GENERATION ASSISTANCE DEVICE, GENERATION ASSISTANCE METHOD, AND GENERATION ASSISTANCE PROGRAM

Abstract

To enable the creation of a related machine learning model to be assisted when a certain machine learning model is altered.SOLUTION: A generation assistance device 1 is constituted to include: a storage device for storing component tables 61, 62 that store an alteration history of a plurality of component groups for a machine learning model, a model component management table 60 that stores the editions of the plurality of component groups; a model alteration recording unit 10 for inputting configuration information on an updated first machine learning model, adding an alteration history to the component tables 61, 62, and recording the editions of updated component groups to the model component management table 60; and a derived model parameter synthesizing unit 20 for generating a model parameter that includes components corresponding to the edition of updated component groups of the first machine learning model, for a second machine learning model that includes component groups of the same edition as that of the component groups of the first machine learning model before updating.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to a technique for assisting in generating a related machine learning model when a machine learning model is changed. [Background technology]

[0002] Machine learning models are used to monitor anomalies in multiple devices in factories, power facilities, etc. In order to monitor anomalies in multiple devices, a machine learning model tuned for each device is required.

[0003] For example, if an algorithm is changed, it becomes necessary to tune and learn the parameters of the machine learning model individually for each device.

[0004] For example, there is known a technique for recording and managing the history of changes (changes in data, parameters, and features) regarding such a machine learning model (see, for example, Patent Document 1). [Prior art documents] [Patent documents]

[0005] [Patent Document 1] U.S. Patent No. 9,996,804 Summary of the Invention [Problem to be solved by the invention]

[0006] As mentioned above, in order to monitor anomalies in multiple devices, a machine learning model tuned for each device is required, but most of the algorithms can often be shared between multiple devices.

[0007] In such a case, when the algorithm of a machine learning model of a certain device is changed, it is preferable to identify related machine learning models and generate those machine learning models.

[0008] For example, in the technology disclosed in Patent Document 1, it is possible to understand past machine learning models from the history of a certain machine learning model, but it is difficult to identify related machine learning models.

[0009] The present invention has been made in consideration of the above circumstances, and its purpose is to provide a technology that can assist in creating a related machine learning model when a certain machine learning model is changed. [Means for solving the problem]

[0010] In order to achieve the above-mentioned object, a generation support device according to one aspect is a generation support device that generates configuration information for constructing a machine learning model, and includes a memory unit that stores a plurality of component information items that store a change history for each of a component group, which is a group of a plurality of components for the machine learning model, and component management information that stores versions for the plurality of component groups that constitute the machine learning model, a recording unit that inputs configuration information for an updated first machine learning model, adds a change history to the component information based on the configuration information, and records the version of the component group of the updated first machine learning model in the component management information, and a configuration information generation unit that generates configuration information including components corresponding to the version of the updated component group of the first machine learning model after the update for a second machine learning model that includes a component group of the same version as the component group before the update of the first machine learning model. Effect of the Invention

[0011] According to the present invention, when a certain machine learning model is changed, it is possible to provide support for creating a related machine learning model. [Brief description of the drawings]

[0012] [Figure 1] FIG. 1 is a diagram illustrating an overall configuration of a generation support device according to an embodiment. [Diagram 2]FIG. 2 is a hardware configuration diagram of a generation support device according to an embodiment. [Diagram 3] FIG. 3 is a flowchart of a machine learning model generation support process according to an embodiment. [Figure 4] FIG. 4 is a flowchart of a configuration element update process according to one embodiment. [Diagram 5] FIG. 5 is a flowchart of a derived model parameter synthesis process according to an embodiment. [Figure 6] FIG. 6 is a diagram illustrating an example of machine learning model generation and a specific example of a machine learning model generation support process in one embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0013] The following embodiments will be described with reference to the drawings. Note that the following embodiments do not limit the scope of the invention, and not all of the elements and combinations thereof described in the embodiments are necessarily essential to the solution of the invention.

[0014] In the following description, information may be described using the expression "AAA table", but the information may be expressed in any data structure. In other words, to show that the information does not depend on the data structure, the "AAA table" may be called "AAA information".

