Training data generation device, training data generation method, and training data generation program

Abstract

A learning data generation apparatus according to one embodiment of the present invention comprises: a first acquisition unit that acquires topology information indicating topology of a network including a plurality of network devices; and a connection pattern extraction unit that extracts, from the topology information, a connection pattern indicating the connection relationship between a first network device to which a fault is to be inserted and a second network device which is connected to the first network device.

Description

[Technical Field]

[0001] The present invention relates to a technique for generating learning data for training an AI (Artificial Intelligence) model that handles network failures. [Background technology]

[0002] AI models for dealing with network faults, such as fault detection models that detect network faults, require a large amount of fault data for model training. However, there is little fault data that can be collected in the target network environment where the AI ​​model will be applied.

[0003] To overcome this lack of fault data, a method for artificially generating fault data is being used. In this method, a network environment identical to the target network environment is constructed, and fault data is acquired by artificially generating network faults in that network environment. However, this method has the problem of high costs for constructing and maintaining the network environment. [Prior art documents] [Non-patent literature]

[0004] [Non-Patent Document 1] D. Li, K. Akashi, H. Nozue and K. Tayama, "A Mirror Environment to Produce Artificial Intelligence Training Data," in IEEE Access, Vol. 10, 2022, pp. 24578-24586, doi: 10.1109 / ACCESS.2022.3154825. [Non-patent document 2] H. Ikeuchi, J. Ge, Y. Matsuo and K. Watanabe, "A Framework for Automatic Failure Recovery in ICT Systems by Deep Reinforcement Learning," 2020 IEEE 40th International Conference on Distributed Computing Systems (ICDCS), 2020, pp. 1310-1315. Summary of the Invention [Problem to be solved by the invention]

[0005] An object of the present invention is to provide a technique that makes it possible to obtain, at low cost, failure data equivalent to that obtained in a target network environment. [Means for solving the problem]

[0006] A training data generation device according to one embodiment of the present invention includes a first acquisition unit that acquires topology information indicating the topology of a network including a plurality of network devices, and a connection pattern extraction unit that extracts, from the topology information, a connection pattern that indicates a connection relationship between a first network device into which a fault is to be inserted and a second network device connected to the first network device. [Effects of the Invention]

[0007] According to the present invention, a technique is provided that makes it possible to obtain, at low cost, failure data equivalent to that obtained in a target network environment. [Brief explanation of the drawings]

[0008] [Figure 1] FIG. 1 is a diagram for explaining an overview of a training data generation method according to an embodiment. [Figure 2] FIG. 2 is a diagram for explaining an overview of the training data generation method according to the embodiment. [Figure 3] FIG. 3 is a functional block diagram illustrating a training data generation device according to an embodiment. [Figure 4] FIG. 4 is a block diagram showing a hardware configuration of the training data generation device according to the embodiment. [Figure 5] FIG. 5 is a flowchart showing a training data generating method according to the embodiment. [Figure 6] FIG. 6 is a flowchart showing a training data generating method according to the embodiment. [Figure 7] FIG. 7 is a diagram for explaining a method for extracting a connection pattern according to the embodiment. [Figure 8] FIG. 8 is a diagram for explaining a method for extracting a connection pattern according to the embodiment. [Figure 9] FIG. 9 is a diagram for explaining a method for extracting a connection pattern according to the embodiment. [Figure 10] FIG. 10 is a diagram for explaining a method for extracting a connection pattern according to the embodiment. [Figure 11] FIG. 11 is a diagram for explaining a method for extracting a connection pattern according to the embodiment. [Figure 12] FIG. 12 is a diagram for explaining a method for extracting a connection pattern according to the embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0009] Hereinafter, an embodiment of the present invention will be described with reference to the drawings.

[0010] First, an overview of a training data generation method according to an embodiment will be described with reference to Figures 1 and 2. The training data generation method according to an embodiment generates training data used to train an AI model that can deal with network failures that occur in a network. This network is a network to which the trained AI model will be applied, and will hereinafter also be referred to as a target network.

