Model inference device, method, and program

Abstract

To suppress the cost of updating a model by executing the updating of the model more appropriately.SOLUTION: In order to solve the above problem, the present invention provides a model update / deployment system 1 which is a model inference device for inferring a model in machine learning. The system includes: a storage unit 40 configured to store a pre-update model which is a model before updating, and a post-update model; a feature quantity data creation unit 33 configured to create a feature quantity indicating a feature related to the relevancy between the pre-update model and the post-update model in data used when updating the model; an update learning model construction unit 34 configured to construct an update learning model that learns an update method of the updating based on the pre-update model, the post-update model, and the feature quantity; and an update learning model inference unit 36 configured to infer a predetermined model based on the update learning model.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present invention relates to the management of so-called machine learning models. Among them, it relates to a technology for managing models in various operations such as manufacturing, design, and maintenance operations. Note that the management of the model includes updating and deploying.

Background Art

[0002] Currently, machine learning models (hereinafter referred to as models) constructed on a computer system are used to support various operations. Also, for more efficient operation support, the models are updated and deployed to other sites, other operations, etc.

[0003] In such updates and deployments, models are constructed based on data obtained in operations and know-how held by engineers, workers, etc. Then, when applied to operations, if the accuracy of the model deteriorates, the model is updated for the purpose of improving model accuracy. Also, the constructed model is utilized in operations, and the obtained results are deployed to other sites, other operations, etc.

[0004] As a prior art document regarding such model updates and deployments, there is Patent Document 1. Patent Document 1 describes that the results obtained in past analysis operations are stored on a database, and when updating the model, cases similar to the subject case are searched from the database, and model updates and deployments are performed based on the search results.

[0005] Also, as another prior art document, there is Patent Document 2. Patent Document 2 describes that a graphical model called a Bayesian network is created with prior knowledge, and when a failure event occurs, the Bayesian network of the sub-graph corresponding to the subject event is constructed and updated, and then the probability distribution of the previously created model is updated.

Prior Art Documents

Patent Documents

[0006] [Patent Document 1] Japanese Patent Application Laid-Open No. 2019-79216 [Patent Document 2] Japanese Patent Application Laid-Open No. 2009-223362 [Summary of the Invention] [Problems to be Solved by the Invention]

[0007] Here, in the management of models including model update and deployment, parameter management is required according to the situation of the application destination. For example, in the manufacturing industry, models called statistical models or machine learning models are used to solve problems, but it is necessary to update and manage the parameters used in the models according to the type of device, usage environment, etc.

[0008] Therefore, in the management of models, it is necessary to perform management including parameter update. In particular, if the application destination of the model is a group of devices with a small quantity and a large variety of items, it is necessary to manage and update parameters for each device. For this reason, the number of parameters to be managed becomes enormous, and as a result, the management cost becomes high.

[0009] Here, in Patent Document 1, a plurality of data resources and analysis results are input, converted into event data that is easy to calculate by a computer, and stored on a database. Then, at the time of model update, cases similar to the target case can be searched from the database, and the results can be reflected in the model update. However, in Patent Document 1, it is necessary to store all cases similar to the target case in the database in advance. If the target case does not exist in the database, it is necessary to reconstruct the model, and the number of parameters according to the type of device, usage environment, etc. cannot be reduced.

[0010] In Patent Document 2, it is possible to reduce the management cost of parameters by constructing a graphical model that includes the target in advance, updating its partial model, and reflecting the update result in the original model. However, in Patent Document 2, it is necessary to construct a model that includes each target model, and for each parameter setting of the model, it is necessary to repeat the loop of parameter setting and evaluation, resulting in a high management cost.

[0011] As described above, in Patent Document 1 and Patent Document 2, it is possible to update parameters. However, for various conditions, repeated operations of model reconstruction and parameter adjustment are required, and the cost of model update becomes enormous.

[0012] Therefore, an object of the present invention is to more appropriately execute model update and suppress its cost.

Means for Solving the Problems

[0013] To solve the above problems, in the present invention, the model update itself is learned. As an example of this, in the present invention, the determination itself of the conditions for parameter adjustment, management, and update is expressed by a statistical model or a machine learning model, and this is learned. Then, this result can be reflected in the update and deployment of the model.

[0014] More specifically, in a model inference device that infers a model in machine learning, a storage unit that stores a pre-update model, which is a model before update, and a post-update model, and a feature amount that indicates a feature regarding the relevance between the pre-update model and the post-update model in the data used when updating the model As differential information indicating the difference between the pre-update model and the post-update model A feature amount data creation unit that creates By performing supervised learning with the post-update model as the training data using the pre-update model, the data used when constructing the pre-update model, and the data used when constructing the post-update model An update learning model construction unit that learns the update method of the update, and an update learning model inference unit that infers a predetermined model based on the update learning model It is a GCN It is a model inference device having Then, during the learning, the feature data creation unit creates, as the differential information, the difference between the edge information, which is the estimated parameter information, and the edge information given as the training data. The updated learning model construction unit constructs the updated learning model by feeding back the differential information to the updated learning model

[0015] ​The present invention also includes a model inference method using this model inference device. Furthermore, program products such as a program for causing a computer, which is the model inference device, to function and a storage medium storing the program are also included in the present invention.