[0015] FIG. 1 is a diagram illustrating an overall configuration of a generation support device according to an embodiment.

[0016] The generation support device 1 has a model change recording unit 10 as an example of a recording unit, a derived model parameter synthesis unit 20 as an example of a configuration information generation unit, a classification confirmation receiving unit 30, a source machine learning learning condition 40, a model state management table 50 as an example of state management information, a model component management table 60 as an example of component management information, a common component table (TB) 61 as an example of component information, an individualized component table (TB) 62 as an example of component information, an operation change element table (TB) 63 as an example of learning element information, and a derived machine learning model 70.

[0017] The model change recording unit 10 performs a process of recording the values ​​of each item (each element) of the source machine learning conditions 40 in the model component management table 60, the common component table 61, the individualized component table 62, and the operation change element table 63.

[0018] The derived model parameter synthesis unit 20 uses the model component management table 60, the common component table 61, the individualized component table 62, and the operation change element table 63 to create configuration information (model parameters) of the machine learning model that reflects the changed elements for the newly changed and recorded machine learning model (first machine learning model) and the machine learning model (second machine learning model) corresponding to the common common component, the individualized component, or the operation change element, and creates the machine learning model based on the configuration information. Note that the derived model parameter synthesis unit 20 may receive a designation of an application target for which a machine learning model is to be created, create configuration information only of the machine learning model for the application target for which the designation was received, and create the machine learning model based on the configuration information.

[0019] The classification confirmation receiving unit 30 displays the contents recorded by the model change recording unit 10 in the model component management table 60, the common component table 61, the individualized component table 62, and the operation change element table 63, receives instructions from the user regarding confirmation of the table to which the values ​​of each item of the recording source machine learning learning conditions 40 are recorded, changing the recording table, etc., and performs processing such as confirming the recording destination or changing the recording destination based on the instructions. Note that if no designation from the user is received, the classification confirmation receiving unit 30 does not need to be provided.

[0020] The source machine learning learning conditions 40 include learning conditions such as configuration information of the updated machine learning model that is the recording source and the dataset used for learning, and specifically, includes information on each item managed in the common component table 61, the individualized component table 62, and the operation change element table 63. The source machine learning learning conditions 40 may be obtained, for example, from a system that manages the editing history of the machine learning model.

[0021] The model state management table 50 stores records corresponding to each machine learning model. The records of the model state management table 50 store items of a model identifier, an application target, and a model state. The model identifier item stores an identifier (model identifier) ​​for identifying the machine learning model corresponding to the record. The application target item stores an identifier of an application target of the machine learning model corresponding to the record, for example, an equipment or the like for which anomaly detection is performed by the machine learning model. The model state item stores an application state of the machine learning model corresponding to the record. As the application state, for example, as shown in FIG. 6 described later, there are a state (development state) in which a common configuration of the machine learning model is developed by a data scientist (DS), a state (individualization state) in which the machine learning model is tuned to individual equipment or the like, and a state (operation state) in which the machine learning model is deployed and actually operated.

[0022] The model component management table 60 stores records corresponding to machine learning models. Note that the model component management table 60 may store multiple records for the same machine learning model. A record in the model component management table 60 includes fields for registration date and time, model identifier, common element Ver, individualization element Ver, and operation change element Ver. The registration date and time field stores the date and time when the record is stored. The model identifier field stores the model identifier of the machine learning model corresponding to the record. The common Ver field stores the version (version) of the common component (common component group: component group) constituting the machine learning model corresponding to the record. The individualization element Ver field stores the version (version) of the individualization component (individualization component group: component group) constituting the machine learning model corresponding to the record. The operation change element Ver field stores the version (version) of the element (learning element: operation change element group) to be changed in the operation of the machine learning model corresponding to the record.

[0023] The common component table 61 is a table that manages the change history of the common components that constitute the machine learning model, and stores records corresponding to each version of the common components. The common components include components that tend to be determined in the development stage of the machine learning model. The records of the common component table 61 include items such as common element Ver, pre-processing logic ID, learning algorithm ID, post-processing logic ID, and model learning parameters.