[0011] Fig. 1 schematically illustrates a communication network 10, which is a non-limiting example of a target network. As shown in Fig. 1, the communication network 10 is a mesh network including a plurality of network devices 11, specifically, network devices 11-1 to 11-8. The communication network 10 can be used, for example, in an ICT (Information and Communication Technology) system.

[0012] Each network device 11 is configured to issue an alarm notifying the occurrence of a network failure when a network failure occurs in the network device 11. The issuance of the alarm makes it possible to carry out tasks such as estimating the location of the failure.

[0013] When a fault occurs in a certain network device 11, an alarm is detected not only in this network device 11 but also in the network devices 11 connected to this network device 11. For example, as shown in Figure 1, when a fault occurs in network device 11-1, network devices 11-2, 11-3, and 11-4 also detect an alarm. Furthermore, as shown in Figure 2, when faults occur simultaneously in network devices 11-1 and 11-2, network devices 11-3, 11-4, and 11-5 also detect an alarm. Hereinafter, network devices connected to the network device in which a fault has occurred will also be referred to as adjacent devices.

[0014] That is, a network failure has the characteristic of affecting not only the network device 11 in which the failure occurs, but also adjacent devices 11. The learning data generation method according to the embodiment utilizes this characteristic to build an environment in which data equivalent to failure data obtained in a target network can be artificially generated and acquired at low cost.

[0015] In the training data generation method according to the embodiment, an environment that matches the connection patterns of each device included in the target network is constructed, rather than constructing an environment identical to the entire target network. Then, by artificially generating failures in the constructed environment, failure data to be used as training data is generated. This allows the entire target network to be simulated using fewer network devices than the target network, making it possible to artificially generate training data at low cost.

[0016] 3 is a schematic diagram of a training data generation device 30 according to an embodiment. As shown in FIG. 3, the training data generation device 30 includes a first acquisition unit 31, a second acquisition unit 32, a connection pattern extraction unit 33, a fault insertion unit 34, and a third acquisition unit 35. The training data generation device 30 may be implemented by a single information processing device (computer) or by multiple information processing devices. For example, the training data generation device 30 may be implemented by two information processing devices 36 and 37, with the first acquisition unit 31, the second acquisition unit 32, and the connection pattern extraction unit 33 being provided in the information processing device 36, and the fault insertion unit 34 and the third acquisition unit 35 being provided in the information processing device 37.

[0017] The first acquisition unit 31 acquires topology information indicating the topology of the target network. For example, the first acquisition unit 31 receives the topology information from a network management device 38 that manages the target network. Alternatively, the first acquisition unit 31 may acquire the topology information from a computer-readable recording medium such as a USB memory.

[0018] The second acquisition unit 32 acquires simultaneous fault insertion information indicating the maximum number of network devices into which faults are to be simultaneously injected. For example, the second acquisition unit 32 receives the simultaneous fault insertion information from the network management device 38. Alternatively, the simultaneous fault insertion information may be input to the learning data generation device 30 by a human operator. Hereinafter, the maximum number of network devices into which faults are to be simultaneously injected is also referred to as the maximum number of fault insertions, and is represented as P. The maximum number of fault insertions P is an integer equal to or greater than 1.

[0019] The connection pattern extraction unit 33 extracts a connection pattern of network devices in the target network from the topology information acquired by the first acquisition unit 31, in accordance with the simultaneous fault insertion information acquired by the second acquisition unit 32. If p is an integer between 1 and P, the connection pattern of network devices in the target network includes a connection pattern that indicates a connection relationship between p adjacent network devices in the target network and one or more network devices connected to these p network devices. The p adjacent network devices in the target network are set as fault-occurrence devices. The fault-occurrence devices refer to network devices into which faults will be inserted in order to obtain fault data.