Effects of the Invention

[0016] According to a typical embodiment of the present invention, in the present invention, it is possible to more appropriately execute the update of the model and suppress the cost of updating the model.

[0017] Problems, configurations, and effects other than those described above will be clarified by the description of the following embodiments.

Brief Description of the Drawings

[0018]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Figure 9

Figure 10

Figure 11

Figure 12

Figure 13

Figure 14

Figure 15

Figure 16

Embodiments for Carrying Out the Invention

[0019] Hereinafter, one embodiment of the present invention will be described with reference to the drawings. In each figure, the same parts are generally denoted by the same reference numerals, and repeated explanations thereof are omitted. In this embodiment, a graphical model is taken as an example of the model to be updated, and a graph neural network is adopted as the updated learning model, but this is not restrictive. (1) System Configuration FIG. 1 is a diagram showing a configuration example of the model update / deployment system 1 of this embodiment. That is, the model update / deployment system 1 is an example of the model inference device of the present invention. This model update / deployment system 1 can be configured by a general computer (such as a PC), and realizes the processing described later by executing a software program, for example.

[0020] The model update and deployment system 1 is composed of an input / output unit 11, a communication unit 12, a display unit 20, a control unit 30, a storage unit 40, and a bus or the like. Here, the input / output unit 11 receives user operations and outputs information to the user. That is, the input / output unit 11 only needs to have an input function and an output function. Therefore, the input / output unit 10 can be realized by individual devices, an input device having an input function and an output device having an output function.

[0021] For example, as the input function, the following inputs are received. (A) Specification of the model to be updated (B) Setting items of the model used when updating the model called the updated learning model (C) Input of items on a graphical user interface (GUI), which is a type of output device (D) Selection of the model to be updated and the updated learning model (E) Input of observation data measured in the environment where the model to be updated is used Also, as an example of the output function, output of learning of the updated learning model and inference results using the updated learning model can be mentioned. And the input / output unit 11 can be realized by, for example, a keyboard, a mouse, a display, a printer, etc.

[0022] Also, the communication unit 12 includes a communication interface device for an external communication network of the model update and deployment system 1 and communicates with an external server, a manufacturing device, or the like. For this purpose, the communication unit 12 may acquire and refer to monitor data, manufacturing process information, etc. from an external server, a manufacturing device, or the like according to the control from the control unit 30.

[0023] In addition, the model update and deployment system 1 may be configured to form a GUI on the screen of the display unit 20 and display various types of information. Note that the display unit 20 may be realized as an output device or may be configured separately. When configured separately, the input / output unit 11 will have the function of connecting to a terminal device used by the user. In this case, the display unit 20 can be omitted. In this case, it is possible to configure the input / output unit 11 and the communication unit 12 as one part.

[0024] The control unit 30 is composed of known elements such as a processor such as a CPU and memories such as RAM and ROM. That is, the control unit 30 realizes the processing of each part described later according to the program stored in the memory. That is, the control unit 30 has an input data processing unit 31, an edge processing unit 32, a feature amount data creation unit 33, an updated learning model construction unit 34, a model display processing unit 35, an updated learning model inference unit 36, and an updated learning model deployment unit 37, and performs the main processing of this embodiment. (2) Display content Next, with reference to FIGS. 2 to 6, the display content in this embodiment will be described. In this embodiment, a GUI screen example displayed on the display unit 20 will be described, but the same output and display are possible via the input / output unit 11 and the communication unit 12.

[0025] The display unit 20 is mainly composed of a model display unit 21, a parameter display unit 22, an updated learning model selection unit 23, an updated learning model display unit 24, and an updated learning model parameter display unit 25. However, these may be realized as functions of the control unit 30. That is, the GUI screen of the display unit 20 can be created by the control unit 30 according to the program that realizes each part and displayed on the display unit 20.

[0026] First, FIG. 2 is a diagram showing an example of the model to be updated and the updated model in the model display unit 21. The model display unit 21 displays a tabular screen called a model management screen (2-1). This model management screen (2-1) is a tabular screen composed of header information (2-2) regarding the model and values (2-3) for the header. The header information (2-2) mainly consists of a header (2-4) representing the row number, a label (2-5) for identifying the model, the number of data items (2-6) used in the model, the creation date and time (2-7) of the model, and a detailed setting link (2-8). The header information (2-2) is only an example and is not limited to this.