[0024] The common element Ver field stores the version of the common component corresponding to the record. The pre-processing logic ID field stores the identifier of the pre-processing logic (pre-processing logic ID) that constitutes the common component of the version corresponding to the record. The learning algorithm ID field stores the identifier of the learning algorithm (learning algorithm ID) that constitutes the common component of the version corresponding to the record. The post-processing logic ID field stores the identifier of the post-processing logic (post-processing logic ID) that constitutes the common component of the version corresponding to the record. The model learning parameters field stores the model learning parameters that constitute the common component of the version corresponding to the record.

[0025] The individualization component table 62 is a table for managing the history of the individualization components that constitute the machine learning model, and stores records corresponding to each version of the individualization component. The individualization components include components that may be determined at the stage of individualizing the machine learning model to suit the application target. Note that the same components may be included between the common components and the individualization components. The records of the individualization component table 62 include items such as the individualization element Ver, the model learning parameters, the pre-processing logic ID, the post-processing logic ID, the application target, and the reference destination.

[0026] The individualization element Ver field stores the version of the individualization component corresponding to the record. The model learning parameter field stores the model learning parameters constituting the individualization component of the version corresponding to the record. The pre-processing logic ID field stores the pre-processing logic ID of the pre-processing logic constituting the individualization component of the version corresponding to the record. The post-processing logic ID field stores the post-processing logic ID of the post-processing logic constituting the individualization component of the version corresponding to the record. The application target field stores an identifier of the application target of the machine learning model composed of the individualization component corresponding to the record. The reference field stores information indicating other records to be referenced as the value of the individualization component of the version corresponding to the record. Note that, when the reference field has a reference record setting, the value of an item in which a value is not set in the record becomes the value of the corresponding item in the reference record, or the value of the corresponding item in the reference record that is further connected via the reference. According to this reference field, the value of the item in a specific record can be commonly used for a group of related application targets.

[0027] The operation change element table 63 is a table that manages the history of operation change elements related to the learning of the machine learning model, and stores a record corresponding to each version of the operation change element. The operation change elements include elements (learning elements) that may be used in the operation stage of the machine learning model. The records of the operation change element table 63 include items such as operation change element Ver, used dataset ID, learning result, and application target.

[0028] The operation change element Ver field stores the version of the operation change element corresponding to the record. The used dataset ID field stores the identifier (used dataset ID) of the used dataset that constitutes the version of the operation change element corresponding to the record. The used dataset is a dataset used for training the machine learning model. The training result field stores the training results when the machine learning model is trained using the used dataset corresponding to the record. The applied target field stores the identifier of the target to which the operation change element corresponding to the record is applied.

[0029] The derived machine learning model 70 includes a created machine learning model (derived machine learning model) and information (model parameters) related to the configuration and learning of the derived machine learning model. The model parameters include a model identifier, a pre-processing logic ID, a learning algorithm ID, a post-processing logic ID, model learning parameters, and learning results.

[0030] The model identifier is a model identifier that identifies the derived machine learning model. The pre-processing logic ID is the pre-processing logic ID of the pre-processing logic that constitutes the derived machine learning model. The learning algorithm ID is the learning algorithm ID of the learning algorithm that constitutes the derived machine learning model. The post-processing logic ID is the post-processing logic ID of the post-processing logic that constitutes the derived machine learning model. The model learning parameters are the model learning parameters used in the derived machine learning model. The learning result is the result of learning the derived machine learning model using the corresponding usage dataset.

[0031] FIG. 2 is a hardware configuration diagram of a generation support device according to an embodiment.

[0032] The generation support device 1 is configured by a computer such as a PC (Personal Computer), a server, etc. The generation support device 1 includes a communication interface (I / F) 101, a CPU (Central Processing Unit) 102, an input device 103, a storage device 104 as an example of a storage unit, a memory 105, and a display device 106.

[0033] The communication interface 101 is, for example, an interface such as a wired LAN card or a wireless LAN card, and communicates with other devices via a network.

