Machine learning devices, machine learning systems, machine learning methods, and programs

Abstract

The machine learning device (2) includes a data model construction unit (11) that constructs a data model having a data structure that conforms to the structure of the system to be learned; a metadata generation unit (13) that generates metadata of operational data acquired from the system based on the data model; a determination unit (14) that determines the similarity between the metadata of the training data used to generate an existing model, which is an existing trained model, and the metadata of the operational data acquired from the system, and determines from among a plurality of existing models which existing model will be used as an applicable model for learning about the system based on the similarity; and a learning unit (15) that generates a trained model about the system by learning with the applicable model applied.

Description

Technical Field

[0001] The present disclosure relates to a machine learning device, a machine learning system, a machine learning method, and a program.

Background Art

[0002] As one of the machine learning techniques, there is known a technique of applying a learned model for a certain domain to learning for another domain to generate a new model. According to such a technique, since the already learned knowledge can be utilized, it becomes possible to generate a new model by inputting a small amount of learning data. However, when there are a plurality of learned models for different domains, there is a problem that it is difficult to determine which learned model should be applied in order to generate a highly accurate model.

[0003] Patent Document 1 discloses a model selection device that selects a learned model optimal for transfer learning from among a plurality of learned models with different domains, regarding transfer learning which is one of the techniques for generating a new model by applying an existing learned model. In Patent Document 1, the learned model is a model that identifies an identification target by using data indicating the identification target as an input. The model selection device according to Patent Document 1 calculates the probability distribution of input data for each of the learned models selected as candidates for application to transfer learning, and calculates an information matrix based on the input data and the probability distribution. The model selection device according to Patent Document 1 inputs the information matrix, calculates a domain score indicating the variance of the result of identifying the identification target, and selects a learned model optimal for application to transfer learning from among the learned models selected as candidates based on the domain score.

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Summary of the Invention

[0005] The method disclosed in Patent Document 1, which involves selecting a pre-trained model from among multiple pre-trained models to be applied to the generation of a new model, can be described as a so-called data-driven method. Therefore, the technology disclosed in Patent Document 1 has the problem that it is difficult to make the learning process using existing pre-trained models understandable to users who do not have specialized knowledge of machine learning or data science.

[0006] This disclosure is made in view of the above, and aims to provide a machine learning device that enables learning using existing pre-trained models through a process that is easy for users to understand. [Means for solving the problem]

[0007] To solve the above-mentioned problems and achieve the objective, the machine learning device according to this disclosure comprises: a data model construction unit that constructs a data model having a data structure in accordance with the structure of the system to be learned; a metadata generation unit that generates metadata of operational data acquired from the system based on the data model; a determination unit that determines the similarity between the metadata of the training data used to generate an existing model, which is an existing trained model, and the metadata of the operational data acquired from the system, and determines from among a plurality of existing models to be used as an applicable model for learning about the system based on the similarity; and a learning unit that generates a trained model about the system by learning with the applicable model applied. [Effects of the Invention]

[0008] The machine learning device described herein has the effect of enabling learning using existing pre-trained models through a process that is easy for the user to understand. [Brief explanation of the drawing]

[0009] [Figure 1] A diagram showing an example configuration of the machine learning system according to Embodiment 1. [Figure 2] This figure shows an example of a data model constructed by the data model construction unit of the machine learning device according to Embodiment 1. [Figure 3] This figure shows an example of operational data collected by the machine learning device according to Embodiment 1. [Figure 4] This figure shows an example of operational metadata generated by the metadata generation unit of the machine learning device according to Embodiment 1. [Figure 5] This figure shows an example of a graph network generated by the graph network generation unit of the machine learning device according to Embodiment 1. [Figure 6] Flowchart showing an example of the operation procedure of the machine learning device according to Embodiment 1 [Figure 7] A diagram showing an example configuration of the machine learning system according to Embodiment 2. [Figure 8] Flowchart showing an example of the operation procedure of the machine learning device according to Embodiment 2 [Figure 9] A diagram showing an example of the configuration of the control circuit according to Embodiment 1 or 2. [Modes for carrying out the invention]

[0010] The machine learning apparatus, machine learning system, machine learning method, and program according to the embodiment will be described in detail below with reference to the drawings.

[0011] Embodiment 1. Figure 1 is a diagram showing an example configuration of machine learning system 1 according to Embodiment 1. Machine learning system 1 is a system that performs machine learning using machine learning device 2. Machine learning device 2 receives operational data of the system to be studied and performs learning targeting that system. Hereinafter, the system that is the target of learning by machine learning device 2 will be referred to as the target system. Machine learning device 2 receives operational data, which is data related to the operation of the target system. One example of a target system is a factory automation (FA) system installed in a factory. In the following description, the target system will be assumed to be a factory FA system.