[0027] Also, the header (2-4) representing the row number represents the row number when displaying each saved model. The label (2-5) indicates information about each model, which is a label for identifying the models stored in the model storage unit 41 and the parameter storage unit 44. The number of data items (2-6) indicates the number of nodes in the network structure described in 11-1 of the model storage unit 41 and the number of data items of the headers (14-3 and 14-4) described in the parameter storage unit 44.

[0028] The creation date and time (2-7) indicates the date and time when the model was created. On the detailed setting link (2-8), a button for transitioning to a screen for displaying the detailed settings of each model is displayed. On the model detailed setting screen (2-9) of each model, the network structure (2-11) of the target model is displayed. The network structure (2-11) to be displayed shows the information stored in the model storage unit 41. The label (2-10) in the network structure corresponds to the label (2-5) of the model management screen (2-1). In the network structure (2-11), a network structure composed of nodes (2-12) representing each data item and edge information (2-13) representing the relationship between data items is displayed.

[0029] Further, the node (2-12) corresponds to 11-3 of the model storage unit 41, and the edge information corresponds to 11-4 of the model storage unit 41. The parameter display button (2-14) is a button that transitions to the parameter display unit 22, which is a screen for displaying the parameters of the target model.

[0030] Next, the parameter display unit 22 will be described with reference to FIG. 3. FIG. 3 is a diagram showing a display example of parameter information on the parameter display unit in the present embodiment. The parameter display unit 22 displays the parameters of the model to be updated and the updated model.

[0031] In FIG. 3, on the parameter display unit 22, when the parameter display button (2-14) on the model detail setting screen (2-9) of the model management screen (2-1) is pressed, the screen transitions. Also, the parameter display unit 22 displays the label (3-1) representing the target model and the parameter information (edge information in this example) in a table format (3-2).

[0032] Here, the table format (3-2) displays the edge information (2-13) between the data items of the model display unit 21 in a table format. The first row (3-3) and the first column (3-4) of the table are the data items used in the target model, corresponding to 14-3 and 14-4 of the parameter storage unit 44. The edge information (3-5) between the data items indicates the information of 14-5 stored in the parameter storage unit 44.

[0033] Next, the updated learning model selection unit 23 will be described with reference to FIG. 4. FIG. 4 is a diagram showing a display example of information for managing the updated learning model on the updated learning model selection unit 23 in the present embodiment. Here, the updated learning model selection unit 23 refers to the model to be updated as the updated learning model and displays a screen for selecting a plurality of updated learning models.

[0034] The update learning model selection unit 23 displays an update learning model management screen (4-1) described in tabular form. Here, the update learning model management screen (4-1) shows a table composed of header information (4-2) and values (4-3) corresponding to the header. The header information (4-2) includes the row number (4-4) of each update learning model, a label (4-5) for identifying the update learning model, the number of data items (4-6) handled by the update learning model, the creation date and time (4-7) when the update learning model was created, and detailed settings (4-8) that describe a button for transitioning to a screen where detailed information about the update learning model is described.

[0035] This header information (4-2) is merely an example and is not limited to this. The row number (4-4), which is a header representing the number in the row direction, represents the row number when displaying each stored update learning model. The label (4-5) indicates the label for identifying the model stored in the model storage unit 41 and the update learning model parameter storage unit 45. The number of data items (4-6) describes the number of nodes in the input layer of the graphical neural network structure described in 12-1 of the model storage unit 41. The creation date and time (4-7) indicates the date and time when the update learning model was created. The detailed settings (4-8) describe a button for transitioning to a screen where the detailed settings of each model are displayed.

[0036] Next, the update learning model display unit 24 will be described with reference to FIG. 5. FIG. 5 is a diagram showing an example of the display in the structural form of the update learning model in the update learning model display unit 24 in the present embodiment.

[0037] The updated learning model display unit 24 displays the updated learning model. Specifically, the updated learning model display unit 24 displays the label (5-1) and the structural format (5-2) of the target updated learning model stored in the updated learning model storage unit 42. Here, the label (5-1) corresponds to the label (12-1) of the updated learning model storage unit. Also, the structural format (5-2) corresponds to the structural format (12-2) of the updated learning model storage unit 42. The structural format (5-2) is composed of the input layer (5-3) of the updated learning model, the output layer (5-4) of the updated learning model, the intermediate layer (5-5) of the updated learning model, the layer number (5-6) of the intermediate layer of the updated learning model, and the edge information (5-7) representing the connection between each node of the updated learning model.

[0038] Also, the input layer (5-3) of the updated learning model corresponds to 12-3 of the updated learning model storage unit 42. The output layer (5-4) of the updated learning model corresponds to 12-4 of the updated learning model storage unit 42. The intermediate layer (5-5) of the updated learning model corresponds to 12-5 of the updated learning model storage unit 42. The layer number (5-6) of the intermediate layer of the updated learning model corresponds to 12-6 of the updated learning model storage unit 42. The edge information (5-7) representing the connection between the nodes of the updated learning model corresponds to 12-7 of the updated learning model storage unit 42. The parameter display button (5-8) is a button for transitioning to the updated learning model parameter display unit 25, which is a screen having a table describing the weights between each node, which are the parameters of the updated learning model.