[0034] The CPU 102 executes various processes in accordance with the programs stored in the memory 105 and / or the storage device 104. In this embodiment, the CPU 102 executes a generation support program described later, thereby configuring a model change recording unit 10, a derived model parameter synthesis unit 20, and a classification confirmation receiving unit 30.

[0035] The memory 105 is, for example, a RAM (RANDOM ACCESS MEMORY), and stores the programs executed by the CPU 102 and other necessary information.

[0036] The storage device 104 is, for example, a hard disk or a flash memory, and stores a program executed by the CPU 102 (for example, a generation support program), various logics constituting the machine learning model, data used by the CPU 102 (for example, a data set used for learning the machine learning model), etc. In this embodiment, the storage device 104 stores the memory source machine learning learning conditions 40, a model state management table 50, a model component management table 60, a common component table 61, an individualized component table 62, an operation change element table 63, and a derived machine learning model 70.

[0037] The input device 103 is, for example, a mouse, a keyboard, etc., and accepts information input by a user of the generation support device 1. The display device 106 is, for example, a display, and displays and outputs various types of information.

[0038] Next, the machine learning model generation support process will be described.

[0039] FIG. 3 is a flowchart of a machine learning model generation support process according to an embodiment.

[0040] The model change recording unit 10 inputs the record source machine learning learning conditions 40 (step S1), and executes a component update process (step S2).

[0041] In the component update process, the model change recording unit 10 refers to a record in the model state management table 50 using the model identifier (target model identifier) ​​of the machine learning model corresponding to the recording source machine learning learning conditions 40, identifies an element table in which to record each item of the recording source machine learning learning conditions 40 based on the contents of the record, and if there is a change, adds a new record to the identified element table and records in the record information of the new version and the contents of the items of the recording source machine learning learning conditions 40. Note that a more detailed explanation of the component update process will be given later with reference to FIG. 4.

[0042] Next, the model change recording unit 10 obtains the latest record corresponding to the same model identifier as the target model identifier from the model component management table 60, extracts changes by comparing it with the items of the original machine learning learning conditions 40, duplicates the latest record, and updates the version information of the changed element table of this record to the information of the changed version (step S3).

[0043] Next, the derived model parameter synthesizing unit 20 executes a derived model parameter synthesizing process (step S4).

[0044] In the derived model parameter synthesis process, the derived model parameter synthesis unit 20 extracts records including versions of the common components, individualized components, and operation change elements before the change of the changed machine learning model from the model component management table 60, and synthesizes model parameters (configuration information) of the derived machine learning model based on the contents of the common components, individualized components, and operation change elements of the changed versions and the contents of the elements referenced by the extracted records. Furthermore, the derived model parameter synthesis unit 20 generates a derived machine learning model based on the model parameters, and performs learning of the generated derived machine learning model.

[0045] Next, the component update process (step S2) will be described in detail.

[0046] FIG. 4 is a flowchart of a configuration element update process according to one embodiment.

[0047] In the component update process, the model change recording unit 10 extracts the values ​​of each item from the record source machine learning learning conditions 40 (also simply referred to as learning conditions) (step S11).

[0048] Next, the model change recording unit 10 performs the process of loop 1 (steps S12 to S20) for each item of the learning conditions as a processing target. Here, the item as a processing target is referred to as a target item.

[0049] First, the model change recording unit 10 judges whether the target item is an item specific to any of the element tables (the common element table 60, the individualized element table 61, and the operation change element table 62) (step S12).

[0050] As a result, if the target item is not an item unique to the element table (step S12: No), the model change recording unit 10 refers to the record corresponding to the model identifier of the machine learning model that corresponds to the learning condition in the model state management table 50, obtains the value of the model state item of this record, determines the element table corresponding to the value of the model state item as the storage destination (step S13), and proceeds to step S15.

[0051] On the other hand, if the target item is an item unique to the element table (step S12: Yes), the model change recording unit 10 determines the element table in which the target item is a unique item as the storage destination (step S14), and proceeds to step S15.

[0052] In step S15, the model change recording unit 10 determines whether the storage destination of the target item is the individualized component table 62 and whether the latest version record corresponding to the target to which the learning condition is applied has a reference destination specified.