[0020] The simultaneous fault insertion information is not essential for extracting a connection pattern. For example, in a situation where simultaneous insertion of multiple faults is not expected, the simultaneous fault insertion information can be omitted.

[0021] The connection pattern extraction unit 33 outputs connection pattern information indicating the extracted connection patterns. For example, the connection pattern extraction unit 33 transmits the connection pattern information to the network management device 38. Then, a network environment set 39 is constructed that has environments that match the connection patterns extracted by the connection pattern extraction unit 33. The environments that match each connection pattern are part of the target network, and are hereinafter also referred to as partial networks or partial network environments.

[0022] The fault insertion unit 34 inserts faults into each of the partial networks included in the network environment set 39. The third acquisition unit 35 acquires fault data resulting from the insertion of faults into each of the partial networks as training data. The fault insertion unit 34 and the third acquisition unit 35 are collectively referred to as a training data generation unit. The training data generation unit acquires training data for training an AI model to deal with faults that occur in the target network by injecting faults into environments that match the connection patterns extracted by the connection pattern extraction unit 33. Fault insertion and training data acquisition can be performed using well-known techniques. For example, chaos engineering can be used. Therefore, a detailed description of fault insertion and training data acquisition will be omitted.

[0023] Fig. 4 schematically illustrates an example of the hardware configuration of the training data generation device 30. As illustrated in Fig. 4, the training data generation device 30 includes a CPU (Central Processing Unit) 41, a RAM (Random Access Memory) 42 as a main storage device, an auxiliary storage device 43, and a communication interface 44.

[0024] The CPU 41 is an example of a general-purpose processor capable of executing programs. The RAM 42 includes a volatile memory such as a Synchronous Dynamic Random Access Memory (SDRAM) and is used as a work area for the CPU 41. The auxiliary storage device 43 includes a non-volatile memory such as a Hard Disk Drive (HDD) or a Solid State Drive (SSD) and stores a group of programs including a learning data generation program, data, and the like.

[0025] The CPU 41 operates in accordance with a program stored in the auxiliary storage device 43. When executed by the CPU 41, the learning data generation program causes the CPU 41 to perform the processes described with respect to the learning data generation device 30. For example, the CPU 41 functions as a first acquisition unit 31, a second acquisition unit 32, a connection pattern extraction unit 33, a fault insertion unit 34, and a third acquisition unit 35 in accordance with the learning data generation program.

[0026] The communication interface 44 is an interface for communicating with an external device. For example, the CPU 41 communicates with the network management device 38 via the communication interface 44.

[0027] Note that a dedicated processor such as an FPGA (Field Programmable Gate Array) or an ASIC (Application Specific Integrated Circuit) may be used instead of or in addition to the general-purpose processor. The processing circuit refers to a general-purpose processor, a dedicated processor, or a combination of a general-purpose processor and a dedicated processor. The processing circuit is configured to function as a first acquisition unit 31, a second acquisition unit 32, a connection pattern extraction unit 33, a fault insertion unit 34, and a third acquisition unit 35.

[0028] A program such as the training data generation program may be provided to the training data generation device 30 in a state where it is stored on a computer-readable recording medium. In this case, the training data generation device 30 has a drive that reads data from the recording medium and acquires the program from the recording medium. Examples of recording media include magnetic disks, optical disks (CD-ROM, CD-R, DVD-ROM, DVD-R, etc.), magneto-optical disks (MO, etc.), and semiconductor memories. The program may also be distributed via a communication network. Specifically, the program may be stored on a server on the communication network, and the training data generation device 30 may download the program from the server.

[0029] Next, the operation of the training data generating device 30 will be described.

[0030] 5 and 6 are diagrams illustrating an example of a procedure for a training data generation method according to an embodiment. Each process shown in FIG. 5 and FIG. 6 is executed by the training data generation device 30.

[0031] 5, the first acquisition unit 31 acquires topology information indicating the topology of the target network. In step S52, the second acquisition unit 32 acquires simultaneous fault insertion information indicating the maximum number P of fault insertions.