[0039] Next, the updated learning model parameter display unit 25 will be described with reference to FIG. 6. FIG. 6 is a diagram showing a display example of displaying the parameter information of the updated learning model in the updated learning model parameter display unit 25 in the present embodiment.

[0040] The updated learning model parameter display unit 25 displays the parameters within the updated learning model. Specifically, the updated learning model parameter display unit 25 is composed of a label (6-1) for identifying the updated learning model and a screen (6-2) that stores the parameters of the updated learning model in tabular form. Here, the label (6-1) is the label of the target model, corresponding to the label (5-1) of the model displayed when the parameter display button (5-8) is pressed by the updated learning model display unit 24, and corresponding to the label (15-1) stored in the updated learning model parameter storage unit 45. The screen (6-2) in tabular form is composed of two tables, namely, a table (6-3) of layer numbers and the number of nodes within the layer, and a table (6-4) having layer numbers, weight numbers, and the values of the weights corresponding to the weight numbers. The table (6-3) of layer numbers and the number of nodes within the layer has the layer number (6-5) and the number of nodes within the layer (6-6) as headers, and the respective values are described in 6-10. The table (6-4) having layer numbers, weight numbers, and the values of the weights corresponding to the weight numbers has the layer number (6-7), the weight number (6-8), and the weight (6-9) as headers, and the respective values are described in 6-11. (3) Control unit 30 Next, the control unit 30 will be described. As shown in FIG. 1, the control unit 30 mainly consists of an input data processing unit 31, an input data processing unit 31, an edge processing unit 32, a feature data creation unit 33, an updated learning model construction unit 34, a model display processing unit 35, an updated learning model inference unit 36, and an updated learning model deployment unit 37. These are each functional blocks, and their processing can also be executed according to the program as described above. This will be described later.

[0041] Here, the updated learning model targeted in this embodiment has three phases called learning, inference, and deployment. Learning is the phase of learning the optimal parameters of the updated learning model using a group of past models and a group of data. Inference is the phase of inferring and outputting information that requires updating from the input model and data using the updated learning model whose parameters have been adjusted by learning. Deployment is the phase of managing the updated learning model created in the past and applying it to other models.

[0042] Next, an example of the processing flow of the control unit 30 in the learning phase will be described with reference to FIG. 7. FIG. 7 is a flowchart showing the processing flow of learning of the updated learning model in the present embodiment.

[0043] First, the input data processing unit 31 performs processing on the target model and its data input to the model update and deployment system 1 (step S1). Here, the input data processing unit 31 receives, as teacher data for updating the updated learning model, the model before update (pre-update model), the data used for constructing the pre-update model (pre-update model construction data), the data actually obtained by operating the pre-update model (post-update model data), and the input of the model updated using the post-update model data (post-update model).

[0044] Here, as the number of inputs, it is possible to input a plurality of pre-update models, pre-update model construction data, post-update models, and post-update model construction data. A plurality of models will be hereinafter referred to as a model group, and a plurality of data will be referred to as a data group. Note that the data processing in this step refers to various processes such as statistical data processing such as pre-processing of data, processing and complementing of missing values, processing of outliers, and error checking.

[0045] Next, the edge processing unit 32 acquires edge information, which is graph information, from the pre-update model group and the post-update model group (step S2). Next, the feature data creation unit 33 creates feature data using the pre-update model construction data and the post-update model construction data (step S3). Here, the feature data indicates features related to the relevance between the pre-update model and the post-update model. As an example of this, the difference between the pre-update model and the post-update model or the difference between the pre-update model construction data and the post-update model construction data can be used.

[0046] Next, the updated learning model construction unit 34 performs learning of the updated learning model (step S4). For this purpose, the updated learning model construction unit 34 is based on a graph neural network. Then, the updated learning model construction unit 34 takes the edge information of the pre-update model group processed in step S2 as input, uses the feature data created in step S3, and performs learning of a model using the edge information of the post-update model group processed in step S2 as teacher data. Here, the model to be learned is called the updated learning model. Regarding step S4, a detailed flow will be described with reference to FIG. 8.

[0047] Finally, the model display processing unit 35 outputs the updated learning model for which learning has been completed on the screen (step S5). Note that the updated learning model may be output by the updated learning model display unit 24 of the display unit 20.

[0048] Next, a detailed flow of step S4 will be described with reference to FIG. 8. FIG. 8 is a flowchart showing the detailed processing flow of step S4 (learning of the updated learning model) in the present embodiment. First, the updated learning model construction unit 34 defines a model derived from a graph neural network as the updated learning model (step S6). Here, Graph Convolutional Neural Network (GCN) is taken as an example.

