TRANSFER LEARNING DEVICE AND TRANSFER LEARNING PROCEDURES

Abstract

Transfer learning device, including: a processing data acquisition unit (11) for acquiring a plurality of processing data elements including data relating to time series processing conditions in order to cause a processing machine (4) to perform processing for a plurality of steps and data relating to a processing state to be performed by the processing machine (4) for the plurality of steps on the basis of the processing conditions; an analysis unit (13) for calculating an importance with respect to each part of the processing data captured by the processing data acquisition unit (11), based on a plurality of processing data elements captured by the processing data acquisition unit (11); an extraction unit (14) for extracting a plurality of transfer learning data elements for performing transfer learning of a first machine learning model, which receives as input the data relating to the processing conditions in time series for the plurality of steps, which are generated in advance from a plurality of processing data elements, which are captured by the processing data acquisition unit on the basis of the importance calculated by the analysis unit (13), and which outputs data relating to a processing state after execution for the plurality of steps; and a learning unit (153) for performing transfer learning of the first machine learning model using a plurality of transfer learning data elements extracted from the extraction unit (14), and for generating a second machine learning model which receives as input data relating to the processing conditions in time series for the plurality of steps and outputs data relating to a processing state after execution for the plurality of steps.

Description

TECHNICAL AREA

[0001] The present disclosure relates to a transfer learning device and a transfer learning method for generating a learned model (hereinafter referred to as the “machine learning model”) to assist in setting a machining condition when a machining machine is prompted to perform a machining operation. STATE OF THE ART

[0002] When machining a workpiece, the machine tool follows the machining conditions in each stage of the process. A machining condition includes, for example, a value that specifies a position or speed of the machining operation. These conditions are either set by a machine tool operator or by a control device, such as one that controls the machine tool, according to a program. However, in a machine tool, it is possible that the machining process will not proceed according to the specified conditions due to disturbances such as friction, temperature, or vibration.Therefore, a conventional technique is known that makes it possible to set a machining condition suitable for a machining machine in order to obtain a machining result according to the instructions by using a machine learning model that takes data relating to the machining condition as input and outputs data relating to a machining state that is executed by the machining machine based on the machining condition.

[0003] A technique known as transfer learning is now available, in which a machine learning model is applied to learning in a different domain. Patent literature 1, for example, discloses a technique for performing additional learning of a common model using sampling data collected by a device in a server facility. This device selects an optimal common model and delivers it to various devices with different environments, conditions, or similar conditions. Although the learning is performed in advance, the common model is in a situation where the learning does not occur in an environment and under conditions specific to the device. Therefore, the server facility performs additional learning to fine-tune the common model to enable high-accuracy identification, classification, and similar tasks.The server setup receives the effect of an action known as transfer learning by performing additional learning based on the shared model. LIST OF COUNTERPOINTMENTS

[0004] Patent literature 1: WO 2018 / 173121 DEMOLITION TECHNICAL PROBLEM

[0005] The characteristics of a machine tool change depending on its operating conditions, such as aging, workpiece modifications, or changes in the environment. Therefore, any machine learning model used to adjust machining conditions to achieve a predetermined machining result must be appropriate for the current operating environment.

[0006] However, if the user starts from scratch according to the current situation of the processing machine and creates a new machine learning model, it takes time.

[0007] For example, it is conceivable, as in the method disclosed in the patent literature 1 described above, to generate a machine learning model suitable for the current operating situation of the processing machine (hereinafter referred to as the “second machine learning model”) on the basis of a machine learning model previously created using a so-called transfer learning technique (hereinafter referred to as the “first machine learning model”).

[0008] Here, for example, data obtained from the operation of the machine tool in the current situation must be collected as data (hereinafter referred to as "transfer learning data") for the creation of the second machine learning model. This data originates, for example, from a control device that controls the machine tool. The control device takes a time-series sampling of machining conditions and machining states during a process in which the machine tool performs machining operations within a predetermined time period. Generally, however, the sampling time of the control device is short, e.g., 1 ms. Furthermore, the data collected by the control device may include data specifying different machining conditions for handling various machining tasks on different workpieces.This means that the amount of data captured by the control device during the operation of the processing machine is enormous.

[0009] Even when a so-called transfer learning method is used, it therefore takes a long time to create the second machine learning model, because the amount of data collected through the operation of the processing machine is enormous.

[0010] As described above, the problem is that it is very time-consuming to create the second machine learning model for setting the machining conditions that are suitable for the situation of the machining machine at the location where the machining machine actually works, using all the data collected at the machining location.

[0011] The present disclosure was made to solve the above-mentioned problems, and one of its aims is to provide a transfer learning device capable of generating a second machine learning model to set a machining condition suitable for a machining situation at a location where the machining machine is actually operating, from a portion of the data without requiring all the data collected at the machining location. SOLUTION TO THE TASK

[0012] A transfer learning device according to the present disclosure comprises: a machining data acquisition unit for acquiring a plurality of machining data elements, including data relating to time-series machining conditions, to cause a machining machine to perform machining for a plurality of steps and data relating to a machining state executed by the machining machine for the plurality of steps based on the machining conditions; an analysis unit for calculating an importance with respect to each part of the machining data acquired by the machining data acquisition unit, based on a plurality of machining data elements acquired by the machining data acquisition unit;an extraction unit for extracting a multitude of transfer learning data elements for performing transfer learning of a first machine learning model, which receives as an input the data relating to the processing conditions in time series for the multitude of steps, which are generated in advance from a multitude of processing data elements, which are captured by the processing data acquisition unit on the basis of the importance calculated by the analysis unit, and which outputs data relating to a processing state after execution for the multitude of steps;and a learning unit for performing transfer learning of the first machine learning model using a plurality of transfer learning data elements extracted by the extraction unit, and for generating a second machine learning model that receives as input data relating to the processing conditions in time series for the plurality of steps, and outputs data relating to a processing state after execution for the plurality of steps. ADVANTAGEOUS EFFECTS OF THE INVENTION

[0013] According to the present disclosure, the above configuration makes it possible to generate the second machine learning model for setting the machining conditions suitable for the situation of the machining machine at the actual place of use from a portion of the data, without requiring all the data collected at the machining site. Thus, the second machine learning model can be created without significant time expenditure. BRIEF DESCRIPTION OF THE DRAWINGS Fig. Figure 1 is a representation showing a configuration example of a transfer learning device according to a first embodiment. Fig. Figure 2 is a representation showing a detailed configuration example of a processing data acquisition unit in the transfer learning device according to the first embodiment. Fig. Figure 3 is a representation showing a detailed configuration example of an analysis unit in the transfer learning device according to the first embodiment. Fig. Figure 4 is a flowchart to explain the operation of the transfer learning device according to the first embodiment. Fig. Figure 5 is a flowchart to explain in detail the operation of the analysis unit of the transfer learning device according to the first embodiment. Fig. Figure 6 is a flowchart to explain in detail the operation of the machine learning of the transfer learning device according to the first embodiment. Fig. 7A and Fig. Figure 7B shows an example of a hardware configuration of the transfer learning device according to the first embodiment. DESCRIPTION OF THE EXECUTION FORMS

[0014] An embodiment of the present disclosure is described in detail below with reference to the drawings. First embodiment.

[0015] Fig. Figure 1 is a representation showing a configuration example of a transfer learning device 1 according to a first embodiment.

[0016] Fig. Figure 2 is a representation showing a detailed configuration example of a processing data acquisition unit 11 in the transfer learning device 1 according to the first embodiment.

[0017] Fig. Figure 3 is a representation showing a detailed configuration example of an analysis unit 13 in the transfer learning device 1 according to the first embodiment.

[0018] The transfer learning device 1 according to the first embodiment generates a learned model (hereinafter referred to as the “machine learning model”) for defining a machining condition when a machining machine 4 is prompted to perform a machining operation.

[0019] The machining tool 4 is an industrial device that performs machining according to the machining conditions. In the first embodiment, it is assumed that the machining tool 4 is, for example, a numerically controlled (NC) machining tool. It should be noted that this is merely an example, and it is sufficient that the machining tool 4 is a motor-driven machining tool.