[0032] In step S53, the connection pattern extraction unit 33 extracts a connection pattern of network devices in the target network from the topology information acquired in step S51, according to the maximum number P of fault insertions indicated in the simultaneous fault insertion information acquired in step S52. The connection pattern of network devices in the target network includes a connection pattern indicating a connection relationship between p first network devices that are adjacent to each other in the target network and set as fault-occurring devices, and a second network device connected to the first network devices. Step S53 includes steps S531, S532, and S533.

[0033] In step S531, the connection pattern extraction unit 33 calculates the maximum number of connections, which indicates the maximum number of network devices connected to the network device set as the failure-occurring device, based on the topology information. The maximum number of connections is denoted as N. For example, in the communication network 10 shown in FIG. 1, network devices 11-1, 11-2, 11-6, and 11-7 are each connected to three network devices, network devices 11-3, 11-4, and 11-5 are each connected to four network devices, and network device 11-8 is connected to two network devices. If the maximum number of failure insertions P is 1, the maximum number of connections N is 4. Furthermore, the connection pattern extraction unit 33 may calculate the total number of devices, which indicates the number of network devices included in the target network, based on the topology information. The total number of devices is denoted as M.

[0034] In step S532, the connection pattern extraction unit 33 detects all connection patterns included in the target network whose number of connections is less than or equal to the maximum number of connections N, according to the maximum number of fault insertions P indicated by the simultaneous fault insertion information. The number of connections indicates the number of network devices connected to a combination of one or more network devices set as the fault-occurring device. Details of the processing in step S532 are shown in FIG. 6.

[0035] 6, the connection pattern extraction unit 33 sets the variable n to 1. In step S62, the connection pattern extraction unit 33 determines whether the variable n is equal to or less than the maximum number of connections N. If the variable n is equal to or less than the maximum number of connections N (step S62; Yes), the flow proceeds to step S63.

[0036] In step S63, the connection pattern extraction unit 33 detects a connection pattern in which the number of connections is n. In step S64, the connection pattern extraction unit 33 increments the variable n by 1. After that, the flow returns to step S62.

[0037] In step S62, if the variable n is equal to or less than the maximum number of connections N, the flow proceeds to step S63, and if the variable n exceeds the maximum number of connections N (step S62; No), the flow ends. Therefore, a series of processes shown in steps S62 to S64 is repeated until the variable n exceeds N. As a result, a connection pattern with the number of connections from 1 to N is detected.

[0038] Referring again to FIG. 5, in step S533, the connection pattern extraction unit 33 selects a representative connection pattern from the connection patterns detected in step S532. Selecting a representative connection pattern includes deleting overlapping connection patterns among the detected connection patterns. For example, if the detected connection patterns include a first connection pattern and a second connection pattern, the connection pattern extraction unit 33 deletes the second connection pattern from the detected connection patterns if the first connection pattern encompasses the second connection pattern, i.e., if a portion of the first connection pattern matches the second connection pattern, and does not delete the second connection pattern if the first connection pattern does not encompass the second connection pattern. The connection pattern extraction unit 33 outputs the representative connection pattern as the connection pattern extracted from the target network.

[0039] A partial network that matches the connection pattern extracted in step S53 is constructed.

[0040] In step S54, the fault insertion unit 34 inserts a fault into the partial network that matches the connection pattern extracted in step S53. In step S55, the third acquisition unit 35 acquires, as learning data, fault data generated by the fault insertion into the partial network.

[0041] The process of extracting connection patterns will be specifically described below.

[0042] Fig. 7 shows a schematic diagram of a communication network 70, which is an example of a target network. As shown in Fig. 7, the communication network 70 includes network devices 71-1 to 71-11. Here, it is assumed that the network devices 71-1 to 71-11 are the same equipment.