[0049] Next, the updated learning model construction unit 34 receives, as input to the updated learning model, the pre-update model construction data created in step S3, the feature data created from the post-update model construction data, and the pre-update model group (step S7). This input means explanatory variables in the field of machine learning.

[0050] Next, the updated learning model construction unit 34 uses the features input in step S7 and the post-update model group corresponding to the pre-update model group as teacher data, and divides the pairs of the pre-update model group, the post-update model group, and the features into a training data group and a verification data group, respectively (step S8).

[0051] Next, the updated learning model construction unit 34 acquires the pre-update model, the post-update model, and the feature data from the training data group (step S9). The acquisition method here can be arbitrary, whether random, ascending order, descending order, or a method specified by the user, and the number of acquisitions is not limited either.

[0052] Next, the updated learning model construction unit 34 inputs the acquired pre-update model and the corresponding feature data into the updated learning model, and creates features from the updated learning model. Then, the updated learning model construction unit 34 estimates the edge appearance probability between each node of the post-update model, which is the target teacher data, from the features (step S10).

[0053] Next, the updated learning model construction unit 34 estimates binary information by determining the existence of an edge according to the estimated edge appearance probability. For this purpose, the updated learning model construction unit 34 compares the value of the edge with the threshold set in advance by the updated learning model parameter display unit 25. If the threshold is exceeded, the updated learning model construction unit 34 determines that the edge exists, and sets the edge information (binary information) between the target nodes as 1 as the edge information. If the threshold is not exceeded, the updated learning model construction unit 34 determines that the edge does not exist, and sets the edge information (binary information) between the target nodes as 0.

[0054] These edge information are referred to as estimated edge information. Note that the threshold can be set by the updated learning model parameter display unit 25, or the updated learning model construction unit 34 can automatically set an appropriate threshold, and the threshold setting method is not limited to these two methods.

[0055] Next, the updated learning model construction unit 34 compares the estimated edge information calculated in step S11 with the edge information processed from the updated model in step S2. Then, the updated learning model construction unit 34 feeds back the difference information indicating the difference to the updated learning model (step S12). Here, as a method of feeding back to the updated learning model, a method can be adopted in which the difference information between the estimated edge information and the edge information is used to calculate a loss value using, for example, cross-entropy, and the value is fed back to the updated learning model by the error backpropagation method. However, this method is only an example and is not limited to this method.

[0056] Next, the updated learning model construction unit 34 repeatedly performs the operations in steps S8 to S10 for each of the training data groups extracted in step S9. By doing this, the updated learning model construction unit 34 performs the operations until the value of the set estimated error function of the model becomes small enough to satisfy a predetermined condition (step S13). Regarding the number of learning times, it is possible to set it in advance in the updated learning model parameter display unit 25 or to automatically set it in the updated learning model construction unit 34, and it is not limited to these two methods. Also, satisfying the predetermined condition includes that the value of the estimated error function is less than or equal to a threshold value or that the number of repetitions of steps S8 to S11 is equal to or more than a certain number of times.

[0057] Next, the updated learning model construction unit 34 repeatedly performs the operations in steps S8 to S10 for each of the verification data extracted in step S9 using the updated learning model learned in steps S6 to S13. By doing this, the updated learning model construction unit 34 outputs the prediction accuracy for the estimation result (step S14).

[0058] Here, the display of the estimated result may be the updated learning model selection unit 23 and the updated learning model display unit 24. Note that the updated learning model construction unit 34 shall save the prediction accuracy for the estimated result in the updated learning model storage unit 42. Further, the updated learning model construction unit 34 shall save information such as parameters regarding the updated learning model for which learning has been completed in the updated learning model parameter storage unit 45. Furthermore, the updated learning model construction unit 34 shall save the pre-update model group, post-update model group, created feature amounts, and learning conditions used for the updated learning model in the updated learning model storage unit 42. This concludes the description of FIG. 8.

[0059] Next, FIG. 9 is a flowchart showing an example of the processing flow of the updated learning model inference unit 36 in the inference phase.

[0060] First, the updated learning model inference unit 36 performs processing on the model to be targeted (update target model) input to the model update / deployment system 1, the data used when constructing the update target model (update target model construction data), and the data obtained when operating the model (data obtained during operation) (step S15). Note that the data processing here refers to statistical data processing such as preprocessing of data, processing and complementing of missing values, and processing of outliers, as well as various processes such as error checking, similar to step S1 in the learning phase.

[0061] Next, the updated learning model inference unit 36 acquires the edge information of the input model (step S16). In this step, the same processing as in step S2 is performed.

[0062] Next, the updated learning model inference unit 36 creates feature amount data using the update target model construction data and the data obtained during operation (step S17). The creation of this feature amount can be realized in the same manner as in step S3 of the learning phase.