[0012] The machine learning system 1 comprises a machine learning device 2, a data acquisition device 3, and a GUI (Graphical User Interface) 4, which is an input / output device. The data acquisition device 3 and GUI 4 are installed in a factory. The machine learning device 2 is implemented, for example, by a cloud server. The cloud server is a server built in a cloud environment that includes computer resources provided on a cloud service platform. The machine learning device 2 is connected to the data acquisition device 3 and GUI 4 via a network. In Embodiment 1, the network is, for example, a WAN (Wide Area Network) such as the Internet, but it may also be a LAN (Local Area Network).

[0013] The data acquisition device 3 acquires operational data, which is data about the operation of each of the multiple pieces of equipment installed in the factory. The data acquisition device 3 sends the acquired operational data to the machine learning device 2. The machine learning device 2 sends the information generated for presentation to the user to the GUI 4. The GUI 4 displays the information received from the machine learning device 2. Here, the user is a person involved in manufacturing in the factory, whether a worker or a manager. The machine learning device 2 is connected to the data acquisition device 3 and GUI 4 in each of the multiple factories. Figure 1 shows one of the multiple factories.

[0014] The machine learning device 2 includes a data model construction unit 11, a storage unit 12, a metadata generation unit 13, a determination unit 14, a learning unit 15, and a graph network generation unit 16.

[0015] The machine learning device 2 receives operation data from the data acquisition device 3 and stores the received operation data in the storage unit 12. The operation data is accumulated in an operation database within the storage unit 12. In FIG. 1 and the following description, the operation database is denoted as the "operation DB". The machine learning device 2 collects operation data by storing the operation data acquired at any time in the storage unit 12.

[0016] The data model construction unit 11 constructs a data model having a data structure conforming to the structure of the target system. The data model constructed by the data model construction unit 11 represents the structure of the target system by defining the relationships between the components in the plurality of components constituting the target system, and is a so-called semantic data model. In the data model constructed by the data model construction unit 11, the association between the components and the data acquired as operation data is also defined.

[0017] As an example, a standard data model used in the field to which the target system belongs is input to the data model construction unit 11. The data model construction unit 11 constructs a data model for the target system by supplementing the standard data model according to the configuration of the target system. Note that the method of constructing the data model by the data model construction unit 11 is arbitrary. For example, the data model construction unit 11 may construct a data model for the target system by analyzing the operation data group collected by the machine learning device 2.

[0018] The metadata generation unit 13 receives the data model constructed by the data model construction unit 11. The metadata generation unit 13 reads the operation data from the storage unit 12. The metadata generation unit 13 generates metadata from the operation data acquired from the target system based on the data model constructed by the data model construction unit 11. The metadata generation unit 13 stores the generated metadata in the metadata database within the storage unit 12. In FIG. 1 and the following description, the metadata database is denoted as "meta DB". Hereinafter, the metadata of the operation data is referred to as operation metadata.

[0019] The existing learned models are stored in the model database within the storage unit 12. In FIG. 1 and the following description, the model database is denoted as "model DB". Hereinafter, the existing learned models stored in the model DB are referred to as existing models. The existing models stored in the model DB are learned models generated by the machine learning device 2 in the past. The existing models stored in the model DB may include learned models generated outside the machine learning device 2 and input into the machine learning device 2.

[0020] The metadata of the learning data used for generating the existing models is stored in the meta DB. The machine learning device 2 aggregates and manages the existing models and the metadata of the learning data used for generating the existing models in the storage unit 12.

[0021] The determination unit 14 reads the operation metadata for the operation data acquired from the target system from the meta DB. For each of the plurality of existing models stored in the model DB, the determination unit 14 reads the metadata of the learning data used for generating the existing model from the meta DB. The determination unit 14 has a function of calculating the similarity between the metadata. By such a function, the determination unit 14 obtains the similarity between the metadata of the learning data used for generating the existing model and the metadata of the operation data acquired from the target system for each of the plurality of existing models.

[0022] The determination unit 14 determines, based on the similarity obtained for each of the multiple existing models, which existing model will be used as the applicable model for learning about the target system. The determination unit 14 determines one or more existing models to be used as the applicable model.

[0023] The learning unit 15 generates a trained model for the target system by applying one or more applicable models to the learning process. The learning unit 15 then saves the generated trained model to the model database.

[0024] The graph network generation unit 16 generates a graph network that represents the relationship between the metadata of the training data used to generate the existing model applied to training in the learning unit 15 and the metadata of the operational data. The graph network generation unit 16 sends the information of the generated graph network to the GUI 4. The GUI 4 receives the information of the graph network and displays the graph network.

[0025] Next, we will describe the data model constructed by the data model construction unit 11. Figure 2 is a diagram showing an example of a data model constructed by the data model construction unit 11 of the machine learning device 2 according to Embodiment 1.