[0043] Network device 71-1 is connected to network device 71-7. Network device 71-2 is connected to network device 71-9. Network device 71-3 is connected to network device 71-8. Network device 71-4 is connected to network device 71-10. Network device 71-5 is connected to network devices 71-7 and 71-8. Network device 71-6 is connected to network devices 71-9 and 71-10. Network device 71-7 is connected to network devices 71-1, 71-5, 71-8, 71-9, and 71-11. Network device 71-8 is connected to network devices 71-3, 71-5, 71-7, 71-10, and 71-11. Network device 71-9 is connected to network devices 71-2, 71-6, 71-7, 71-10, and 71-11. The network device 71-10 is connected to the network devices 71-4, 71-6, 71-8, 71-9, and 71-11. The network device 71-11 is connected to the network devices 71-7, 71-8, 71-9, and 71-10.

[0044] In the communication network 70, the total number of devices M is 11. If the maximum number of fault insertions P is 1, the maximum number of connections N is 5.

[0045] 8 shows all connection patterns in the communication network 70 when the maximum number of fault insertions P is 1. In FIG. 8, connection pattern 8a is a connection pattern when network devices 71-1, 71-2, 71-3, and 71-4 are set as faulty devices. Connection pattern 8b is a connection pattern when network devices 71-5 and 71-6 are set as faulty devices. Connection pattern 8c is a connection pattern when network devices 71-7, 71-8, 71-9, and 71-10 are set as faulty devices. Connection pattern 8d is a connection pattern when network device 71-11 is set as faulty device.

[0046] The maximum number of devices required to construct a partial network that matches these connection patterns is six. On the other hand, to construct a network environment identical to communication network 70, 11 network devices are required. Therefore, the training data generation method according to this embodiment makes it possible to generate training data using fewer network devices.

[0047] Furthermore, connection patterns 8a and 8b are included in connection patterns 8c and 8d, respectively. Therefore, if an environment that matches connection patterns 8c and 8d is constructed, an environment that matches all connection patterns can be used, making it possible to insert faults and generate fault data in the communication network 70.

[0048] Fig. 9 schematically shows a communication network 90, which is an example of a target network. As shown in Fig. 9, the communication network 90 includes network devices A1 to A4, B1 to B7, and C1. The network devices A1 to A4 belong to layer A, the network devices B1 to B7 belong to layer B, and the network device C1 belongs to layer C. Here, different facilities are used for each layer, and the communication network 90 is constructed with three types of facilities: facility A, facility B, and facility C.

[0049] Network device A1 is connected to network devices A3 and A4. Network device A2 is connected to network devices A3 and A4. Network device A3 is connected to network devices A1, A2, B1, and B2. Network device A4 is connected to network devices A1, A2, B3, and B4. Network device B1 is connected to network device A3. Network device B2 is connected to network devices A3 and B5. Network device B3 is connected to network devices A4 and B6. Network device B4 is connected to network devices A4 and B6. Network device B5 is connected to network device B2. Network device B6 is connected to network device B4. Network device B7 is connected to network devices B3 and C1. Network device C1 is connected to network device B7.

[0050] In the communication network 90, the total number of devices M is 12, there are four network devices (facility A) belonging to layer A, seven network devices (facility B) belonging to layer B, and one network device (facility C) belonging to layer C. If the maximum number of fault insertions P is 1, then the maximum number of connections N is 4.

[0051] 10 shows all connection patterns in a communication network 90 when the maximum number of fault insertions P is 1. In FIG. 10, connection pattern 10a includes connection patterns 10a-1, 10a-2, and 10a-3, each of which has a connection number of 1. Connection pattern 10a-1 is a connection pattern when network device B1 is set as the faulty device. Connection pattern 10a-2 is a connection pattern when network device C1 is set as the faulty device. Connection pattern 10a-3 is a connection pattern when network devices B5 and B6 are set as the faulty devices.