[0063] Next, the updated learning model inference unit 36 estimates the updated model using the updated learning model obtained in the learning phase (step S18). As an estimation method, the input obtained in step S16 is used as the edge information and feature data of the model to be updated. Then, the updated learning model inference unit 36 performs the same processing as in steps S10 and S11 of the learning phase to estimate the edge information of the updated model.

[0064] Next, the updated learning model inference unit 36 sets the model having the estimated edge information as the updated model and outputs the model (step S19). At this time, it is desirable to display it on the display unit 20 using the updated learning model display unit 24.

[0065] Next, FIG. 10 is a flowchart showing an example of the processing flow of the control unit 30 in the deployment phase. This flowchart is executed by the updated learning model deployment unit 37.

[0066] First, the updated learning model deployment unit 37 inputs the model to be updated and checks the number of its data items (step S20). That is, the number of data items is measured.

[0067] Next, the updated learning model deployment unit 37 compares the confirmed number of data items with the number of updated learning models (step S21). As a result, if they are the same, the process proceeds to step S22. If they are not the same, the process proceeds to step S23.

[0068] Next, if they are the same, the updated learning model deployment unit 37 determines to use the selected updated learning model (step S22).

[0069] Also, when the numbers are not the same, the updated learning model deployment unit 37 shall use only the parameters of the intermediate layer of the selected updated learning model. Then, the updated learning model deployment unit 37 fine-tunes the parameters between the input layer and the output layer of the selected updated learning model and the intermediate layer with teacher data having the same number of data items as the model to be updated. The updated learning model deployment unit 37 saves this result as a new updated learning model in the storage unit 40 (S22).

[0070] Next, the updated learning model deployment unit 37 shifts to the inference phase (step S24). As a result, in the inference phase, the model to be updated can be updated using the updated learning model determined in step S21 or the updated learning model saved in step S22. (4) Storage unit 40 Next, the storage unit 40 will be described. The storage unit 40 mainly consists of a model storage unit 41, an updated learning model storage unit 42, an input data storage unit 43, a parameter storage unit 44, and an updated learning model parameter storage unit 45. Note that these are classified for convenience according to the information to be stored, and the number thereof is not limited, such as being realized in one housing as a configuration. Next, various information stored in the storage unit 40 will be described.

[0071] First, the content stored in the model storage unit 41 will be described. FIG. 11 is a diagram showing a configuration example of data in the network structure of the model in the present embodiment. Specifically, this data is the pre-update model, the post-update model which are teacher data, and the model estimated by the updated learning model. Here, in the present embodiment, the pre-update model, the post-update model, and the model estimated by the updated learning model are assumed to be Bayesian networks, and their formats are shown in FIG. 11. Note that the assumed model is a Bayesian network model, but it is not limited thereto.

[0072] The data shown in FIG. 11 includes label information (11-1) for identifying a model and model information (11-2) represented by a network. Here, in this figure, using the representation format in the network structure, for each data item, the node (11-3) is indicated by a circle, and the edge information (11-4) between the nodes is indicated by an arrow. In FIG. 11, when an edge is connected from the temperature 1 in the first row to the temperature 2, in the network structure, an edge starting from the node of temperature 1 and ending at the node of temperature 2 is described between the node of temperature 1 and the node of temperature 2.

[0073] Next, FIG. 12 is a diagram showing a configuration example of data in the structure format of the update learning model in the present embodiment. That is, in FIG. 12, the update learning model stored in the update learning model storage unit 42 is shown. In the present embodiment, it is assumed that the update learning model is a graph neural network, and its format is shown in FIG. 12.

[0074] In FIG. 12, a label (12-1) for identifying the update learning model and a structure format (12-2) in which the model is represented in a structure format are described. In this structure format, the update learning model described in a table format is represented in a graph format. Here, in this figure, the nodes (input layer 12-3, output layer 12-4) existing in the input layer and the output layer of the model represented by the graph neural network are each represented by a black circle, and the nodes in the intermediate layer are each represented by a white circle (12-6). Here, the layer number represents the number of the intermediate layer (12-6), and the nodes in the layer represent the number of nodes represented by the white circles of each layer number. Also, the weight number (12-7) for each layer number is an edge between the nodes of the input layer and the intermediate layer, between the nodes of the intermediate layers, and between the nodes of the intermediate layer and the output layer, and there are weights for each of them.

[0075] Next, FIG. 13 is a diagram showing a configuration example of data at the time of constructing the pre-update model, data at the time of constructing the post-update model, or data acquired during operation in the present embodiment. That is, in FIG. 13, the pre-update model construction data, the post-update model construction data, which are teacher data stored in the input data storage unit 43, and the data acquired during operation are represented.