[0026] The data model shown in Figure 2 is a data model for data related to FA systems. In the example shown in Figure 2, "factory," "line," "equipment," and "sensor" are items that represent components of the FA system. In the example shown in Figure 2, the entity representing the factory is associated with the entity representing the line, which is equipment installed in the factory. The entity representing the line is associated with the entity representing the equipment installed in the line. The entity representing the equipment is associated with the entity of the sensor that performs detection on the equipment. Each entity representing a component is assigned information that represents the name of the component.

[0027] For example, suppose a robot arm A is installed on "Assembly Line A" in a factory. "Assembly Line A" is the name of one line installed in the factory. "Robot arm A" is the name of one robot arm installed on "Assembly Line A". In the data model, the relationship between "Assembly Line A" and "Robot arm A" is represented by a line connecting the "Line" entity, which holds information about "Assembly Line A", and the "Equipment" entity, which holds information about "Robot arm A".

[0028] In the example shown in Figure 2, "Motor Data" and "Temperature and Humidity Sensor Data" are items that represent data. "Motor Data" is an item that represents data detected for the motor. "Temperature and Humidity Sensor Data" is an item that represents data detected by the temperature and humidity sensor. Each entity that represents data is assigned information that represents the name of the data. The name of the data may also be the ID (IDentifier) ​​assigned to the data.

[0029] The data that makes up the data model includes data representing the status of the equipment and data representing the environment in which the equipment is installed. "Motor data" is data representing the status of the motor, which is a piece of equipment. If "temperature and humidity sensor data" is data indicating the temperature of the equipment, then this data represents the status of the equipment. If "temperature and humidity sensor data" is data indicating the temperature or humidity of the environment, then this data represents the environment. The data that makes up the data model may also include data related to the operation of the system, such as failure rate data.

[0030] For example, suppose data with ID "1001" is detected for "Motor X," which is part of "Robot Arm A." Let "Motor X" be the name of one of the motors that makes up "Robot Arm A." In the data model, the relationship between "1001" and "Motor X" is represented by an arrow pointing from the "Motor Data" entity, which holds the information for "1001," to the "Sensor" entity, which holds the information for "Motor X."

[0031] A tree-type data model is constructed by representing the relationships between components and the relationships between components and data in the target system as described above. Note that the tree-type data model is merely one example of a data model constructed by the data model construction unit 11. The data model constructed by the data model construction unit 11 can be any model based on the relationships between components and the relationships between components and data, and may be a data model other than a tree type.

[0032] Next, we will describe the operational data collected by the machine learning device 2. Figure 3 is a diagram showing an example of operational data collected by the machine learning device 2 according to Embodiment 1. In Embodiment 1, the operational data is time-series data of a certain signal. Figure 3 shows an example of operational data detected by a single sensor in a factory. In the example shown in Figure 3, the value detected by the sensor and the time at which that value was detected are associated with each other in the operational data.

[0033] Next, the operational metadata generated by the metadata generation unit 13 will be described. Figure 4 is a diagram showing an example of operational metadata generated by the metadata generation unit 13 of the machine learning device 2 according to Embodiment 1.

[0034] In the example shown in Figure 4, "ID," "Signal," "Sensor," "Equipment," and "Line" are each items of operational metadata. In other words, operational metadata includes information about operational data, specifically "ID," "Signal," "Sensor," "Equipment," and "Line." "ID" is the ID of the operational data. "Signal" represents the name of the signal that is the source of the operational data. "Sensor" represents the name of the sensor that detected the signal. "Equipment" represents the name of the equipment on which the sensor is installed. "Line" represents the name of the line on which the equipment is installed.

[0035] In Figure 4, for example, the operational metadata data with ID "1001" is data for the signal "acceleration X," which is detected by "motor X." "Motor X" is installed on "robot arm A," and "robot arm A" is installed on "assembly line A." Thus, operational metadata can be said to be data that represents the attributes of operational data. The metadata generation unit 13 acquires information about each item of operational metadata based on the relationships between components and the relationships between components and data, as shown in the data model. As a result, the metadata generation unit 13 generates operational metadata as shown in Figure 4 based on the data model constructed by the data model construction unit 11.

[0036] Note that the items shown in Figure 4 are merely examples of operational metadata items. Any items can be set in the operational metadata based on the components of the target system and the data obtained about those components.

[0037] The MetaDB stores operational metadata generated by the metadata generation unit 13. The MetaDB also stores metadata for the training data used to generate existing models. The metadata for the training data is assumed to be the same as the operational metadata exemplified in Figure 4.

[0038] Next, the determination by the determination unit 14 will be explained. The determination unit 14 reads operational metadata about the operational data acquired from the target system from the meta DB. For each of the multiple existing models stored in the model DB, the determination unit 14 reads metadata about the training data used to generate the existing model from the meta DB.

[0039] The determination unit 14 calculates the similarity between the operational metadata and the metadata of the training data. The determination unit 14 compares the information for each item in each metadata and detects items where the information matches. The determination unit 14 determines that the two sets of information match if they are substantially the same, differing only in notation. One example of a case where only the notation differs is the difference in the language used. Any method can be used to calculate the similarity.

