflected in the dataset.

[0031] In this invention, we adopt a two-stage error detection method, which has a higher accuracy (probability) of error detection than a direct error detection method. A pre-trained LM (fine-tuned LM) supplemented with a dictionary achieves better performance than a pure LM.

[0032] The above-mentioned "Step 1" and "Step 2" are performed by an external agent or a machine other than the network controller.

[0033] [Error detection and identification] FIG. 4 is a diagram showing a schematic configuration of a configuration error detection system according to the present invention, and FIG. 5 is a flowchart showing the operation of Stage B of the configuration error detection system according to the present invention. After pre-training the language model as described above, the pre-trained language model is deployed to the OTN agent (C5) (or network controller). The OTN agent (C5) acquires configuration information from optical communication devices (C6 to C10) (Step 21). Here, the OTN agent (C5) strengthens the learning model by acquiring the settings (configuration information) from each optical communication device (network equipment). Here, Mode [1] is a passive mode in which control messages are received, the control messages are converted into configuration information, and the converted (generated) configuration is detected. Mode [2] is a proactive mode in which the current configuration is monitored (i.e., the current configuration is acquired from each node) and the acquired configuration is detected.

[0034] The configuration is then determined using EDiMo (C3) (Steps 22 and 23). This means that errors are actually detected. The learning model created in the above steps is applied to check whether the configuration of each network device is correct. If the result is "correct," the process proceeds to Step 24, where the "OTN Agent C5" sends the configuration from C6 to C10 (in mode [1]) and terminates. In mode [2], the determination is based on the configuration information already set in each device, so there is no need to send the configuration to each device. If the result is not "correct," the process proceeds to Step 25 (i.e., when the classification result is <Error>), and the error configuration is labeled using ENaMo (C4) (Step 26). This makes it possible to analyze where errors are most likely to occur. Next, the labeling results (error location probability ranking) are reported to the OTN administrator (Step 26). For example, this is reported via the "Agent or Controller" interface (API: Application Programming Interface).

[0035] As described above, according to this embodiment, it is possible to automate the detection of configuration errors in optical communication devices. Furthermore, because detected errors can be identified, the process of error detection and identification is significantly accelerated compared to manual methods. Furthermore, it is possible to reduce the likelihood of configuration errors occurring and reduce the burden on network administrators involved in identifying errors.

[0036] [Other embodiments] The present invention is not limited to the above-described embodiments, and various modifications and variations are possible without departing from the technical spirit and scope of the present invention. That is, the setting error detection system according to the present invention can be realized as a method invention, and can also be realized by a program executed by one or more computers. This program can be provided or distributed by being recorded on a computer-readable recording medium or via a telecommunications line. [Explanation of symbols]

[0037] C3…EDiMo(Error discrimination language model) C4…ENaMo(Error analysis model) C5…OTN Agent C6~C10...Optical communication equipment< / label> < / label> < / class> < / error> < / correct>

Claims

1. A configuration error detection system for detecting a configuration error in an optical communication device that constitutes an optical transport network, comprising: an acquisition unit that acquires configuration information from the optical communication device; an error detection unit that detects an error in the configuration information by using the configuration information acquired by the acquisition unit as an input of an error identification model that has been pre-trained; an error identification unit that identifies the type of the optical communication device in which the configuration information including the error is set by using the configuration information in which an error has been detected by the error detection unit as an input to an error analysis model that has been pre-trained; an output unit that outputs information indicating the type of the optical communication device identified by the error identification unit; Equipped with the error identification model is generated by a first pre-training that learns a vocabulary related to the optical transport network, a second pre-training that is executed after the first pre-training and that learns sentences related to the optical transport network, and a pre-training that is executed after the second pre-training and that uses the configuration information as an input; A configuration error detection system, wherein the error analysis model is generated by the first pre-training, the second pre-training that is performed after the first pre-training, and the pre-training that is performed after the second pre-training and uses the erroneous configuration information as input.

2. 1. A configuration error detection method executed by a computer for detecting a configuration error in an optical communication device constituting an optical transport network, comprising: The computer an error identification model generated by a first pre-learning to learn vocabulary related to the optical transport network, a second pre-learning to learn sentences related to the optical transport network, which is executed after the first pre-learning, and a pre-learning to learn sentences related to the optical transport network, which is executed after the second pre-learning, and which uses configuration information of the optical communication device as an input; storing information indicating an error analysis model generated by the first pre-learning, the second pre-learning executed after the first pre-learning, and the pre-learning executed after the second pre-learning and using the erroneous configuration information as an input; The setting error detection method includes: acquiring the configuration information from the optical communication device; detecting an error in the configuration information by using the configuration information acquired from the optical communication device as an input to the error identification model; a step of identifying the type of the optical communication device in which the configuration information including the error is set by inputting the configuration information in which the error is detected into the error analysis model; outputting information indicating the type of the identified optical communication device; A configuration error detection method comprising at least:

3. A setting error detection program that causes a computer to execute the setting error detection method described in claim 2.

Images