[0076] In FIG. 13, label information (13-1) for identifying data and data information (13-2) described in tabular form are described. In the data information (13-2) in tabular form, the first row is the header information of each data item. And the first column is the acquisition time (13-3), and the columns after the second column are data items (13-4). Note that the information and arrangement of the headers are not limited to the example shown in FIG. 13. Also, from the second row onwards, information (13-5) about the items described in each header is described. Further, in the present embodiment, each data item for each acquisition time is described, but data described at different granularities such as information about each data item for each product ID etc. may also be used.

[0077] Next, FIG. 14 is a diagram showing an example of the data configuration of edge information which is a parameter of the model in the present embodiment. That is, FIG. 14 shows the edge information stored in the parameter storage unit 44. Here, in the present embodiment, it is assumed that the pre-update model, the post-update model, and the model estimated by the updated learning model are Bayesian networks, and FIG. 14 shows a format defining edge information which is an example of the parameters thereof.

[0078] In this FIG. 14, the model is expressed in tabular form. In this tabular form, it is described as a label (14-1) for identifying the model and a table (14-2) including the edge information of the model. In the data information shown in this table (14-2), the first row is the data item (14-3) constituting the target model, and the first column is also the data item (14-4) constituting the target model.

[0079] The edge information (14-5) between the data items described in this first row and the first column is described in the second row and the columns after the second column. Here, for the data item in the first column, the data items in the first row are each called a parent node. When an edge is connected as a means for describing the edge information, 1 is described, and when the edge is not connected, 0 is described.

[0080] Next, FIG. 15 is a diagram showing an example of the data configuration of the parameter information of the update learning model in the present embodiment. That is, in FIG. 15, it is a diagram showing the parameters of the update learning model stored in the update learning model parameter storage unit 45. Here, in the present embodiment, it is assumed that the update learning model is a graph neural network, and its format is described in FIG. 15.

[0081] In this FIG. 15, a diagram representing these models in tabular form is shown. In this tabular form, a label (15-1) for identifying the update learning model and a graph neural (15-2) are described. The graph neural (15-2) is composed of a table (15-3) representing the layer number on the graph neural network and the number of nodes in the layer, and a table (15-4) identifying the layer number, the number for identifying their weights, and holding the weights.

[0082] Also, in the table (15-3), in the first row, the layer number (15-5) for identifying the layers in the graph neural network and the number of nodes in the layer (15-6) used for each layer number are used as headers, and from the second row onwards, the layer number and the values of the number of nodes in the layer (15-10) are included. In the table (15-4), in the first row, the layer number (15-7), the weight number (15-8) for identifying the weights in the layer, and the weight (15-9) are used as headers, and from the second row onwards, the layer number, the weight number, and the values of the weights (15-11) are described. (5) Implementation Example Next, an implementation example in the present embodiment will be described. FIG. 16 is a diagram showing the implementation example in the present embodiment. This implementation example is an example of realizing the model update and deployment system 1 as a cloud (server). In this implementation example, the model update and deployment system 1 is connected to the business system A, the business system B, the mobile terminal 101, and the terminal device 102 via the network 70. Hereinafter, each of these devices will be described.

[0083] First, the model update and deployment system 1 can be implemented on a server, which is a type of computer. The model update and deployment system 1 includes a processing device 301, a communication unit 12, an input / output unit 11, a memory 302, and a storage unit 40, which are connected to each other via a communication path such as a bus.

[0084] Here, the processing device can be implemented by a processor such as a CPU, and executes processing according to each program described later. This processing is the processing of the control unit 30 described above.

[0085] Also, the communication unit 12 and the input / output unit 11 are the same as those shown in FIG. 1. The communication unit 12 is connected to a network 70 such as the Internet. The input / output unit 11 is connected to user terminals 100-1 and 100-2 used by the user. These user terminals 100-1 and 100-2 can be implemented by computers such as smartphones, tablets, and PCs. Note that the user terminals 100-1 and 100-2 may be connected to the model update and deployment system 1 via the network 70 or a dedicated network.

[0086] In addition, various programs stored in a storage medium such as the storage unit 40 and information used for processing are expanded in the memory 302. These programs include an input data processing program 311, an edge processing program 321, a feature data creation program 331, an updated learning model construction program 341, a model display processing program 351, an updated learning model inference program 361, and an updated learning model deployment program 371. These are programs for executing the same processing as the input data processing unit 31, the edge processing unit 32, the feature data creation unit 33, the updated learning model construction unit 34, the model display processing unit 35, the updated learning model inference unit 36, and the updated learning model deployment unit 37, respectively. That is, the processing of the control unit 30 is executed by the processing device 301 and the memory 302. In this implementation example, various processes are realized based on software called programs, but dedicated hardware or an FPGA (field-programmable gate array) may execute the processing of each part shown in FIG. 1.

[0087] Next, the storage unit 40 is as shown in FIG. 1 and stores the various types of information and programs described above. Note that the storage unit 40 may be implemented in a housing separate from the model update and deployment system 1.