[0020] The processing machine 4, for example, cuts or polishes a metal object (hereinafter referred to as the "workpiece") or removes excess material using energy, e.g., electricity, to bring the workpiece into the desired shape. It should be noted that the workpiece is not limited to metal.

[0021] Examples of processing machine 4 are a laser processing machine, a spark erosion machine, a cutting processing machine, a polishing processing machine, an electrolytic processing machine, an ultrasonic processing machine and an electron beam processing machine.

[0022] The machining conditions are configured by a combination of several control parameters used to control the machining center 4. For example, the machining conditions include a numerical value specifying a command position, command speed, command acceleration, return position, return speed, return acceleration, or current return, among other control parameters. The machining conditions can also specify the workpiece type or a numerical value indicating the machining process for that workpiece.

[0023] It should be noted that a control device 2 is connected to the processing machine 4 and controls the processing machine 4 to process a workpiece according to the processing conditions.

[0024] In the first embodiment, it is assumed that the machining conditions are set by the control device 2 depending on the machining process, i.e., on each machining step.

[0025] For example, the user on site operates an input device (not shown) contained in the control device 2 to enter a type of workpiece to be machined by the processing machine 4 and an instruction for the machining content.

[0026] For example, the control device 2 contains a program that specifies which type of machining step is performed when a particular machining operation is carried out on a specific workpiece. When the user enters an instruction specifying the type of workpiece to be machined by the machine tool 4 and the machining operation, the control device 2 sets a machining condition for each machining step according to the stored program and causes the machine tool 4 to execute each machining step according to the set machining condition.

[0027] The control device 2 controls the machining center 4 and takes samples of data relating to the machining conditions set by the control device 2, and of data relating to a machining state in a process in which the machining center 4 performs machining at a machining location for a preset duration (hereinafter referred to as the "sampling time"). The machining state refers to a condition of a result of the machining of a workpiece and is, for example, a defect in a machining result that is achieved by the actual operation of the machining center 4 with respect to a machining result in the manner instructed by the machining center 4.In the first embodiment, it is assumed that the machining state is a positional error of the machining position actually performed by the machining machine 4 in relation to the position at which the machining machine 4 is instructed to perform the machining.

[0028] The processing machine 4 performs a large number of processing steps within a sampling time.

[0029] To initiate machining operation on the machine tool 4, the control device 2 collects data relating to the machining conditions for a multitude of steps and data relating to the machining state being performed for the multitude of steps over a sampling period. Hereinafter, the data relating to the machining conditions and the data relating to the machining state, which are acquired by the control device 2 from the machine tool 4 at the location, are referred to as "machining data".

[0030] The data contained in certain processing data relating to processing conditions includes time-series data of control parameters for a variety of steps performed during the sampling time. That is, the data relating to processing conditions and contained in certain processing data includes time-series data specifying the command position, time-series data specifying the command rate, time-series data specifying the command acceleration, or similar data for a variety of steps.

[0031] Furthermore, the machining state, for data relating to the machining process performed for a multitude of steps contained in the machining data, is a state of the result of machining a workpiece at a time point several steps after the machining process for the multitude of steps has been carried out. Specifically, certain machining data includes data on the machining conditions for n steps up to a specific time t and data on a machining state at time t+k.

[0032] It should be noted that the sampling time of the control device 2 is generally about 1 ms.

[0033] The transfer learning device 1 receives a multitude of machining data elements from the control device 2. The transfer learning device 1 generates a machine learning model for setting a machining condition suitable for a situation at a location where the machining machine 4 operates, such as aging, a change in a workpiece, or the environment of the location where the machining machine 4 operates, based on a multitude of machining data elements acquired by the control device 2.

[0034] In particular, the transfer learning device 1 uses a variety of machining data elements to perform transfer learning based on a machine learning model created in advance at the time of factory delivery of the machining machine 4 or the like, thereby generating a machine learning model for setting a machining condition suitable for the situation of the site where the machining machine 4 is operating.

[0035] In the first embodiment, a machine learning model created in advance at the time of delivery of the processing machine 4 or the like is referred to as the "first machine learning model," and a machine learning model generated by the transfer learning device 1 based on the first machine learning model is referred to as the "second machine learning model." In the first embodiment, it is assumed that the first machine learning model and the second machine learning model are configured by a neural network.

[0036] The first machine learning model receives as input data relating to time-series processing conditions for a multitude of steps and outputs data relating to a processing state after execution for the multitude of steps.

[0037] The first machine learning model is generated by learning, for example by using a large number of machining data elements as training data, which are obtained by the machining machine 4 performing basic machining on a hypothetical basic workpiece before shipping (hereinafter referred to as "preliminary training data").

[0038] For example, a manufacturer or similar entity of the machine tool 4 operates an input device contained in the control device 2 to input instructions regarding the type of workpiece to be subjected to basic machining by the machine tool 4 and the basic machining content. When instructions regarding the type of workpiece to be subjected to basic machining by the machine tool 4 and basic machining content are entered by a manufacturer or similar entity, the control device 2 sets machining conditions at each machining step and causes the machine tool 4 to execute each machining step according to the set machining conditions. The control device 2 causes the machine tool 4 to execute each step, thereby acquiring a variety of preliminary learning data elements.It should be noted that the processing status data included in the preliminary learning data is a teacher label. The generated first machine learning model is delivered together with the machining center 4 and the control device 2.

[0039] In machine tool 4, it is possible that the machining process may deviate from the specified machining conditions due to disturbances such as friction, temperature, or vibration. Since the first machine learning model is created in advance, the control device 2 can estimate a positional error that will occur when machine tool 4 is operated under the set machining conditions, for example, based on the first machine learning model and the data related to the machining conditions. The control device 2 can then correct the machining condition, taking the estimated positional error into account.In addition, the user is provided with information about the position error estimated by the control device 2, and the user can correct the program stored in the control device 2 or adjust the machining conditions taking into account the estimated position error.

[0040] However, the characteristics of the machining tool 4 change depending on the actual conditions at its operating location. Therefore, even if the initial machine learning model is created, it may not be suitable for the specific location, and the machining tool 4—more precisely, the control device 2 that controls the machining tool 4—may not perform the machining operation intended by the user. This is because data relating to machining conditions not previously learned can be input into the initial machine learning model.

[0041] Therefore, according to the first embodiment, the transfer learning device 1 performs transfer learning based on the first machine learning model and generates a second machine learning model for setting a machining condition suitable for the location where the machining center 4 operates. The transfer learning device 1 according to the first embodiment can create the second machine learning model efficiently and without significant time expenditure. The details are described below.

[0042] The transfer learning device 1 according to the first embodiment is connected to the control device 2 and the storage device 3.

[0043] As described above, the control device 2 controls the processing machine 4 and collects data relating to the processing status in the process in which the processing machine 4 performs a processing operation with the sampling time.

[0044] The storage facility 3 includes a storage unit 31 for preliminary learning data and a storage unit 32 for preliminary learning models.

[0045] The preliminary training data storage unit 31 stores a variety of preliminary training data elements used at the time of creating the first machine learning model. The preliminary training model storage unit 32 stores the first machine learning model.

[0046] It should be noted that the storage device 3 can be located in the processing machine 4 or in the control device 2. Furthermore, the control device 2 can be located in the processing machine 4.

[0047] The transfer learning device 1 generates the second machine learning model by performing transfer learning of the first machine learning model based on the multiple machining data elements acquired by the control device 2 and the first machine learning model stored in the storage device 3. As described above, the machining data received by the transfer learning device 1 from the control device 2 includes data relating to the machining conditions, including time-series data for each control parameter for a multitude of steps that the control device 2 directs the machining machine 4 to perform, and data relating to the machining state.

[0048] The transfer learning device 1 comprises a processing data acquisition unit 11, a unit 12 for acquiring preliminary learning data, an analysis unit 13, an extraction unit 14 and a machine learning unit 15.

[0049] The processing data acquisition unit 11 comprises a processing data preprocessing unit 111 and a processing data storage unit 112.

[0050] The analysis unit 13 comprises a compression unit 131 and an analysis unit 132.

[0051] The machine learning unit 15 comprises a selection unit 151, a model reading unit 152, a learning unit 153, an evaluation unit 154, a model storage unit 155, an evaluation value storage unit 156 and a weighting update unit 157.