[0053] As a result, if no reference destination is specified (step S15: No), the model change recording unit 10 advances the process to step S18.

[0054] On the other hand, if a reference is specified (step S15: No), the model change recording unit 10 accepts a specification from the user as to whether to change the contents of the referenced record or the contents of the referenced record (step S16). If the value of an item is not set in the referenced record, the record in which the item is set is identified by tracing the references of each record in order until a record in which the value of the item is set is reached. In step S16, the content of which record is to be changed may be determined according to a predetermined process, rather than a user specification.

[0055] Next, the model change recording unit 10 determines a storage destination for the target item according to the user's designation (step S17), and advances the process to step S18.

[0056] In step S18, the model change recording unit 10 obtains the latest element version in the storage destination element table.

[0057] Next, the model change recording unit 10 determines whether the learning condition item has been changed from the contents of the record corresponding to the latest version in the storage destination element table (step S19). If the item has not been changed (step S19: No), the processing for the target item is terminated. If the item has been changed (step S19: Yes), the model change recording unit 10 merges the changes to the target item into the contents of the record corresponding to the latest version, creates a new record with the latest version as its version, and adds this record to the storage destination element table (step S20), thereby terminating the processing for the target item.

[0058] Next, the model change recording unit 10 performs the process of loop 1 (steps S12 to S20) for each item as a processing target, and when the process of loop 1 has been completed for all items as processing targets, ends the component update process.

[0059] Next, the derived model parameter synthesis process (step S4) will be described in detail.

[0060] FIG. 5 is a flowchart of a derived model parameter synthesis process according to an embodiment.

[0061] In the derived model parameter synthesis process, the derived model parameter synthesis unit 20 judges which element table was added in the element update process (step S41). As a result, when the derived model parameter synthesis unit 20 adds to the common element table 61 (step S41: common element TB), it executes the process of steps S42 to S47, when the derived model parameter synthesis unit 20 adds to the individualized element table 62 (step S41: individualized element TB), it executes the process of steps S48 to S54, and when the derived model parameter synthesis unit 20 adds to the operation change element table 63 (step S41: operation change element TB), it executes the process of steps S55 to S59. Note that when the derived model parameter synthesis unit 20 adds to multiple element tables, it executes the process corresponding to the common element table 61, the individualized element table 62, and the operation element table 63 in that order for the multiple element tables to which the derived model parameter synthesis unit 20 adds.

[0062] When an addition is made to the common component table 61 (step S41: common component TB), the derived model parameter synthesis unit 20 extracts records that refer to the version of the common component before the change from the model component management table 60 (step S42).

[0063] Next, the derived model parameter synthesis unit 20 performs the processing of loop 2 (steps S43 to S47) on each extracted record as a processing target. Here, in the explanation of the processing of this loop, the record as a processing target is referred to as a target record.

[0064] First, the derived model parameter synthesizing unit 20 identifies the version of the individualized component of the target record, and extracts the record (individualized component record) of the corresponding version from the individualized component table 62 (step S43).

[0065] Next, the derived model parameter synthesis unit 20 adds the values ​​of each item of the changed record in the common component table 61 to the machine learning model settings (configuration information: model parameters), and overwrites those items that are common to the individualized component record with the values ​​of the individualized component record (step S44).

[0066] Next, the derived model parameter synthesis unit 20 identifies the version of the operation change element of the target record, refers to the record of the corresponding version (operation change element record) in the operation change element table 63, and adds the value of this record to the model parameters of the machine learning model, overwriting common items with the values ​​of the operation change element record (step S45).

[0067] Next, the derived model parameter synthesis unit 20 creates a machine learning model (derived machine learning model) according to the model parameters of the machine learning model, and also performs machine learning on the created machine learning model using the set usage dataset, and adds the learning results to the model parameters of the machine learning model (step S46).

[0068] Next, the derived model parameter synthesis unit 20 adds a record set consisting of the version of the changed common component, the version of the extracted individualized component record, and the version of the extracted operation change element record to the model component management table 60 (step S47).