[0052] Connection pattern 10b includes connection patterns 10b-1, 10b-2, and 10b-3, each with two connections. Connection pattern 10b-1 is a connection pattern when network device B7 is set as the failed device. Connection pattern 10b-2 is a connection pattern when network devices B2, B3, and B4 are set as the failed devices. Connection pattern 10b-3 is a connection pattern when network devices A1 and A2 are set as the failed devices.

[0053] The connection pattern 10c includes a connection pattern 10c-1 in which the number of connections is 4. The connection pattern 10c-1 is a connection pattern in which the network devices A3 and A4 are set as failed devices.

[0054] The numbers of pieces of equipment required to construct these connection patterns 10a, 10b, and 10c are three pieces of equipment A, two pieces of equipment B, and one piece of equipment C. On the other hand, the numbers of pieces of equipment required to reproduce an environment identical to that of communication network 90 are four pieces of equipment A, seven pieces of equipment B, and one piece of equipment C. Therefore, the training data generation method according to this embodiment makes it possible to generate training data using fewer network devices.

[0055] Furthermore, connection pattern 10a-1 is included in connection pattern 10b-2, connection patterns 10a-2 and 10a-3 are included in connection pattern 10b-1, and connection pattern 10b-3 is included in connection pattern 10c-1. Therefore, if an environment that matches connection patterns 10b-1, 10b-2, and 10c-1 is constructed, an environment that matches all connection patterns can be used, making it possible to insert faults and generate fault data in communication network 90.

[0056] Referring again to the communication network 70 shown in FIG. 7, a case where the maximum number of fault insertions P is 2 will be described.

[0057] If there are 11 network devices 71, there are 55 possible combinations in which faults can be simultaneously inserted into two network devices. Of these, there are 32 combinations in which faults can be simultaneously inserted into two adjacent network devices, as shown in FIG. 11. In FIG. 11, the notation (i)-(j) indicates the combination of network device 71-i and network device 71-j. When duplicate combinations are eliminated, the following 16 possible combinations remain.

[0058] (1)-(7);(2)-(9);(3)-(8);(4)-(10); (5)-(7);(5)-(8);(6)-(9);(6)-(10); (7)-(8);(7)-(9);(7)-(11); (8)-(10);(8)-(11); (9)-(10);(9)-(11); (10)-(11);

[0059] 12 shows connection patterns of the communication network 70 when the maximum number of fault insertions P is 2. In FIG. 12, connection pattern 12a-1 is a connection pattern in which network devices 71-1 and 71-7, or network devices 71-2 and 71-9, or network devices 71-3 and 71-8, or network devices 71-4 and 71-10 are set as faulty devices. Therefore, combinations (1)-(7), (2)-(9), (3)-(8), and (4)-(10) can be simulated by connection pattern 12a-1. Connection pattern 12b-1 is a connection pattern in which network devices 71-5 and 71-7, or network devices 71-6 and 71-9 are set as faulty devices. Therefore, combinations (5)-(7) and (6)-(9) can be simulated by connection pattern 12b-1. Connection pattern 12b-2 is a connection pattern when network devices 71-5 and 71-8 or network devices 71-6 and 71-10 are set as failed devices. Therefore, combinations (5)-(8) and (6)-(10) can be simulated by connection pattern 12b-2. Connection pattern 12c-1 is a connection pattern when network devices 71-7 and 71-8 or network devices 71-9 and 71-10 are set as failed devices. Therefore, combinations (7)-(8) and (9)-(10) can be simulated by connection pattern 12c-1. Connection pattern 12c-2 is a connection pattern when network devices 71-7 and 71-9 or network devices 71-8 and 71-10 are set as failed devices. Therefore, combinations (7)-(9) and (8)-(10) can be simulated by connection pattern 12c-2. The connection pattern 12c-3 is a connection pattern when the network devices 71-7 and 71-11, or the network devices 71-8 and 71-11, or the network devices 71-9 and 71-11, or the network devices 71-10 and 71-11 are set as the failed devices. Therefore, the combinations (7)-(11), (8)-(11), (9)-(11), and (10)-(11) can be simulated by the connection pattern 12c-3.