[0052] The processing data acquisition unit 11 acquires and stores a large number of processing data elements from the control device 2.

[0053] The processing data preprocessing unit 111 of the processing data acquisition unit 11 acquires processing data from the control device 2 and shapes the acquired processing data by applying a low-pass filter or the like. It should be noted that the processing data preprocessing unit 111 acquires processing data from the control device 2 for a preset period, e.g., one hour.

[0054] The processing data preprocessing unit 111 stores the processed data after shaping in the processing data storage unit 112.

[0055] The processing data storage unit 112 stores a variety of processing data elements that were formed by the processing data preprocessing unit 111.

[0056] The processing data acquisition unit 11 stores a selection of processing data from the multitude of processing data elements in the processing data storage unit 112. These processing data elements are selected separately from the multitude of other processing data elements after the processing data has been formatted. The processing data elements to be evaluated can be one or more. The processing data acquisition unit 11 can select processing data for evaluation from a multitude of processing data elements either randomly or according to a specific rule. For example, the defined rule might state that processing data recorded in the second half of a 10-minute period, such as an hour, is used as processing data for evaluation.

[0057] It should be noted that the processing data storage unit 112 is contained within the transfer learning device 1 here, but this is only an example. The processing data storage unit 112 can be located outside the transfer learning device 1 at a location that the transfer learning device 1 can access.

[0058] Unit 12 for capturing preliminary learning data captures a variety of preliminary learning data elements that are stored in the preliminary learning data storage unit 31 of storage facility 3.

[0059] Unit 12, for capturing preliminary training data, outputs the multitude of received preliminary training data elements to analysis unit 13.

[0060] Analysis Unit 13 calculates the importance for the multitude of post-shaping machining data elements based on the multitude of post-shaping machining data elements stored in Machining Data Storage Unit 112 of Machining Data Acquisition Unit 11, and the multitude of preliminary training data elements output by Unit 12 for preliminary training data acquisition. Hereinafter, the post-shaping machining data will simply be referred to as machining data. Analysis Unit 13 calculates the importance for each machining data element.

[0061] In the first embodiment, the importance of the machining data calculated by the analysis unit 13 indicates the degree of learning effectiveness when the machining data is used as training data in the creation of the second machine learning model.

[0062] It should be noted that the numerous machining data elements after shaping, which are captured by the analysis unit 13 from the machining data storage unit 112, are different machining data than the machining data to be evaluated. The machining data stored in the machining data storage unit 112 for evaluation purposes is used when the evaluation unit 154 evaluates the second machine learning model. Details of the evaluation unit 154 are described later.

[0063] First, the compression unit 131 in the analysis unit 13 compresses a large number of editing data elements and a large number of preliminary training data elements. The analysis unit 13 can compress a large number of editing data elements and a large number of preliminary training data elements by using a known dimensionality reduction technique such as an autoencoder (AE), principal component analysis (PCA), or T-distributed stochastic neighbor embedding (t-SNE).

[0064] Since the compression unit 131 compresses the multitude of processing data elements and the multitude of preliminary learning data elements, the calculation of the importance of the processing data in the analysis unit 13 can be reduced.

[0065] Analysis Unit 132 calculates the importance for each post-compression edit data element based on a multitude of post-compression edit data elements (hereinafter referred to as "post-compression edit data elements") after compression by Compression Unit 131 and a multitude of preliminary training data elements (hereinafter referred to as "preliminary training data elements") after compression by Compression Unit 131.

[0066] Analysis Unit 132 calculates the importance of post-compression machining data by detecting steep changes in the data related to the machining conditions contained within the post-compression machining data, for example, using a change point detection method. Specifically, Analysis Unit 132 captures a value indicating a steep change in the time-series instruction velocity data or the time-series instruction acceleration data contained within the data relating to the machining condition. Analysis Unit 132 calculates the importance of the post-compression machining data such that the higher the captured score, the greater the importance. Analysis Unit 132 can use the determined score as the importance value or calculate the importance by inserting the determined score into a predefined formula.Velocity or acceleration is a control parameter that causes a position error. Analysis Unit 132 calculates the severity depending on whether there is a steep change in the control parameter causing the position error. It should be noted that Analysis Unit 132 does not require a large number of preliminary training data elements when calculating the severity using the post-compression change point detection method.

[0067] Furthermore, the analysis unit 132 can, for example, determine a data distance between a large number of editing data elements after compression and a large number of preliminary training data elements after compression, and calculate the data distance as an importance.

[0068] Specifically, the analysis unit 132 performs an “interdata distance calculation” for each processing data element after compression.

[0069] During the "interdata processing" process, the analysis unit 132 selects specific preliminary training data (hereinafter referred to as "preliminary target training data") from the multitude of preliminary training data elements after compression and calculates an interdata distance between the post-compression processing data and the preliminary target training data. The analysis unit 132 repeats the calculation of the interdata distance between the post-compression processing data and the preliminary training data elements until all processing data elements have been selected as preliminary training data. Then, the analysis unit 132 determines the maximum interdata distance among the calculated interdata distances as the interdata distance between the post-compression processing data and the multitude of preliminary training data elements after compression.

[0070] Analysis Unit 132 performs the "Interdata Distance Calculation Process" for all post-compression processing data elements and calculates the interdata distance between each post-compression processing data element and a multitude of post-compression preliminary training data elements. Analysis Unit 132 calculates the post-compression importance of the processing data such that the greater the interdata distance between the post-compression processing data and the multitude of post-compression preliminary training data elements, the greater the importance.

[0071] It should be noted that the analysis unit 132 is only required in the “interdata distance calculation processing” to calculate at least one interdata distance between the data relating to the processing condition contained in the processing data after compression and the data relating to the processing condition contained in the multitude of preliminary training data elements after compression.

[0072] Furthermore, the analysis unit 132 can, for example, perform an “inference error calculation operation” in which the importance is calculated based on an error (hereinafter referred to as “inference error”) between data relating to an operation state (hereinafter referred to as “first inference result”) obtained by input into the first machine learning model, data relating to an operation condition contained in a variety of operation data elements stored in the operation data storage unit 112 in the operation data acquisition unit 11, and data relating to an operation state contained in a variety of operation data elements stored in the operation data storage unit 112 in the operation data acquisition unit 11.

[0073] Analysis unit 132 performs an "inference error calculation processing" for each processing data element.

[0074] During the "Inference Error Calculation Processing" operation, Analysis Unit 132 calculates an inference error between the initial inference result and the data contained in the processing data that relates to the processing state. Analysis Unit 132 performs the "Inference Error Calculation Processing" on all processing data elements and calculates an inference error for each processing data element.

[0075] The analysis unit 132 calculates the importance of the processing data in such a way that the importance is higher the greater the interference error corresponding to the processing data.

[0076] As described above, when calculating importance through the "inference error calculation processing", the analysis unit 132 performs the "inference error calculation processing" on each processing data element before compression, rather than on each processing data element after compression, to calculate the importance for each processing data element.

[0077] In the case of performing inference error calculation processing, the analysis unit 132 can, for example, read the first machine learning model from the storage unit 32 for preliminary learning models of the storage device 3.

[0078] Furthermore, the analysis unit 132 can, for example, calculate the importance from the result of classifying a large number of post-compression processing data elements and a large number of post-compression preliminary training data elements using a classification procedure via a so-called decision tree. The analysis unit 132 assigns a higher importance to the post-compression processing data because the post-compression processing data contains more preliminary training data elements that do not fall into the same classification.

[0079] It should be noted that the analysis unit 132 is only required to classify at least the data relating to the processing condition contained in the processing data after compression, and the data relating to the processing condition contained in the multitude of preliminary training data elements after compression.

[0080] When the analysis unit 132 calculates the importance for the multitude of processing data elements after compression, the analysis unit 13 outputs each processing data element after compression and its importance in relation to each other to the extraction unit 14.

[0081] It should be noted that when the analysis unit 132 calculates the importance for a large number of processing data elements by performing the "inference error calculation processing", the analysis unit 13 outputs each processing data element and its importance in relation to each other to the extraction unit 14.