[0088] Next, the business system A and the business system B are systems that support various operations by using the updated learning model and the model itself constructed in the present embodiment. In the present embodiment, a model related to the operations in the business system A can be deployed (used) in the business system B, which is a different business. For example, the model of "maintenance of power generation facilities" in the business system A can be deployed in the "software management operation" in the business system B.

[0089] Here, the business system A includes a server 51 including a storage device 511 and a processing device 512, and a terminal device 52. The business system B includes a server 61 including a storage device 611 and a processing device 612, and a terminal device 62. That is, these can be realized by a so-called computer system.

[0090] Also, the mobile terminal 101 and the terminal device 102 are used for the various operations described above. For example, the mobile terminal 101 and the terminal device 102 receive various processing results using the deployed model and display them.

[0091] Although specifically described based on the embodiments of the present invention above, it goes without saying that the present invention is not limited to the above-described embodiments and can be variously modified without departing from the gist thereof. In particular, the update of the model may be executed by a device other than the model update and deployment system 1.

Explanation of Reference Numerals

[0092] 1 ··· Model update and deployment system 11 ··· Input / output unit 12 ··· Communication unit 20 ··· Display unit 21 ··· Model display unit 22 ··· Parameter display unit 23 ··· Update learning model selection section 24 ··· Update learning model display section 25 ··· Update learning model parameter display section 30 ··· Control section 31 ··· Input data processing section 32 ··· Edge processing section 33 ··· Feature data creation section 34 ··· Update learning model construction section 35 ··· Model display processing section 36 ··· Update learning model inference section 37 ··· Update learning model deployment section 40 ··· Memory section 41 ··· Model memory section 42 ··· Update learning model memory section 43 ··· Input data memory section 44 ··· Parameter memory section 45 ··· Update learning model parameter memory section

Claims

1. In a model inference device that infers a model in machine learning, a storage unit that stores a pre-update model, which is a model before update, and an updated model; a feature quantity data creation unit that creates difference information indicating the difference between the pre-update model and the updated model as a feature quantity indicating the relevance between the pre-update model and the updated model in the data used when updating the model; an updated learning model construction unit that constructs an updated learning model, which is a GCN that learns the update method of the update by performing supervised learning using the pre-update model, the data used when constructing the pre-update model, and the data used when constructing the updated model, with the updated model as teacher data; an updated learning model inference unit that infers a predetermined model based on the updated learning model; the feature quantity data creation unit creates, as the difference information, the difference between the edge information, which is the estimated parameter information, and the edge information given as teacher data during the learning; the updated learning model construction unit constructs the updated learning model by feeding back the difference information to the updated learning model, the model inference device.

2. In the model inference device according to Claim 1, further comprising an updated learning model deployment unit that deploys a model in a predetermined operation based on the updated learning model.

3. In a model inference method executed by a model inference device that infers a model in machine learning, store a pre-update model, which is a model before update, and an updated model in a storage unit; a feature quantity data creation unit creates difference information indicating the difference between the pre-update model and the updated model as a feature quantity indicating the relevance between the pre-update model and the updated model in the data used when updating the model; an updated learning model construction unit constructs an updated learning model, which is a GCN that learns the update method of the update by performing supervised learning using the pre-update model, the data used when constructing the pre-update model, and the data used when constructing the updated model, with the updated model as teacher data; an updated learning model inference unit infers a predetermined model based on the updated learning model; When the feature amount data creation unit creates the difference between the edge information, which is the estimated parameter information, and the edge information given as teacher data during the learning as the difference information. A model inference method in which the updated learning model construction unit constructs the updated learning model by feeding back the difference information to the updated learning model.

4. In the model inference method according to claim 3, Furthermore, a model inference method in which an updated learning model deployment unit deploys a model in a predetermined operation based on the updated learning model.

5. A computer that is a model inference device for inferring a model in machine learning, A storage unit that stores a pre-update model, which is a model before update, and an updated model; A feature amount data creation unit that creates difference information indicating the difference between the pre-update model and the updated model as a feature amount indicating the relevance between the pre-update model and the updated model in the data used during the update of the model; An updated learning model construction unit that constructs an updated learning model, which is a GCN that learns the update method of the update by performing supervised learning using the pre-update model, the data used when constructing the pre-update model, and the data used when constructing the updated model, with the updated model as teacher data; Functioning as an updated learning model inference unit that infers a predetermined model based on the updated learning model; The feature amount data creation unit creates the difference between the edge information, which is the estimated parameter information, and the edge information given as teacher data during the learning as the difference information. A program in which the updated learning model construction unit constructs the updated learning model by feeding back the difference information to the updated learning model.

6. In the program according to claim 5, Furthermore, a program that causes the computer, which is the model inference device, to function as an updated learning model deployment unit that deploys a model in a predetermined operation based on the updated learning model.

Images