[0069] Next, the model change recording unit 10 performs the process of loop 2 (steps S43 to S47) for each record as a processing target, and when the process of loop 2 has been completed for all records as processing targets, ends the derived model parameter synthesis process.

[0070] When an addition is made to the individualized component table 62 (step S41: individualized component TB), the derived model parameter synthesis unit 20 extracts records that reference the individualized component version before the change from the model component management table 60 (step S48).

[0071] Next, the derived model parameter synthesis unit 20 performs the process of loop 3 (steps S49 to S54) on each extracted record as a processing target. Here, in the explanation of the process of this loop, the record as a processing target is referred to as a target record.

[0072] First, the derived model parameter synthesis unit 20 identifies the version of the common components and the version of the operation change elements of the target record, and extracts the record of the corresponding version (common component record) from the common component table 61 and the record of the corresponding version (operation change element record) from the operation change element table 63 (step S49).

[0073] Next, the derived model parameter synthesizing unit 20 extracts all individualized component records referenced by the extracted individualized component record and all individualized component records that reference the version of the individualized component before the change (step S50).

[0074] Next, the derived model parameter synthesis unit 20 adds the values ​​of each item of the record in the common component table to the settings (model parameters) of the machine learning model, and further arranges the extracted individualized component records and the changed individualized component records in order of reference priority, and overwrites the values ​​of the items contained in each individualized component record to the items that are common to the common component record (step S51).

[0075] Next, the derived model parameter synthesis unit 20 overwrites items that are common to the model parameters of the machine learning model with values ​​of the operation change element record (step S52).

[0076] Next, the derived model parameter synthesis unit 20 creates a machine learning model according to the model parameters of the machine learning model, and also performs machine learning on the created machine learning model using the set usage dataset, and adds the learning results to the model parameters of the machine learning model (step S53).

[0077] Next, the derived model parameter synthesis unit 20 adds a set of records consisting of the changed individualized component version, the version of the extracted common component record, and the version of the extracted operation change element record to the model component management table 60 (step S54).

[0078] Next, the model change recording unit 10 performs the process of loop 3 (steps S49 to S54) for each record, and when the process of loop 3 has been completed for all records, ends the derived model parameter synthesis process.

[0079] When an addition is made to the operation change element table 63 (step S41: operation change element TB), the derived model parameter synthesis unit 20 extracts records that reference the operation change element version before the change from the model component management table 60, extracts the common component records and individual component records of the versions contained in the extracted records, and extracts the necessary items from the common component records and individual component records (step S55).

[0080] Next, the derived model parameter synthesis unit 20 adds the values ​​of each item of the record in the common component table to the machine learning model settings (model parameters), and overwrites those items that are common to the individualized component records with the values ​​of the individualized component records (step S56).

[0081] Next, the derived model parameter synthesis unit 20 overwrites items that are common to the model parameters of the machine learning model with values ​​of the operation change element record (step S57).

[0082] Next, the derived model parameter synthesis unit 20 creates a machine learning model according to the model parameters of the machine learning model, and also performs machine learning on the created machine learning model using the set usage dataset, and adds the learning results to the model parameters of the machine learning model (step S58).

[0083] Next, the derived model parameter synthesis unit 20 adds a set of records including the changed operation change element version, the version of the extracted common component record, and the version of the extracted individualized component record to the model component management table 60 (step S59), and terminates the derived model parameter synthesis process.

[0084] According to the above-described machine learning model generation support process, for a machine learning model that used a pre-change element corresponding to a changed element, a machine learning model can be created in which the pre-change element is replaced with the changed element.

[0085] 6 is a diagram for explaining an example of machine learning model generation and a specific example of machine learning model generation support processing in one embodiment. Fig. 6 shows an example of generating a machine learning model for detecting anomalies in equipment 1 to 3 in a certain factory.

[0086] The application states of a machine learning model, from its development to its operation, include the development stage, in which common components are designed regardless of the equipment to which the machine learning model is to be applied; the personalization stage, in which the machine learning model designed in the development stage is modified to suit the individual equipment to which it is to be applied; and the deployment and operation stage (simply referred to as the operation stage), in which the machine learning model generated in the personalization stage is modified to suit the operational situation, etc., when it is put into operation.