[0082] The extraction unit 14 extracts from the multitude of machining data elements captured by the machining data acquisition unit 11, based on the importance calculated by the analysis unit 13, a multitude of machining data elements for carrying out the transfer learning of the first machine learning model, in other words, transfer learning data.

[0083] For example, information that can specify the processing data, such as the date and time of its capture, is assigned to the processing data. Extraction unit 14 then combines the processing data output by analysis unit 13, after compression, with the processing data captured by processing data capture unit 11, using the information that can specify the processing data. Consequently, extraction unit 14 can determine the importance corresponding to the multitude of processing data elements captured by processing data capture unit 11. When the processing data is output by analysis unit 13, extraction unit 14 can link the processing data output by analysis unit 13 with the processing data captured by processing data capture unit 11.

[0084] Extraction unit 14 assigns the multitude of processing data elements captured by processing data acquisition unit 11 to the importance calculated by analysis unit 13 and sorts the multitude of processing data elements in descending order of importance. Then, extraction unit 14 extracts, for example, a predefined number of processing data elements in descending order of importance as transfer training data.

[0085] Furthermore, extraction unit 14 also extracts a set of features when extracting the transfer learning data.

[0086] For example, for each processing data element, extraction unit 14 extracts a set of features using a known moving average method with respect to data relating to a processing condition contained in the processing data, and uses processing data, including data relating to the processing condition, after extraction of the set of features and data relating to a processing state, as transfer learning data.For example, extraction unit 14 can extract, for each machining data element, a maximum and a minimum value of data relating to a machining condition contained in the machining data as feature sets and use the machining data, including the data relating to the machining condition, as transfer learning data after extracting the feature sets and the data relating to the machining state. For example, considering the speed data contained in the machining condition-related data, extraction unit 14 extracts the maximum and minimum values ​​of the speed data in time series as feature sets. Extraction unit 14 extracts a maximum and a minimum value as feature sets for each data element that specifies the control parameter contained in the machining condition-related data.

[0087] Furthermore, Extraction Unit 14, for example, calculates the importance of the feature set using LightGBM for the data relating to the processing condition contained in each processing data element. It then multiplies the data relating to the processing condition that is assumed to be unimportant based on the calculated importance of the feature set—in other words, it multiplies the data relating to the processing condition where the importance of the feature set is "0" by "0" to apply a mask. To describe a specific example, it is now assumed that the data relating to the processing condition includes, for example, command velocity data, command acceleration data, current command data, and command position data.It is then assumed that, as a result of calculating the importance of the feature sets with respect to the data relating to the machining condition using LightGBM, the importance of the feature sets with respect to the position data is "0", and the importance of the feature sets with respect to the other velocity data, acceleration data, and current data is "1" or more. In this case, extraction unit 14 multiplies the position data by "0" and applies a mask. Extraction unit 14 uses the machining data, including the data relating to the machining condition after the mask has been applied, and the data relating to the machining state, as transfer learning data.

[0088] The extraction unit 14 outputs a large number of transfer learning data elements to the machine learning unit 15, to which the importance calculated by the analysis unit 13 has been assigned and from which the feature set has been extracted.

[0089] The machine learning unit 15 performs transfer learning of the first machine learning model using the transfer learning data elements output by the extraction unit 14 and generates the second machine learning model.

[0090] Machine learning unit 15 is described in detail.

[0091] Selection unit 151 selects from the multitude of transfer learning data elements a multitude of transfer learning data elements (hereinafter referred to as "selected transfer learning data") for the creation of the second machine learning model.

[0092] In particular, the selection unit 151 assigns a priority to each of the plurality of transfer learning data elements output by the extraction unit 14 and selects the chosen transfer learning data based on the assigned priority. In the first embodiment, the priority assigned to the transfer learning data by the selection unit 151 indicates the extent to which the transfer learning data to which the priority has been assigned should be given priority for the creation of the second machine learning model.

[0093] When a large number of transfer learning data elements are output by extraction unit 14, selection unit 151 first assigns priority to these elements. For example, selection unit 151 uniformly assigns an initial priority value to each transfer learning data element. This initial priority value can be predefined. Furthermore, selection unit 151 can, for example, use the importance associated with a large number of transfer learning data elements as a priority. This processing, in which selection unit 151 assigns a priority to multiple transfer learning data elements when they are output by extraction unit 14, is also referred to as "processing with priority initial settings."

[0094] When assigning a priority to each transfer learning data element in "Editing with priority initial settings", selection unit 151 selects the selected transfer learning data elements based on the assigned priority.

[0095] For example, selection unit 151 chooses a preset number of transfer learning data elements as the selected transfer learning data in descending order of the assigned priority from among the multitude of transfer learning data elements. In a case where the same priority is assigned uniformly, selection unit 151 can, for example, select the transfer learning data elements randomly.

[0096] Furthermore, in a case where the selected transfer learning data, in which the assigned priority is updated, is output by the weighting update unit 157, selection unit 151 includes the selected transfer learning data in the sample based on the updated priority. For example, selection unit 151 includes in the sample the selected transfer learning data with a priority equal to or higher than a predefined threshold for the updated priority.

[0097] The details of weighting update unit 157 will be described later. Details regarding the sampling of the selected transfer learning data by selection unit 151 will also be described later, along with the description of weighting update unit 157.

[0098] Selection unit 151 outputs the selected transfer learning data to learning unit 153 when the multitude of transfer learning data elements is output by extraction unit 14.

[0099] In a case where the selected transfer learning data are sampled based on the selected transfer learning data output by the weighting update unit 157, the selection unit 151 outputs the selected transfer learning data to the learning unit 153 after sampling.

[0100] The model reading unit 152 reads the first machine learning model that is stored in the preliminary learning model storage unit 32 of the storage device 3.

[0101] It should be noted that the learning unit 153 can take over the function of the model reading unit 152. In this case, the transfer learning device 1 can be configured so that it does not contain the model reading unit 152.

[0102] Learning Unit 153 uses the multitude of selected transfer learning data elements output by Selection Unit 151 to perform transfer learning of the first machine learning model, which is read by Model Reading Unit 152. This generates the second machine learning model, which receives as input data relating to a time-series processing condition for a multitude of steps and outputs data relating to a processing state after execution for the multitude of steps. That is, the data relating to the processing state and contained in the selected transfer learning data is the teacher label.

[0103] The learning unit 153 stores the generated second machine learning model in the model storage unit 155.

[0104] Furthermore, learning unit 153 outputs the generated second machine learning model, along with the selected transfer learning data used in generating the second machine learning model, to evaluation unit 154. Learning unit 153 assigns the priority to the selected transfer learning data that is assigned by selection unit 151 when the selected transfer learning data is output.

[0105] The machine learning method used by learning unit 153 can be any method, as long as supervised learning is possible. Learning unit 153 generates the second machine learning model through so-called supervised learning.

[0106] It should be noted that if extraction unit 14 uses a known moving average method or a method for extracting the maximum and minimum values ​​of the data with respect to the processing condition during feature set extraction, the selected transfer learning data will have a data configuration that differs from that of the preliminary learning data when the first machine learning model is generated. That is, the selected transfer learning data does not match the input of the first machine learning model. Therefore, in this case, learning unit 153, for example, feeds the selected transfer learning data into a machine learning model (hereafter referred to as the "third machine learning model") that has been prepared in advance and differs from the first machine learning model in order to perform dimensionality reduction (feature set extraction).In parallel, learning unit 153 inputs the selected transfer learning data into the first machine learning model, which is to be learned halfway through, in a state where the feature set is not extracted, and performs a dimensionality reduction (extraction of the feature set). It should be noted that learning unit 153 can capture the selected transfer learning data in a state where the feature set is not extracted from extraction unit 14 via selection unit 151.Then, learning unit 153 performs transfer learning, combining data obtained through partial learning with the first machine learning model and data obtained from the third machine learning model. This combined data is then fed into a machine learning model (hereinafter referred to as the "fourth machine learning model") that differs from the first and third machine learning models, which were pre-prepared to obtain data relating to the processing state. The fourth machine learning model is a machine learning model that receives as input data obtained by combining data obtained through partial learning of the first machine learning model and data obtained from the third machine learning model, and outputs data relating to a processing state.