[0040] The determination unit 14 calculates the similarity between the training data metadata for each of the multiple existing models and the operational metadata. In one example, the determination unit 14 determines that the one with the highest calculated similarity among the multiple existing models will be the existing model to be applied. Alternatively, the determination unit 14 may determine that two or more existing models will be the models to be applied. For example, the determination unit 14 may determine that the three existing models with the highest calculated similarity among the multiple existing models will be the existing models to be applied. In this case, the number of existing models to be applied may be predetermined, or it may be determined based on the similarity calculated for each of the multiple existing models.

[0041] Next, the learning process performed by the learning unit 15 will be explained. If the determination unit 14 determines that one existing model is an applicable model, the learning unit 15 generates a trained model for the target system by applying that one existing model, which is the applicable model, to the learning process. If the determination unit 14 determines that two or more existing models are applicable models, the learning unit 15 generates a trained model for the target system by applying those two or more existing models, which are the applicable models, to the learning process.

[0042] Here, we will explain three examples of the learning method used by the learning unit 15. The first and second examples are learning examples when one existing model is used as the application model. The third example is a learning example when two or more existing models are used as the application model.

[0043] The first example is transfer learning. In the first example, the learning unit 15 generates a trained model for the target system by transfer learning to which the applied model is applied. The specific method of transfer learning is arbitrary. For example, the learning unit 15 deletes the last layer of an existing neural network, which is the applied model, and adds a new layer to the neural network. The learning unit 15 then performs training only on the newly added layer. By completing training on the new layer, the learning unit 15 completes the neural network, which is the trained model.

[0044] The second example is fine-tuning. In this second example, the learning unit 15 generates a trained model for the target system by fine-tuning the applied model. The learning unit 15 uses the weights of the existing neural network, which is the applied model, as initial values ​​and fine-tunes the weights of each layer of the neural network through retraining. By completing the retraining, the learning unit 15 completes the neural network, which is the trained model.

[0045] A third example is ensemble learning. In this third example, the learning unit 15 generates a trained model for the target system through ensemble learning, applying two or more application models. The learning unit 15 performs bagging, for example, one of the ensemble learning techniques. In this case, the learning unit 15 constructs a model that combines the outputs of each of the two or more application models to derive a single output. In this way, the learning unit 15 completes the trained model.

[0046] The first to third examples above are merely examples of learning methods used by the learning unit 15. The learning unit 15 may also generate a trained model from the applied model using methods other than those described above.

[0047] Next, we will describe the graph network generated by the graph network generation unit 16. The graph network generated by the graph network generation unit 16 represents the relationship between the metadata of the training data used to generate the existing model applied to training in the learning unit 15 and the operational metadata.

[0048] Figure 5 shows an example of a graph network generated by the graph network generation unit 16 of the machine learning device 2 according to Embodiment 1. The graph network shown in Figure 5 is a part of a graph network that represents the relationship between operational metadata and training data metadata.

[0049] In Figure 5, "1001" represents the ID of a data item included in the operational data. "2001" represents the ID of a data item included in the training data. The graph network shown in Figure 5 represents the relationship between the metadata for the data item "1001" and the metadata for the data item "2001".

[0050] In Figure 5, the entities "Robot Arm A" and "Assembly Line A" are linked to entities "1001" and "2001," respectively. This indicates that "Robot Arm A" and "Assembly Line A" each have common metadata in the data for "1001" and the data for "2001."

[0051] In Figure 5, the "Motor X" entity is connected to the "1001" entity but not to the "2001" entity. This indicates that "Motor X" is metadata for the "1001" data, but not for the "2001" data.

[0052] In Figure 5, the entity for "Temperature and Humidity Sensor A" is connected to the entity for "2001" but not to the entity for "1001". This indicates that "Temperature and Humidity Sensor A" is metadata for the data for "2001" and not metadata for the data for "1001".

[0053] The graph network generation unit 16 constructs a graph network as shown in Figure 5 for all data included in the operational data and all data included in the training data. This generates a graph network representing the relationship between the training data metadata and the operational metadata. The graph network generation unit 16 sends the generated graph network information to the GUI 4. The GUI 4 receives the graph network information and displays the graph network.

[0054] Users can easily see the relationship between the operating data of the target system and the training data that forms the basis of the existing model applied to the learning process, from the graph network displayed in GUI4.

[0055] Next, the operation procedure of the machine learning device 2 according to Embodiment 1 will be described. Figure 6 is a flowchart showing an example of the operation procedure of the machine learning device 2 according to Embodiment 1.

[0056] In step S1, the machine learning device 2 acquires operational data by receiving operational data from the data acquisition device 3. The machine learning device 2 stores the acquired operational data in the storage unit 12.