[0087] First, in the development stage, a data scientist (DS) designs a machine learning model that is common to the target equipment 1 to 3. In this development stage, the pre-processing logic, learning algorithm, post-processing logic, etc. in the machine learning model are developed. The elements determined in this development stage are the common components. As a result, the machine learning model V1 is created.

[0088] Next, in the individualization stage, the developer performs a process of updating the machine learning model to one suitable for each piece of equipment at the site where the equipment to be applied is located. In this individualization stage, changes or modifications are made to the pre-processing logic, post-processing logic, etc. determined in the development stage, and parameters in the model learning parameters are changed, etc., according to each piece of equipment. In this individualization stage, components that may be changed, etc. are individualized components. In this individualization stage, a machine learning model V1.1 to which an individualized component corresponding to equipment 1 is applied, a machine learning model V1.0 to which an individualized component corresponding to equipment 2 is applied, and a machine learning model V1.2 to which an individualized component corresponding to equipment 3 is applied are created.

[0089] Next, in the operation stage, the developer performs a process of updating the machine learning model to one that corresponds to the operation of each piece of equipment at the site where the equipment to be applied is located. In this operation stage, a data set (usage data set) for learning the machine learning model is created in accordance with the operation of each piece of equipment, the machine learning model is learned using the usage data set, and the result of the learning (learning result) is obtained. In this operation stage, an element that may be changed or the like is an operation change element. In this operation stage, a machine learning model of V1.1.1 using the operation change element of equipment 1, a machine learning model of V1.0.1 using the operation change element corresponding to equipment 2, and a machine learning model of V1.2.1 using the operation change element corresponding to equipment 3 are created. After this, when the operation change element corresponding to equipment 3 is changed, a machine learning model of V1.2.2 using the changed operation change element is created.

[0090] In the state where the machine learning model is created in this way, when a new common component (v2) is developed for the common component (v1), the model parameters of the machine learning model using the updated common component (v2) are created for the machine learning model that used the common component (v1) by executing the above-mentioned derived model parameter synthesis process, and a corresponding machine learning model is generated. Specifically, as a machine learning model for facility 1, a machine learning model of V2.1.1 is generated based on the common component (v2) and the individualized components and operation change elements used for the common component (v1). Similarly, as a machine learning model for facility 2, a machine learning model of V2.0.1 is generated based on the common component (v2) and the individualized components and operation change elements used for the common component (v1). As a machine learning model for facility 3, a machine learning model of V2.2.2 is generated based on the common component (v2) and the individualized components and operation change elements used for the common component (v1). Note that, for the operation change elements, a machine learning model is generated using the latest operation change elements for the application target.

[0091] In this way, according to the above-described derived model parameter synthesis process, when a common component is updated, a machine learning model can be generated in which the common component is changed and used as the common component for a plurality of machine learning models generated using the changed common component. Note that, in the derived model parameter synthesis process, when an individualization component is updated, a machine learning model using the changed individualization component can be generated for a machine learning model that used the individualization component before the change. Also, in the derived model parameter synthesis process, when an operation change element is updated, a machine learning model using the changed operation change element can be generated for a machine learning model that used the operation change element before the change.

[0092] The present invention is not limited to the above-described embodiment, and can be modified as appropriate without departing from the spirit of the present invention.

[0093] For example, in the above embodiment, the components of the machine learning model are divided into two groups: common components and individualized components. However, the present invention is not limited to this, and the components may be divided into three or more groups.