[0107] Evaluation unit 154 evaluates the second machine learning model generated by learning unit 153.

[0108] In the first embodiment, the evaluation unit 154 evaluates the second machine learning model using two methods.

[0109] As a first evaluation procedure, evaluation unit 154 evaluates the second machine learning model by assessing each of a variety of selected transfer learning data elements used when the second machine learning model is generated by learning unit 153. This evaluation procedure is also referred to as the "initial assessment." Suppose learning unit 153 generates the second machine learning model using five selected transfer learning data elements; then evaluation unit 154 assesses each of the five selected transfer learning data elements.

[0110] As a second evaluation procedure, evaluation unit 154 assesses the second machine learning model by evaluating the entire set of selected transfer learning data elements used by learning unit 153 in creating the second machine learning model. This evaluation procedure is also referred to as the "second evaluation." In the example above, evaluation unit 154, for instance, assesses all five selected transfer learning data elements.

[0111] The evaluation unit 154 evaluates the multitude of selected transfer learning data elements by calculating evaluation values ​​for the multitude of selected transfer learning data elements, e.g. using an error (e.g. the mean squared error or the mean absolute error).

[0112] Specifically, the evaluation unit 154 first inputs data relating to the processing condition, contained in the multitude of selected transfer learning data elements, into the second machine learning model and receives data relating to the processing state. At this point, processing state-related data is determined for the number of selected transfer learning data elements. The evaluation unit 154 calculates an evaluation score (hereinafter referred to as the "individual evaluation score") for each selected transfer learning data element by calculating a discrepancy between the processing state-related data obtained from the second machine learning model and the corresponding processing state-related data contained in the multitude of transfer learning data elements. This is the initial evaluation.

[0113] In a concrete example, assume there are five selected transfer learning data elements (first selected transfer learning data element, second selected transfer learning data element, third selected transfer learning data element, fourth selected transfer learning data element, and fifth selected transfer learning data element), and that the data relating to the processing state contained in each selected transfer learning data element are "0.15", "0.32", "0.10", "-0.52", and "-0.11", respectively. It is important to note that each of these processing state data elements is data that has become a teacher label when Learning Unit 153 creates the second machine learning model.

[0114] On the other hand, it is assumed that the processing condition-related processing data elements, which are provided by the evaluation unit 154 by inputting the processing condition-related processing data contained in the first selected transfer learning data, the processing condition-related processing data contained in the second selected transfer learning data, the processing condition-related processing data contained in the third selected transfer learning data, the processing condition-related processing data contained in the fourth selected transfer learning data, and the processing condition-related processing data contained in the fifth selected transfer learning data, to the second machine learning model “0.10”, “0.28”, “0.14”, “-0.14” and “-0.19”, respectively.

[0115] In this case, the assessment unit 154 calculates an individual score of "0.05" (absolute error between "0.15" and "0.10") for the first selected transfer learning data. Furthermore, the assessment unit 154 calculates an individual score of "0.04" (absolute error between "0.32" and "0.28") for the second selected transfer learning data. Furthermore, the assessment unit 154 calculates an individual score of "0.04" (absolute error between "0.10" and "0.14") for the third selected transfer learning data. Furthermore, the assessment unit 154 calculates an individual score of "0.38" (absolute error between "-0.52" and "-0.14") for the fourth selected transfer learning data. Additionally, the assessment unit 154 calculates an individual score of "0.08" (absolute error between "-0.11" and "-0.19") for the fifth selected transfer learning data.

[0116] Next, the assessment unit 154 calculates an average of the assessment scores for each of the selected transfer learning data elements, calculated by performing the initial assessment, thereby calculating a single assessment score (hereafter referred to as the "overall assessment score") for the entire set of selected transfer learning data elements. This is the second assessment.

[0117] In the specific example described above, the evaluation unit 154 calculates an overall evaluation value of “0.118” (the average absolute error of “0.05”, “0.04”, “0.04”, “0.38” and “0.08”) for the entire set of selected transfer learning data elements (first selected transfer learning data, second selected transfer learning data, third selected transfer learning data, fourth selected transfer learning data and fifth selected transfer learning data).

[0118] The evaluation unit 154 outputs each selected transfer learning data element and the corresponding individual evaluation value for the multitude of selected transfer learning data elements in conjunction with the weighting update unit 157.

[0119] Furthermore, the evaluation unit 154 stores the multitude of selected transfer learning data elements and the entire evaluation score in the evaluation score storage unit 156 in conjunction with each other. For example, the evaluation unit 154 can store the multitude of selected transfer learning data elements, the entire evaluation score, and the second machine learning model in the evaluation score storage unit 156 in conjunction with each other.

[0120] It should be noted that the evaluation unit 154 calculates an overall score for the entire set of selected transfer learning data elements and also calculates a score (hereinafter referred to as the "overall test score") for all the processing data for evaluation. The processing data for evaluation is stored by the processing data acquisition unit 11 in the processing data storage unit 112. The evaluation unit 154 generates the second machine learning model based on the processing data for evaluation and the first machine learning model. The evaluation unit 154 can then calculate the overall test score for all the processing data for evaluation using the same procedure as the second evaluation performed for the set of selected transfer learning data elements.

[0121] The evaluation unit 154 stores the processing data for the evaluation and the overall test score in the evaluation score storage unit 156 in conjunction with each other. For example, the evaluation unit 154 can store the processing data for the evaluation, the overall test score, and the second machine learning model in the evaluation score storage unit 156 in conjunction with each other.

[0122] The model storage unit 155 stores the second machine learning model generated by the learning unit 153.

[0123] It should be noted that here the model storage unit 155 is contained within the transfer learning device 1, but this is only an example. The model storage unit 155 can be provided at a location to which the transfer learning device 1 can refer outside of the transfer learning device 1. For example, the model storage unit 155 can be housed in the control device 2 or in the storage device 3.

[0124] The score storage unit 156 stores the multitude of selected transfer learning data elements associated with the total score output by the score unit 154. In addition, the score storage unit 156 stores the processing data for assessment output by the score unit 154 and associated with the total test score.

[0125] It should be noted that here the evaluation value storage unit 156 is contained within the transfer learning device 1, but this is only an example. The evaluation value storage unit 156 can be located outside the transfer learning device 1 at a location that the transfer learning device 1 can access. For example, the evaluation value storage unit 156 could be located in the control device 2 or in the storage device 3.

[0126] The weighting update unit 157 updates the priority assigned to each selected transfer learning data element based on the evaluation result of the second machine learning model performed by the evaluation unit 154.

[0127] The evaluation result of the second machine learning model, which is carried out by the evaluation unit 154, consists of a large number of selected transfer learning data elements that are assigned to individual evaluation values ​​issued by the evaluation unit 154.

[0128] The weighting update unit 157 updates the priority for each part of the selected transfer learning data.

[0129] For example, the Weighting Update Unit 157 updates the priority assigned to each selected transfer learning data item with each of the individual scores associated with the set of selected transfer learning data. Specifically, Weighting Update Unit 157 can calculate the priority by inserting the individual score into a predefined conditional expression for calculating the priority and updating the priority assigned to the selected transfer learning data with the calculated priority. Weighting Update Unit 157 updates the priority in such a way that the priority assigned to the selected transfer learning data becomes higher when the associated individual score is lower.

[0130] The weighting update unit 157 outputs a variety of selected transfer learning data elements, to which the updated priority is assigned to the selection unit 151.

[0131] When the multitude of selected transfer learning data elements to which the updated priority is assigned is output by the weighting update unit 157, the selection unit 151 takes the multitude of selected transfer learning data elements into samples based on the priority.

[0132] For example, selection unit 151 takes the selected transfer learning data with a priority equal to or higher than a preset threshold (hereinafter referred to as the "sampling threshold").

[0133] Subsequently, selection unit 151 outputs the selected transfer learning data to learning unit 153 after sampling. The priority updated by weighting update unit 157 is assigned to the selected transfer learning data that were sampled by selection unit 151.