[0057] In step S2, the data model construction unit 11 constructs a data model having a data structure that conforms to the structure of the target system. The data model construction unit 11 outputs the constructed data model to the metadata generation unit 13. The data model construction unit 11 may also store the constructed data model in the storage unit 12.

[0058] In step S3, the metadata generation unit 13 generates operational metadata based on the data model constructed in step S2. The metadata generation unit 13 stores the generated operational metadata in the storage unit 12.

[0059] In step S4, the determination unit 14 calculates the similarity between the metadata of the training data used to generate the existing model and the operational metadata generated in step S3, and determines which existing model to apply from among multiple existing models based on the similarity. The determination unit 14 outputs information indicating the existing model to apply to the learning unit 15.

[0060] In step S5, the learning unit 15 generates a trained model for the target system by applying the applicable model to the learning unit. The learning unit 15 obtains information from the determination unit 14 indicating an existing model to be used as the applicable model, and reads the applicable model, which is the existing model indicated in that information, from the model DB. The learning unit 15 generates a trained model for the target system by applying the read-out applicable model to the learning unit. The learning unit 15 stores the generated trained model in the storage unit 12.

[0061] In step S6, the graph network generation unit 16 generates a graph network that represents the relationship between the metadata of the training data used to generate the existing model, which is the applied model, and the operational metadata.

[0062] In step S7, the graph network generation unit 16 outputs the graph network generated in step S6 to the GUI4. With this, the machine learning device 2 completes its operation according to the procedure shown in Figure 6.

[0063] According to Embodiment 1, the machine learning device 2 includes a data model construction unit 11 that constructs a data model having a data structure in accordance with the structure of the system to be learned; a metadata generation unit 13 that generates metadata of operational data acquired from the system based on the data model; a determination unit 14 that determines the similarity between the metadata of the training data used to generate an existing model, which is an existing trained model, and the metadata of the operational data acquired from the system, and determines from among a plurality of existing models to be used as an applicable model for learning about the system based on the similarity; and a learning unit 15 that generates a trained model about the system by learning with the applicable model applied. The machine learning device 2 can achieve a highly interpretable determination for the user by determining the existing model to be used as an applicable model based on the similarity between the metadata. The user can easily understand the reason why the existing model determined to be an applicable model was determined to be an applicable model. As a result, the machine learning device 2 has the effect of being able to perform learning with an existing trained model applied in a process that is easy for the user to understand.

[0064] The learning unit 15 may also generate a trained model of the system by transfer learning with the applied model. This allows the machine learning device 2 to generate a trained model of the system to be learned by learning with an existing model.

[0065] The learning unit 15 may generate a trained model for the system by fine-tuning the applied model. This allows the machine learning device 2 to generate a trained model for the system being studied by learning with an existing model.

[0066] The determination unit 14 may determine from among multiple existing models to be applied as the model. The learning unit 15 may generate a trained model of the system by learning by applying two or more application models. As a result, the machine learning device 2 can generate a trained model that reflects the learning results for each of the two or more existing models.

[0067] The machine learning device 2 may further include a graph network generation unit 16 that generates a graph network representing the relationship between the metadata of the training data used to generate the existing model applied to the learning unit 15 and the metadata of the operational data. This allows the machine learning device 2 to present the relationship between the operational data and the training data to the user in an easily understandable manner.

[0068] Embodiment 2. Figure 7 shows an example configuration of the machine learning system 1A according to Embodiment 2. The machine learning system 1A comprises a machine learning device 2A, a data acquisition device 3, and a GUI 4. The machine learning device 2A receives operational data of the system to be studied and performs learning targeting that system. In Embodiment 2, the same reference numerals are used for components identical to those in Embodiment 1, and the configuration that differs from Embodiment 1 will be described in detail.

[0069] The machine learning device 2A comprises a data model construction unit 11, a storage unit 12, a metadata generation unit 13A, a determination unit 14A, a learning unit 15, a graph network generation unit 16, and a metadata analysis unit 23. The metadata generation unit 13A comprises an operational metadata generation unit 21 and a model metadata generation unit 22.

[0070] The operational metadata generation unit 21 generates operational metadata in the same manner as the metadata generation unit 13 shown in Figure 1. The operational metadata generation unit 21 stores the generated operational metadata in the meta DB within the storage unit 12. The model metadata generation unit 22 generates model metadata, which is metadata for existing models. In one example, the model DB stores information equivalent to model metadata along with existing models. The model metadata generation unit 22 generates model metadata by reading the information equivalent to model metadata from the model DB. The model metadata generation unit 22 stores the generated model metadata in the meta DB within the storage unit 12.