[0060] The maximum number of devices required to construct a partial network that matches these connection patterns is 9. Therefore, the training data generation method according to this embodiment makes it possible to generate training data using fewer network devices.

[0061] As described above, the training data generation device 30 according to the embodiment acquires topology information indicating the topology of a target network, acquires simultaneous fault insertion information indicating the maximum number of fault insertions P, and extracts, from the topology information, a connection pattern indicating the connection relationships between a network device into which a fault will be inserted and network devices connected to the network device according to the simultaneous fault insertion information. This configuration simulates the connection pattern of the target network using a smaller number of network devices than the number of network devices included in the target network, making it possible to artificially obtain fault data equivalent to that obtained in the target network. Therefore, even if it is not possible to construct a network environment exactly identical to the target network, it is possible to obtain fault data used to train an AI model that can handle faults that occur in the target network environment.

[0062] The process of extracting connection patterns includes detecting connection patterns that indicate connection relationships between p adjacent network devices in the target network and network devices connected to the p adjacent network devices, and deleting overlapping connection patterns from the detected connection patterns. For example, when the detected connection patterns include a first connection pattern and a second connection pattern, deleting the connection patterns includes deleting the second connection pattern when a part of the first connection pattern matches the second connection pattern. This configuration makes it possible to avoid overlapping connection patterns.

[0063] The present invention is not limited to the above-described embodiments, and various modifications can be made in the implementation stage without departing from the spirit of the invention. Furthermore, the embodiments may be implemented in appropriate combinations, in which case the combined effects can be obtained. Furthermore, the above-described embodiments include various inventions, and various inventions can be extracted by combining selected components from the disclosed components. For example, if the problem can be solved and the effects can be obtained even if some components are removed from all the components shown in the embodiments, the configuration from which these components are removed can be extracted as an invention. [Explanation of symbols]

[0064] 10. Communication Network 11...Network equipment 30...Learning data generation device 31…First acquisition part 32…Second acquisition part 33...Connection pattern extraction unit 34...Fault insertion section 35...Third Acquisition Department 36, 37...Information processing device 38...Network management device 39...Network environment set 41...CPU 42...RAM 43…Auxiliary storage device 44...Communication interface 70...Communication Network 71...Network equipment 90...Communication Network A1~A4, B1~B7, C1...Network devices

Claims

1. a first acquisition unit that acquires topology information indicating a topology of a network including a plurality of network devices; a connection pattern extraction unit that extracts, from the topology information, a connection pattern that indicates a connection relationship between a first network device into which a fault is to be inserted and a second network device connected to the first network device; A training data generation device comprising:

2. a second acquisition unit that acquires simultaneous fault injection information indicating a maximum number of network devices into which faults are to be injected simultaneously; Let P be the maximum number of network devices into which faults can be simultaneously inserted, and p be an integer between 1 and P. Extracting the connection pattern includes detecting a connection pattern indicating a connection relationship between p adjacent network devices in the network and a network device connected to the p adjacent network devices; removing duplicate connection patterns from among the detected connection patterns; The training data generating device according to claim 1 , comprising:

3. the detected connection pattern includes a first connection pattern and a second connection pattern; 3. The training data generation device according to claim 2, wherein deleting the connection pattern comprises deleting the second connection pattern from the detected connection patterns when a part of the first connection pattern matches the second connection pattern.

4. 2. The training data generation device according to claim 1, further comprising a training data generation unit that obtains training data for training a model to deal with failures that occur in the network by artificially inserting failures into an environment that matches the extracted connection pattern.

5. A computer-implemented training data generation method, comprising: obtaining topology information indicating a topology of a network including a plurality of network devices; extracting, from the topology information, a connection pattern that indicates a connection relationship between a first network device into which a fault is to be inserted and a second network device connected to the first network device; A training data generation method comprising:

6. A training data generation program for causing a computer to function as each unit included in the training data generation device according to any one of claims 1 to 4.

Images