[0134] When the selected transfer learning data is output by selection unit 151 after sampling, learning unit 153 re-teaches the second machine learning model. As learning unit 153 performs re-learning, evaluation unit 154 assesses the second machine learning model generated by the re-learning, and weight update unit 157 updates the priority. While the selected transfer learning data is output by selection unit 151 after sampling, the learning process by learning unit 153, the assessment of the second machine learning model by evaluation unit 154, and the priority update by weight update unit 157 are repeated.

[0135] It should be noted that selection unit 151 performs a sampling of the selected transfer learning data after determining whether or not relearning by learning unit 153 is required.

[0136] Selection unit 151 stops sampling the selected transfer learning data when it determines that relearning by learning unit 153 is not necessary. This completes the generation of the second machine learning model by learning unit 153.

[0137] For example, if the priority assigned to each selected transfer learning data element updated by the weight update unit 157 is equal to or less than a preset threshold (referred to below as the "first stop threshold"), the selection unit 151 determines that relearning by the learning unit 153 is unnecessary. Similarly, in a case where both the total score and the total test score stored in the score storage unit 156 are equal to or greater than a preset threshold (referred to below as the "second stop threshold"), the selection unit 151 may determine that relearning by the learning unit 153 is not required.

[0138] Once the generation of the second machine learning model by learning unit 153 is complete, control device 2, for example, uses the second machine learning model to infer a positional error in the machining process when it instructs machining machine 4 to perform a machining operation. Control device 2 then adjusts the machining condition so that the positional error can be corrected based on the inference result. Furthermore, control device 2 can inform the user of the inference result, for example, by causing a display device (not shown) contained within control device 2 to display the inference result.

[0139] The operation of the transfer learning device 1 according to the first embodiment is described.

[0140] Fig. Figure 4 is a flowchart to explain the operation of the transfer learning device 1 according to the first embodiment.

[0141] The processing data acquisition unit 11 acquires and stores a large number of processing data elements from the control device 2 (step ST1).

[0142] Specifically, in step ST1, the processing data acquisition unit 111 of the processing data acquisition unit 11 captures the processing data from the control device 2 and formats the captured processing data. Subsequently, the processing data preprocessing unit 111 stores the processed data in the processing data storage unit 112 after formatting.

[0143] Unit 12 for capturing preliminary learning data captures a variety of preliminary learning data elements stored in the preliminary learning data storage unit 31 of storage device 3 (step ST2).

[0144] Unit 12, for capturing preliminary training data, outputs the multitude of received preliminary training data elements to analysis unit 13.

[0145] Analysis Unit 13 calculates the importance for the multitude of post-capture processing data elements based on the multitude of post-capture processing data elements stored in Process Data Storage Unit 112 in Process Data Acquisition Unit 11, and the multitude of preliminary training data elements output by Unit 12 for preliminary training data acquisition (Step ST3).

[0146] The analysis unit 13 outputs each processing data element (processing data after compression) and its importance in relation to each other to the extraction unit 14.

[0147] The extraction unit 14 extracts a variety of machining data elements for the purpose of performing transfer learning of the first machine learning model, i.e. transfer learning data, from the variety of machining data elements that were captured by the machining data acquisition unit 11 in step ST1 on the basis of the importance calculated by the analysis unit 13 in step ST3 (step ST4).

[0148] Extraction unit 14 also extracts a set of features during the extraction of the transfer learning data.

[0149] The extraction unit 14 outputs a large number of transfer learning data elements to the machine learning unit 15, to which the importance calculated by the analysis unit 13 in step ST3 has been assigned and from which the feature set has been extracted.

[0150] The machine learning unit 15 performs the transfer learning of the first machine learning model using the multitude of transfer learning data elements output by the extraction unit 14 in step ST4, and generates the second machine learning model (step ST5).

[0151] Fig. Figure 5 is a flowchart to explain in detail the operation of the analysis unit 13 of the transfer learning device 1 according to the first embodiment. That is to say, Fig. 5 is a flowchart to explain in detail the processing of step ST3 in Fig. 4.

[0152] First, the compression unit 131 in the analysis unit 13 compresses a large number of processing data elements and a large number of preliminary training data elements (step ST31).

[0153] Analysis Unit 132 calculates the importance for each post-compression processing data element based on the multitude of post-compression processing data elements compressed by Compression Unit 131 in step ST31 and the multitude of post-compression preliminary training data elements compressed by Compression Unit 131 (step ST32).

[0154] When the analysis unit 132 calculates the importance for the multitude of processing data elements after compression, the analysis unit 13 outputs each processing data element after compression and its importance in relation to each other to the extraction unit 14.

[0155] Fig. Figure 6 is a flowchart for a detailed explanation of the operation of the machine learning unit 15 of the transfer learning device 1 according to the first embodiment. That is, Fig. 6 is a flowchart to explain in detail the processing of step ST5 in Fig. 4.

[0156] The model reading unit 152 reads the first machine learning model stored in the preliminary learning model storage unit 32 of the storage device 3 (step ST51).

[0157] When a large number of transfer learning data elements are output by the extraction unit 14, the selection unit 151 performs a "processing with priority initial settings" (step ST52).

[0158] Selection unit 151 determines whether transfer learning is required (step ST53). Specifically, selection unit 151 determines whether generating the second machine learning model based on transfer learning is necessary. If the second machine learning model has not yet been created, selection unit 151 determines that transfer learning is required. Conversely, if the second machine learning model has already been created, selection unit 151 determines that transfer learning is unnecessary. It's important to note that, for example, when creating the second machine learning model, learning unit 153 sets a flag indicating model creation is complete, which machine learning unit 15 can reference. Selection unit 151 can determine whether the second machine learning model has been created based on whether the model creation completion flag is set to "1".The flag indicating completion of model generation is initialized, for example, when the second machine learning model is completed.

[0159] If the selection unit 151 determines that transfer learning is necessary (in the case of “YES” in step ST53), it selects the selected transfer learning data elements based on the priority assigned to each editing data element in step ST52 (step ST54) during “Editing with priority initial settings”.

[0160] The learning unit 153 performs transfer learning of the first machine learning model read by the model reading unit 152 in step ST51, using the multitude of selected transfer learning data elements output by the selection unit 151 in step ST54, and generates the second machine learning model (step ST55).

[0161] The learning unit 153 stores the generated second machine learning model in the model storage unit 155.

[0162] Furthermore, the learning unit 153 outputs the generated second machine learning model, together with the selected transfer learning data used in the generation of the second machine learning model, to the evaluation unit 154.

[0163] Evaluation Unit 154 evaluates the second machine learning model generated by Learning Unit 153 in step ST55 (step ST56). Specifically, Evaluation Unit 154 evaluates each of the multiple selected transfer learning data elements used by Learning Unit 153 in the creation of the second machine learning model and evaluates the entire majority of the selected transfer learning data elements.

[0164] The evaluation unit 154 outputs each selected transfer learning data element and the corresponding individual evaluation value for the multitude of selected transfer learning data elements in conjunction with the weighting update unit 157.

[0165] Furthermore, the evaluation unit 154 stores the multitude of selected transfer learning data elements and the entire evaluation score in the evaluation score storage unit 156 in conjunction with each other. For example, the evaluation unit 154 can store the multitude of selected transfer learning data elements, the entire evaluation score, and the second machine learning model in the evaluation score storage unit 156 in conjunction with each other.

[0166] It should be noted that in step ST56, the assessment unit 154 also calculates an overall test score for all processing data submitted for assessment by the assessment unit 154. The assessment unit 154 stores the processing data for the assessment and the overall test score in the assessment score storage unit 156, linked together.

[0167] The weighting update unit 157 updates the priority assigned to each selected transfer learning data element based on the evaluation result of the second machine learning model, which was performed by the evaluation unit 154 in step ST56 (step ST57). Afterwards, the operation of the machine learning unit 15 returns to step ST53.

[0168] On the other hand, in a case where step ST53 determines that no transfer learning is required (“NO” in step ST53), i.e., in a case where the second machine learning model has been generated, the selection unit 151 determines whether relearning by the learning unit 153 is required or not (step ST58).

[0169] If it is determined that relearning is required by learning unit 153 (in the case of “YES” in step ST58), selection unit 151 takes a sample of the selected transfer learning data (step ST59).

[0170] Selection unit 151 outputs the selected transfer learning data to learning unit 153 after sampling.