[0071] Examples of model metadata generated by the model metadata generation unit 22 include basic information about the trained model, information about the training data used for training, training parameters, information about the evaluation of the trained model, and information about the deployment environment. Basic information about the trained model includes, for example, the name of the trained model, the date the trained model was created, and information indicating the source of the trained model. Information about the training data used for training includes, for example, the name of the dataset, information about the preprocessing of the dataset, and information about the features. Examples of training parameters include parameters indicating the type of trained model, hyperparameters set for training, loss function, and objective function. Information about the evaluation of the trained model includes, for example, information about evaluation metrics, evaluation results, and evaluation accuracy. Information about the deployment environment includes, for example, information about dependent libraries, information about memory usage during training, and information about processing time during inference.

[0072] The model metadata generation unit 22 generates metadata expected for the newly trained model generated by the learning unit 15, based on the operational metadata. The model metadata generation unit 22 may also generate metadata expected for the newly trained model generated by the learning unit 15 based on data input by the user from the GUI 4 to the machine learning device 2A. Hereinafter, the newly trained model generated by the learning unit 15 will be referred to as the new model. The metadata expected for the new model will be referred to as the expected metadata.

[0073] The assumed metadata for a new model is a part of the metadata for the new model generated by the learning unit 15. The assumed metadata for a new model is data that is known before the new model is generated, or data that can be assumed before the new model is generated. The assumed metadata for a new model includes, for example, information about the features used for training, information indicating the type of trained model, loss function, objective function, and evaluation metric information. For example, information that is determined by the actual generation of the trained model, such as evaluation results, evaluation accuracy, and processing time in inference, is not included in the assumed metadata for a new model.

[0074] The metadata analysis unit 23 reads operational metadata and model metadata for operational data acquired from the target system from the metadata database. The metadata analysis unit 23 obtains assumed metadata for the new model from the model metadata generation unit 22. The metadata analysis unit 23 analyzes the operational metadata, the model metadata of the existing model, the metadata of the training data used to generate the existing model, and the assumed metadata to construct a graph network that represents the relationship between the operational data, the existing model, the training data, and the new model.

[0075] The determination unit 14A calculates the similarity between the metadata of the training data used to generate the existing model and the operational metadata, similar to the determination unit 14 shown in Figure 1.

[0076] The determination unit 14A reads the model metadata of the existing model from the metaDB in the storage unit 12. The determination unit 14A obtains the assumed metadata of the new model, which was generated by the model metadata generation unit 22. Based on the graph network constructed by the metadata analysis unit 23, the determination unit 14A calculates the similarity between the model metadata of the existing model and the assumed metadata of the new model.

[0077] The determination unit 14A determines which existing model to apply based on the similarity between the metadata of the training data used to generate the existing model and the operational metadata, and the similarity between the model metadata of the existing model and the assumed metadata of the new model. The determination unit 14A determines which one or more existing models to apply.

[0078] The learning unit 15 generates a trained model for the target system by applying one or more application models to the learning unit 15 shown in Figure 1.

[0079] Next, the operation procedure of the machine learning device 2A according to Embodiment 2 will be described. Figure 8 is a flowchart showing an example of the operation procedure of the machine learning device 2A according to Embodiment 2.

[0080] In step S11, the machine learning device 2A acquires operational data by receiving operational data from the data acquisition device 3. The machine learning device 2A stores the acquired operational data in the storage unit 12.

[0081] In step S12, the data model construction unit 11 constructs a data model having a data structure that conforms to the structure of the target system. The data model construction unit 11 outputs the constructed data model to the metadata generation unit 13A. The data model construction unit 11 may also store the constructed data model in the storage unit 12.

[0082] In step S13, the metadata generation unit 13A generates operational metadata, model metadata for the existing model, and assumed metadata for the new model. The metadata generation unit 13A generates operational metadata using the operational metadata generation unit 21. The metadata generation unit 13A generates model metadata for the existing model using the model metadata generation unit 22. The metadata generation unit 13A stores the generated operational metadata and generated model metadata in the storage unit 12. The metadata generation unit 13A generates assumed metadata for the new model using the model metadata generation unit 22.

[0083] In step S14, the metadata analysis unit 23 analyzes the operational metadata, the model metadata of the existing model, the metadata of the training data used to generate the existing model, and the assumed metadata of the new model to construct a graph network that represents the relationship between the operational data, the existing model, the training data, and the new model.

[0084] In step S15, the determination unit 14A determines which existing model to apply based on the similarity between the metadata of the training data used to generate the existing model and the operational metadata, and the similarity between the model metadata of the existing model and the assumed metadata of the new model. The determination unit 14A outputs information indicating which existing model to apply to the learning unit 15.

[0085] In step S16, the learning unit 15 generates a trained model for the target system by applying the applicable model to the learning unit. The learning unit 15 obtains information from the determination unit 14A indicating an existing model to be used as the applicable model, and reads the applicable model, which is the existing model indicated in that information, from the model DB. The learning unit 15 generates a trained model for the target system by applying the read-out applicable model to the learning unit. The learning unit 15 stores the generated trained model in the storage unit 12.