[0094] In the above embodiment, a part or all of the processing performed by the CPU may be performed by a hardware circuit. The program in the above embodiment may be installed from a program source. The program source may be a program distribution server or a recording medium (e.g., a portable recording medium). [Explanation of symbols]

[0095] 1...generation support device, 10...model change recording unit, 20...derived model parameter synthesis unit, 30...classification confirmation reception unit, 40...recording source machine learning learning conditions, 50...model state management table, 60...model component management table, 61...common component table, 62...individualized component table, 63...operation change element table, 70...derived machine learning model, 101...communication I / F, 102...CPU, 103...input device, 104...storage device, 105...memory, 106...display device

Claims

1. A generation support device that generates configuration information for configuring a machine learning model, A storage unit that stores a plurality of pieces of component information that stores a change history for each of a plurality of component groups that are groups of components for a machine learning model, and component management information that stores versions for the plurality of component groups that constitute the machine learning model; A recording unit that inputs configuration information about an updated first machine learning model, adds a change history to the component information based on the configuration information, and records a version of the component group of the updated first machine learning model in the component management information; A configuration information generation unit that generates configuration information including components corresponding to the updated version of the component group of the first machine learning model after the update for a second machine learning model including a component group of the same version as the component group before the update of the first machine learning model; A generation assistance device having the above configuration.

2. The storage unit further stores learning element information that stores a change history of a learning element used in learning the machine learning model, The component management information further stores a version of the learning element of the machine learning model; The recording unit further inputs a learning element used for learning the updated first machine learning model, adds a change history to the learning element information based on the learning element, and records a version of the learning element of the updated first machine learning model in the component management information. The configuration information generation unit specifies a version of an updated learning element of the first machine learning model after the update for a second machine learning model that uses a learning element of the same version as that of a component group before the update of the first machine learning model, and includes the specified version of the learning element in the configuration information. The generation support device according to claim 1 .

3. The plurality of component groups include a first group including components that are likely to be determined in a development stage of the machine learning model, and a second group including components that may be determined in a personalization stage in which the machine learning model after the development stage is personalized to the application. The generation support device according to claim 1 .

4. The first group includes information for identifying a learning algorithm of the machine learning model, information for identifying a logic used for pre-processing of the learning algorithm, information for identifying a logic used for post-processing of the learning algorithm, and parameters used in the machine learning model; The second group includes information for identifying logic used in pre-processing of the learning algorithm, information for identifying logic used in post-processing of the learning algorithm, parameters used in the machine learning model, and information for identifying an application target of the machine learning model. The generation support device according to claim 3 .

5. The configuration information generation unit generates a machine learning model for the second machine learning model based on configuration information including components corresponding to an updated version of the component group of the first machine learning model after the update. The generation support device according to claim 1 .

6. The configuration information generation unit generates a machine learning model for the second machine learning model based on configuration information including components corresponding to an updated version of the component group of the first machine learning model after the update, and performs learning on the generated machine learning model using the learning elements. The generation support device according to claim 2 .

7. A plurality of the component groups include the same component, The storage unit is Further storing state management information for managing the application state of the machine learning model; The recording unit identifies an application state of the first machine learning model from the state management information, and determines component information for storing components included in updated configuration information for the first machine learning model based on the application state. The generation support device according to claim 1 .

8. A generation support method by a generation support device that generates configuration information for configuring a machine learning model, A plurality of pieces of component information that store change histories for each of the component groups, which are groups of a plurality of components for the machine learning model, and component management information that stores versions for the plurality of component groups that constitute the machine learning model are stored; Input configuration information including a configuration for an updated first machine learning model, add a change history to the component information based on the configuration information, and record a version of the component group of the updated first machine learning model in component management information; For a second machine learning model including a component group of the same version as the component group before the update of the first machine learning model, generating configuration information including components corresponding to the updated version of the component group of the first machine learning model after the update Generation support method.

9. A generation support program for causing a computer to execute a process of generating configuration information for configuring a machine learning model, The computer includes: Storing a plurality of pieces of component information that store change histories for each of the component groups, which are groups of a plurality of components for the machine learning model, and component management information that stores versions for the plurality of component groups that constitute the machine learning model; Input configuration information including a configuration for an updated first machine learning model, add a change history to the component information based on the configuration information, and record a version of the component group of the updated first machine learning model in a component management table; For a second machine learning model including a component group of the same version as the component group before the update of the first machine learning model, configuration information including components corresponding to the updated version of the component group of the first machine learning model after the update is generated. Generation assistance program.

Images