[0171] When the selected transfer learning data are output by the selection unit 151 in step ST59 after sampling, the learning unit 153 re-learns the second machine learning model (step ST60).

[0172] Machine learning unit 15 then proceeds to process step ST56. It should be noted that in this case, evaluation unit 154 evaluates the second machine learning model in step ST56, which was newly learned by learning unit 153. Afterward, step ST56 is processed, and machine learning unit 15 returns to processing step ST53.

[0173] On the other hand, in step ST58, selection unit 151 stops sampling the selected transfer learning data if selection unit 151 determines that relearning by learning unit 153 is not necessary (in the case of "NO" in step ST58). This completes the generation of the second machine learning model by learning unit 153. Subsequently, machine learning unit 15 terminates the process shown in the flowchart of Fig. 6 shown processing.

[0174] Regarding the operation of the transfer learning device 1, which refers to the in Fig. As described in the flowchart shown in section 4, it should be noted that the processing follows the sequence of steps ST1 and ST2 in the flowchart. Fig. Step 4 is performed, but it is not strictly necessary to perform the processing in this order. For example, the processing can be carried out in the order of step ST2 and step ST1, or the processing of step ST1 and the processing of step ST2 can be carried out in parallel.

[0175] Regarding the detailed operation of machine learning unit 15, which is described with reference to the in Fig. The flowchart shown in section 6 describes the following: Fig. 6. First, the processing of step ST51 was carried out, which is only an example. The processing of step ST51 can be completed before the processing of step ST55 is carried out.

[0176] As described above, the transfer learning device 1 calculates the importance of the multitude of machining data elements acquired by the control device 2 and extracts a multitude of transfer learning data elements from this multitude of machining data elements based on their importance. Subsequently, the transfer learning device 1 performs the transfer learning of the first machine learning model using the multiple machining data elements from the extracted transfer learning data and generates the second machine learning model.

[0177] As described above, the amount of machining data acquired by the control device 2 at the machine tool 4 is enormous. Therefore, if the transfer learning device 1 simply performs transfer learning using all machining data elements, it will take some time to create the second machine learning model, even if the transfer learning is efficient. On the other hand, according to the first embodiment, the transfer learning device 1 calculates the importance of each machining data element acquired by the control device 2 and extracts the transfer learning data elements based on this calculated importance.

[0178] This allows the transfer learning device 1 to reduce the amount of data used to create the second machine learning model. In other words, the transfer learning device 1 does not need all the machining data elements collected at the machining location and can generate the second machine learning model for setting the machining conditions from only a portion of the machining data. This model is suitable for the situation of machining machine 4 at the location where it is actually operating. As a result, the transfer learning device 1 can create the second machine learning model efficiently and without significant time expenditure.

[0179] It should be noted that the transfer learning device 1 extracts the transfer learning data based on importance; in other words, it extracts as transfer learning data the processing data that is expected to be highly effective when used as training data in the creation of the second machine learning model. Therefore, the processing data suitable for creating the second machine learning model can be used as transfer learning data, even if the quantity is reduced.

[0180] Furthermore, the transfer learning device 1 assigns a priority to each part of the extracted transfer learning data based on its importance and selects a variety of selected transfer learning data elements from this variety based on priority. Subsequently, the transfer learning device 1 performs the transfer learning of the first machine learning model using the multiple processing data elements of the selected transfer learning data elements.

[0181] In this way, Transfer Learning Device 1 can further reduce the amount of data used to create the second machine learning model. As a result, Transfer Learning Device 1 can create the second machine learning model more efficiently.

[0182] Furthermore, the transfer learning device 1 evaluates the generated second machine learning model and updates the priority assigned to the selected transfer learning data based on the evaluation result. Subsequently, based on the updated priority, the transfer learning device 1 takes samples of the selected transfer learning data and retrains the second machine learning model using the selected transfer learning data after sampling.

[0183] Thus, the transfer learning device 1 can increase the accuracy of the generated second machine learning model.

[0184] Fig. 7A and Fig. Figure 7B shows an example of a hardware configuration of the transfer learning device 1 according to the first embodiment.

[0185] In the first embodiment, the functions of the machining data acquisition unit 11, the unit 12 for acquiring preliminary learning data, the analysis unit 13, the extraction unit 14, and the machine learning unit 15 are implemented by a machining circuit 501. That is, the transfer learning device 1 includes the machining circuit 501 for performing the control to generate the second machine learning model by performing transfer learning based on the first machine learning model, which was generated in advance using the transfer learning data extracted from the multitude of machining data elements acquired by the control device 2.

[0186] The 501 processing circuit can be a special piece of hardware, as in Fig. 7A is shown, or a 504 processor executing a program stored in memory, as in Fig. 7B is shown.

[0187] If the 501 processing circuit is dedicated hardware, then the 501 processing circuit corresponds, for example, to a single circuit, a composite circuit, a programmed processor, a parallel programmed processor, an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or a combination thereof.

[0188] If the processing circuit is the processor 504, the functions of the processing data acquisition unit 11, the preliminary learning unit 12, the analysis unit 13, the extraction unit 14, and the machine learning unit 15 are implemented by software, firmware, or a combination of software and firmware. The software or firmware is described as a program and stored in a memory 505. The processor 504 reads the program stored in the memory 505 and executes it, thereby performing the functions of the processing data acquisition unit 11, the preliminary learning unit 12, the analysis unit 13, the extraction unit 14, and the machine learning unit 15. That is, the transfer learning device 1 contains the memory 505 for storing a program which, when executed by the processor 504, performs the functions described above. Fig.The four steps described in ST1 to ST5 lead to this. Furthermore, it can also be said that the program stored in memory 505 causes a computer to execute a procedure or process that is carried out during the processing of the processing data acquisition unit 11, the preliminary training acquisition unit 12, the analysis unit 13, the extraction unit 14, and the machine learning unit 15. Here, memory 505 corresponds, for example, to non-volatile or volatile semiconductor memory such as RAM, read-only memory (ROM), flash memory, erasable programmable read-only memory (EPROM), or electrically erasable programmable read-only memory (EEPROM), or to a magnetic disk, a flexible disk, an optical disk, a compact disk, a mini-disk, a digital versatile disc (DVD), or similar.

[0189] It should be noted that some of the functions of the processing data acquisition unit 11, the preliminary training unit 12, the analysis unit 13, the extraction unit 14, and the machine learning unit 15 can be implemented by dedicated hardware, and some by software or firmware. For example, the functions of the processing data acquisition unit 11 and the preliminary training unit 12 can be implemented by the processing circuit 501 as dedicated hardware, and the functions of the analysis unit 13, the extraction unit 14, and the machine learning unit 15 can be implemented by the processor 504, which reads and executes programs stored in memory 505.

[0190] The model storage unit 155 and the rating value storage unit 156 include, for example, a memory.

[0191] Furthermore, the transfer learning device 1 includes a device such as the control device 2 or the storage device 3, as well as an input interface device 502 and an output interface device 503, which perform wired or wireless communication.

[0192] It should be noted that, as described above, in the first embodiment, when the analysis unit 13 calculates the importance of a plurality of processing data elements, the analysis unit 13 does not require a plurality of preliminary training data elements when the analysis unit calculates the importance using the change point detection method. Therefore, in this case, the transfer training device 1 can be configured so that it does not include a unit 12 for acquiring preliminary training data.

[0193] In the first embodiment of the transfer learning device 1 described above, the analysis unit 13 also includes the compression unit 131, but this is only an example. The analysis unit 13 does not necessarily include the compression unit 131. In this case, the analysis unit 132 calculates the importance for each processing data element based on a plurality of uncompressed processing data elements and a plurality of uncompressed preliminary learning data elements. The function of the analysis unit 132 may be included in the analysis unit 13, and the analysis unit 13 may not include the analysis unit 132.

[0194] Furthermore, in the first embodiment described above, the transfer learning device 1 includes the evaluation unit 154 and the weight update unit 157, but this is only an example. The transfer learning device 1 can be configured so that it does not include the evaluation unit 154 and the weight update unit 157. In this case, the transfer learning device 1 does not have the function of relearning the second machine learning model.