[0086] In step S17, the graph network generation unit 16 generates a graph network that represents the relationship between the metadata of the training data used to generate the existing model, which is the applied model, and the operational metadata.

[0087] In step S18, the graph network generation unit 16 outputs the graph network generated in step S17 to the GUI4. With this, the machine learning device 2A completes its operation according to the procedure shown in Figure 8.

[0088] According to Embodiment 2, the metadata generation unit 13A includes an operational metadata generation unit 21 that generates metadata for operational data, and a model metadata generation unit 22 that generates model metadata, which is metadata for an existing model. The determination unit 14A determines which existing model to apply based on the similarity between the metadata of the training data used to generate the existing model and the operational metadata, and the similarity between the model metadata of the existing model and the metadata assumed for the newly trained model generated by the training unit 15. Since the machine learning device 2A determines the applicable model based on the similarity between the metadata of the training data and the operational metadata, and the similarity between the model metadata of the existing model and the metadata assumed for the new trained model, it can achieve a determination that is highly interpretable to the user. The user can easily understand the reason why an existing model was determined to be an applicable model. As a result, the machine learning device 2A has the effect of being able to perform learning using an existing trained model in a process that is easy for the user to understand.

[0089] The machine learning device 2A may include a metadata analysis unit 23 that analyzes the metadata of the operational data, the model metadata of the existing model, the metadata of the training data used to generate the existing model, and the assumed metadata of the new model to construct a graph network representing the relationship between the operational data, the existing model, the training data, and the new model. The determination unit 14A may calculate the similarity between the model metadata and assumed metadata of the existing model based on the constructed graph network. As a result, the machine learning device 2A can determine the similarity between the model metadata and assumed metadata of the existing model based on the relationship between the operational data, the existing model, the training data, and the new model.

[0090] Next, the hardware configuration for realizing the machine learning devices 2 and 2A according to Embodiment 1 or 2 will be described. The machine learning devices 2 and 2A are realized by processing circuits. The processing circuit is, for example, a circuit in which a processor executes software.

[0091] When the processing circuit is implemented by software, the processing circuit is, for example, the control circuit 40 shown in Figure 9. Figure 9 is a diagram showing an example configuration of the control circuit 40 according to Embodiment 1 or 2. The control circuit 40 comprises an input unit 41, a processor 42, a memory 43, and an output unit 44. The input unit 41 is an interface circuit that receives data input from outside the control circuit 40 and provides it to the processor 42. The output unit 44 is an interface circuit that sends data from the processor 42 or the memory 43 to the outside of the control circuit 40.

[0092] The machine learning devices 2,2A are implemented by software, firmware, or a combination of software and firmware. The software or firmware is written as a program and stored in memory 43. The control circuit 40 implements the various functions of the machine learning devices 2,2A by having the processor 42 read and execute the program stored in memory 43. In other words, the control circuit 40 has memory 43 for storing the program that will ultimately execute the processing of the machine learning devices 2,2A. This program can also be said to cause the computer system to execute the procedures and methods of processing for the machine learning devices 2,2A. Memory 43 is also used as temporary memory when the processor 42 executes various processes.

[0093] In the machine learning device 2,2A, the processing functions of the data model construction unit 11, metadata generation units 13,13A, determination units 14,14A, learning unit 15, graph network generation unit 16, and metadata analysis unit 23 are realized by using a processor 42 and memory 43. The processor 42 is a CPU (Central Processing Unit). The processor 42 may also be a central processing unit, processing unit, arithmetic unit, microprocessor, microcomputer, processor, or DSP (Digital Signal Processor). The memory 43 includes, for example, non-volatile or volatile semiconductor memory such as RAM (Random Access Memory), ROM (Read Only Memory), flash memory, EPROM (Erasable Programmable Read Only Memory), EEPROM (Registered Trademark) (Electrically Erasable Programmable Read Only Memory), magnetic disks, flexible disks, optical disks, compact disks, minidiscs, or DVDs (Digital Versatile Discs). The storage unit 12 in the machine learning device 2,2A is realized by using memory 43.

[0094] The program according to Embodiment 1 or 2 may be provided on a recording medium such as a CD (Compact Disc)-ROM or DVD-ROM. The program according to Embodiment 1 or 2 may be provided by being stored on a computer connected to a network such as the Internet and downloaded via the Internet or other network. The program according to Embodiment 1 or 2 may be provided or distributed via a network such as the Internet.

[0095] The functions of machine learning device 2,2A may be implemented by dedicated hardware circuits. These dedicated hardware circuits include processing circuits. The processing circuits may be single circuits, composite circuits, programmed processors, parallel programmed processors, ASICs (Application Specific Integrated Circuits), FPGAs (Field Programmable Gate Arrays), or combinations thereof. The functions of machine learning device 2,2A may be implemented by processing circuits individually, or by a single processing circuit encompassing all functions. Furthermore, machine learning device 2,2A may be implemented by combining control circuits and hardware circuits.