[0195] Furthermore, the transfer learning device 1 can function without the selection unit 151. In this case, the learning unit 153 in the transfer learning device 1 can generate the second machine learning model by performing the transfer learning of the first machine learning model using the multitude of transfer learning data elements extracted by the extraction unit 14.

[0196] In the first embodiment of the transfer learning device 1 described above, the machining data acquisition unit 11 also includes the machining data preprocessing unit 111, but this is only an example. The machining data acquisition unit 11 does not necessarily include the machining data preprocessing unit 111. If the machining data preprocessing unit 111 is not present, the machining data acquisition unit 11 stores a plurality of machining data elements acquired by the control device 2 in the machining data storage unit 112.

[0197] Furthermore, in the first embodiment described above, the transfer learning device 1 can be attached, for example, to the control device 2 or the machining center 4, or to a server (not shown) in order to communicate with the control device 2. In addition, for example, part of the machining data acquisition unit 11, the preliminary learning data acquisition unit 12, the analysis unit 13, the extraction unit 14, and the machine learning unit 15 can be contained in the control device 2 or the machining center 4, and the others can be contained in the server.

[0198] Furthermore, in the first embodiment described above, it is assumed that the machining state is a positional error of the machining position actually performed by the machining machine 4 with respect to the machining position specified by the machining machine 4. However, this is only an example. In the first embodiment described above, the machining state could, for example, consist of whether the machining position actually performed by the machining machine 4 has a positional error that is equal to or greater than a specified range with respect to the machining position specified by the machining machine 4.

[0199] As described above, the transfer learning device 1 according to the first embodiment comprises: the machining data acquisition unit 11 for acquiring a plurality of machining data elements, including data relating to time-series machining conditions to cause a machining machine 4 to perform machining for a plurality of steps, and data relating to a machining state executed by the machining machine 4 for the plurality of steps based on the machining conditions; the analysis unit 13 for calculating an importance with respect to each machining data element acquired by the machining data acquisition unit 11, based on a plurality of machining data elements acquired by the machining data acquisition unit 11;the extraction unit 14 for extracting a multitude of transfer learning data elements for performing the transfer learning of a first machine learning model, which receives as input the processing condition-related data in time series for the multitude of steps, which were generated in advance from a multitude of processing data elements that were captured by the processing data acquisition unit 11 on the basis of the importance calculated by the analysis unit 13, and which outputs data relating to a processing state after the execution of the multitude of steps;and the learning unit 153 for performing transfer learning of the first machine learning model using a multitude of transfer learning data elements extracted by the extraction unit 14, and for generating a second machine learning model that receives as input data relating to the machining conditions in time series for the multitude of steps, and outputs data relating to a machining state after execution for the multitude of steps. Consequently, the transfer learning device 1 does not require all the data collected at the machining location and can generate the second machine learning model for setting the machining conditions from a portion of the data, which is suitable for the situation of the machining machine at the location where the machining machine is actually working. Thus, the second machine learning model can be created without significant time expenditure.

[0200] It should be noted that any component of the embodiment can be modified or any component of the embodiment can be omitted. INDUSTRIAL APPLICABILITY

[0201] The transfer learning device according to the present disclosure can be used as a transfer learning device that generates a second machine learning model at a machining location for setting a machining condition suitable for a situation of a machining machine at the location where the machining machine actually works. REFERENCE MARK LIST

[0202] 1: Transfer learning device, 11: Machining data acquisition unit, 111: Machining data preprocessing unit, 112: Machining data storage unit, 13: Analysis unit, 131: Compression unit, 132: Analysis unit, 14: Extraction unit, 15: Machine learning unit, 151: Selection unit, 152: Model reading unit, 153: Learning unit, 154: Evaluation unit, 155: Model storage unit, 156: Evaluation value storage unit, 157: Weight update unit, 2: Control device, 3: Storage device, 31: Preliminary learning data storage unit, 32: Preliminary learning model storage unit, 4: Machining machine, 501: Machining circuit, 502: Input interface device, 503: Output interface device, 504: Processor, 505: Memory

Claims

[1] Transfer learning device comprising: a machining data acquisition unit (11) for acquiring a plurality of machining data items including data relating to time-series machining conditions for causing a machining machine (4) to perform machining for a plurality of steps and data relating to a machining state executed by the machining machine (4) for the plurality of steps based on the machining conditions; an analysis unit (13) for calculating an importance with respect to each piece of the machining data acquired by the machining data acquisition unit (11) based on a plurality of machining data items acquired by the machining data acquisition unit (11); an extraction unit (14) for extracting a plurality of transfer learning data items for performing transfer learning of a first machine learning model, which receives as an input the data relating to the machining conditions in time series for the plurality of steps generated in advance from a plurality of machining data items acquired by the machining data acquisition unit based on the importance calculated by the analysis unit (13), and outputs data relating to a machining state after execution for the plurality of steps; and a learning unit (153) for performing transfer learning of the first machine learning model using a plurality of transfer learning data items extracted by the extraction unit (14), and for generating a second machine learning model that receives as input data related to the machining conditions in time series for the plurality of steps and outputs data related to a machining state after execution for the plurality of steps. [2] The transfer learning apparatus according to claim 1, further comprising a preliminary learning data acquisition unit (12) for acquiring a plurality of preliminary learning data items including data used at the time of generating the first machine learning model and relating to the machining conditions in time series for the plurality of steps, and data relating to a machining state after execution for the plurality of steps, wherein the analysis unit (13) calculates the importance of each piece of the machining data acquired by the machining data acquisition unit based on a plurality of machining data items acquired by the machining data acquisition unit (11) and a plurality of preliminary learning data items acquired by the preliminary learning data acquisition unit (12). [3] The transfer learning apparatus according to claim 1, further comprising a selecting unit (151) for assigning a priority to each piece of transfer learning data extracted by the extracting unit (14) and selecting a plurality of selected transfer learning data items from a plurality of transfer learning data items based on the priority, wherein the learning unit (153) performs the transfer learning of the first machine learning model using a plurality of selected transfer learning data items selected by the selection unit (151). [4] Transfer learning device according to claim 3, further comprising: an evaluation unit (154) for evaluating the second machine learning model generated by the learning unit (153); and a weight updating unit (157) for updating the priority assigned to the selected transfer learning data item based on an evaluation result of the second machine learning model performed by the evaluation unit (154), wherein the selection unit (151) samples the selected transfer learning data based on the priority updated by the weight update unit (157), and the learning unit (153) relearns the second machine learning model based on the selected transfer learning data after the selection unit (151) has performed the sampling. [5] Transfer learning device according to claim 2, wherein the analysis unit (13) comprises: a compression unit (131) for compressing a plurality of processing data items and a plurality of preliminary learning data items; and an analysis unit (132) for calculating the importance of each piece of machining data acquired by the machining data acquisition unit (11) based on a plurality of machining data items and a plurality of preliminary learning data items after being compressed by the compression unit (131). [6] Transfer learning device according to claim 1, wherein the extraction unit (14) extracts feature sets of a plurality of transfer learning data items, and the learning unit (153) performs the transfer learning of the first machine learning model based on the feature set extracted by the extraction unit (14). [7] Transfer learning method, which includes the steps: Acquiring, by a machining data acquiring unit, a plurality of machining data items including data related to time-series machining conditions for causing a machining machine to perform machining for a plurality of steps and data related to a machining state executed by the machining machine for the plurality of steps based on the machining conditions; Calculating, by an analysis unit, an importance with respect to each piece of the machining data acquired by the machining data acquisition unit, based on a plurality of machining data items acquired by the machining data acquisition unit; Extracting, by an extraction unit, a plurality of transfer learning data items for performing transfer learning of a first machine learning model that receives as an input the data related to the machining conditions in time series for the plurality of steps generated in advance from a plurality of machining data items acquired by the machining data acquisition unit based on the importance calculated by the analysis unit, and outputs data related to a machining state after execution for the plurality of steps; and Performing, by a learning unit, transfer learning of the first machine learning model using a plurality of transfer learning data items extracted by the extraction unit, and generating a second machine learning model that receives as input data related to machining conditions in time series for the plurality of steps and outputs data related to a machining state after execution for the plurality of steps.

Images