[0096] In the embodiments 1 and 2 described above, the target system was assumed to be an FA system. However, the target system may be a system other than an FA system. For example, the target system may be a water treatment plant, a power plant, a chemical plant, or building facilities.

[0097] The configurations shown in each of the embodiments described above are examples of the content of this disclosure. The configurations of each embodiment can be combined with other known technologies. The configurations of each embodiment may be combined with each other as appropriate. It is possible to omit or modify parts of the configurations of each embodiment without departing from the gist of this disclosure. [Explanation of symbols]

[0098] 1,1A Machine learning system, 2,2A Machine learning device, 3 Data acquisition device, 4 GUI, 11 Data model construction unit, 12 Memory unit, 13,13A Metadata generation unit, 14,14A Decision unit, 15 Learning unit, 16 Graph network generation unit, 21 Operation metadata generation unit, 22 Model metadata generation unit, 23 Metadata analysis unit, 40 Control circuit, 41 Input unit, 42 Processor, 43 Memory, 44 Output unit.

Claims

1. A data model construction unit that constructs a data model having a data structure that conforms to the structure of the system to be learned, A metadata generation unit generates metadata for operational data acquired from the system based on the data model, A determination unit determines the similarity between the metadata of the training data used to generate an existing model, which is an existing trained model, and the metadata of the operational data obtained from the system, and based on the similarity, determines from among a plurality of existing models which existing model will be used as the applicable model to be applied to learning about the system. The system comprises a learning unit that generates a trained model for the system by learning using the aforementioned application model. A machine learning device characterized by the following features.

2. The metadata generation unit, An operation metadata generation unit that generates metadata for the operation data, It includes a model metadata generation unit that generates model metadata, which is metadata of the existing model, The determination unit determines the existing model to be applied based on the similarity between the metadata of the training data used to generate the existing model and the metadata of the operational data, and the similarity between the model metadata of the existing model and the assumed metadata, which is assumed metadata for a new model that is a newly trained model generated by the training unit. The machine learning device according to feature 1.

3. The system further includes a metadata analysis unit that analyzes the metadata of the operational data, the model metadata of the existing model, the metadata of the training data used to generate the existing model, and the assumed metadata of the new model to construct a graph network representing the relationship between the operational data, the existing model, the training data, and the new model. The determination unit calculates the similarity between the model metadata of the existing model and the assumed metadata based on the constructed graph network. The machine learning apparatus according to feature 2.

4. The learning unit generates a trained model for the system by transfer learning to which the applied model is applied. A machine learning device according to any one of claims 1 to 3.

5. The learning unit generates a trained model for the system by fine-tuning the applied model. A machine learning device according to any one of claims 1 to 3.

6. The determination unit determines two or more existing models to be the applicable model from among the plurality of existing models, The learning unit generates a trained model for the system by learning by applying two or more of the aforementioned application models. A machine learning device according to any one of claims 1 to 3.

7. The learning unit further comprises a graph network generation unit that generates a graph network representing the relationship between the metadata of the training data used to generate the existing model applied to the learning unit and the metadata of the operational data. A machine learning device according to any one of claims 1 to 3.

8. A data acquisition device that acquires operational data of the system to be studied, The system includes a machine learning device that receives the aforementioned operational data and performs learning targeting the aforementioned system, The aforementioned machine learning device is A data model construction unit constructs a data model having a data structure in accordance with the structure of the aforementioned system, A metadata generation unit generates metadata for the operational data based on the data model, A determination unit determines the similarity between the metadata of the training data used to generate an existing model, which is an existing trained model, and the metadata of the operational data obtained from the system, and based on the similarity, determines from among a plurality of existing models which existing model will be used as the applicable model to be applied to learning about the system. The system comprises a learning unit that generates a trained model for the system by learning using the aforementioned application model. A machine learning system characterized by the following features.

9. The steps include: constructing a data model with a data structure that conforms to the structure of the system being studied; The steps include generating metadata for operational data acquired from the system based on the data model, The process involves determining the similarity between the metadata of the training data used to generate an existing model, which is an existing trained model, and the metadata of the operational data obtained from the system, and then, based on the similarity, determining from among a plurality of existing models which existing model will be used as the applicable model for learning about the system. The step includes generating a trained model for the system by learning with the aforementioned application model applied. A machine learning method characterized by the following features.

10. In the computer system, The steps include: constructing a data model with a data structure that conforms to the structure of the system being studied; The steps include generating metadata for operational data acquired from the system based on the data model, The process involves determining the similarity between the metadata of the training data used to generate an existing model, which is an existing trained model, and the metadata of the operational data obtained from the system, and then, based on the similarity, determining from among a plurality of existing models which existing model will be used as the applicable model for learning about the system. The step of generating a trained model for the system by learning using the aforementioned application model is performed. A program characterized by the following features.

Images